feat(core): close live capabilities and cache ownership

Wait for bounded plugin readiness in the generation location before resolving explicit or default models. Add deterministic cold-first-request regressions through the embedded SDK.

.changeset/rename-undo.md

@@ -0,0 +1,5 @@
+---
+"@opencode-ai/core": patch
+---
+
+Include both paths of renamed files in new snapshot change lists so undo restores the original file instead of only deleting the renamed file.

.changeset/session-capabilities.md

@@ -0,0 +1,8 @@
+---
+"@opencode-ai/core": minor
+"@opencode-ai/schema": patch
+---
+
+Open durable sessions with in-process model, tool, instruction, and permission capabilities. Live Sources update at safe boundaries through existing instruction epochs, while capability replacement waits for the next busy period. Capability-owned sessions remain pending after restart until their host reopens and drives them.
+
+Close an open's in-process capabilities after settlement without deleting durable history. Tool executors may yield domain errors, which normalize to tool failures while canonical permission declines retain their interruption behavior.

.changeset/shell-scanner-compatibility.md

@@ -0,0 +1,10 @@
+---
+"@opencode-ai/core": patch
+---
+
+Make the experimental portable shell scanner authoritative, with no Tree-sitter
+fallback. Scan common Bash and PowerShell control flow, heredocs, functions,
+expressions, quoting, and substitutions natively. Preserve existing redirect and
+declaration permission matching, and make PowerShell saved approvals cover the
+original command spelling. Parser failures remain visible without changing the
+permission engine. The default Tree-sitter path is unchanged.

.github/workflows/test.yml

@@ -237,9 +237,9 @@ jobs:
         working-directory: packages/app
         run: bunx playwright install chromium
 
-      - name: Run app e2e tests
+      - name: Run app e2e tests against production build
         if: env.E2E_ENABLED == 'true'
-        run: bun --cwd packages/app test:e2e:local
+        run: bun --cwd packages/app test:e2e:built
         env:
           CI: true
         timeout-minutes: 30

bun.lock

@@ -125,7 +125,7 @@
         "@effect/platform-node": "catalog:",
         "@opencode-ai/client": "workspace:*",
         "@opencode-ai/plugin": "workspace:*",
-        "@opencode-ai/pty": "0.1.10",
+        "@opencode-ai/pty": "0.1.12",
         "@opencode-ai/schema": "workspace:*",
         "@opencode-ai/server": "workspace:*",
         "@opencode-ai/tui": "workspace:*",
@@ -364,7 +364,7 @@
         "@opencode-ai/ai": "workspace:*",
         "@opencode-ai/codemode": "workspace:*",
         "@opencode-ai/plugin": "workspace:*",
-        "@opencode-ai/pty": "0.1.10",
+        "@opencode-ai/pty": "0.1.12",
         "@opencode-ai/schema": "workspace:*",
         "@opencode-ai/util": "workspace:*",
         "@parcel/watcher": "2.5.1",
@@ -553,6 +553,20 @@
         "@typescript/native-preview": "catalog:",
       },
     },
+    "packages/latex": {
+      "name": "@opencode-ai/latex",
+      "version": "0.0.0",
+      "dependencies": {
+        "@opencode-ai/plugin": "workspace:*",
+        "@opentui/core": "catalog:",
+        "string-width": "catalog:",
+      },
+      "devDependencies": {
+        "@tsconfig/bun": "catalog:",
+        "@types/bun": "catalog:",
+        "@typescript/native-preview": "catalog:",
+      },
+    },
     "packages/merman": {
       "name": "@opencode-ai/merman",
       "version": "0.0.0",
@@ -879,6 +893,7 @@
       "dependencies": {
         "@opencode-ai/client": "workspace:*",
         "@opencode-ai/core": "workspace:*",
+        "@opencode-ai/latex": "workspace:*",
         "@opencode-ai/merman": "workspace:*",
         "@opencode-ai/plugin": "workspace:*",
         "@opencode-ai/schema": "workspace:*",
@@ -2150,6 +2165,8 @@
 
     "@opencode-ai/httpapi-codegen": ["@opencode-ai/httpapi-codegen@workspace:packages/httpapi-codegen"],
 
+    "@opencode-ai/latex": ["@opencode-ai/latex@workspace:packages/latex"],
+
     "@opencode-ai/merman": ["@opencode-ai/merman@workspace:packages/merman"],
 
     "@opencode-ai/plugin": ["@opencode-ai/plugin@workspace:packages/plugin"],
@@ -2158,19 +2175,19 @@
 
     "@opencode-ai/protocol": ["@opencode-ai/protocol@workspace:packages/protocol"],
 
-    "@opencode-ai/pty": ["@opencode-ai/pty@0.1.10", "", { "optionalDependencies": { "@opencode-ai/pty-darwin-arm64": "0.1.10", "@opencode-ai/pty-darwin-x64": "0.1.10", "@opencode-ai/pty-linux-arm64-gnu": "0.1.10", "@opencode-ai/pty-linux-arm64-musl": "0.1.10", "@opencode-ai/pty-linux-x64-gnu": "0.1.10", "@opencode-ai/pty-linux-x64-musl": "0.1.10" }, "bin": { "opencode-pty": "bin/opencode-pty.js" } }, "sha512-cEJT1ADtmnb+df2wrlUcsGny6Q7pTe9Sa7keISzCO0xN1FrL1aS6+eleBPpDimHjgM/sXqvLwJv0UiAeiAvgxQ=="],
+    "@opencode-ai/pty": ["@opencode-ai/pty@0.1.12", "", { "optionalDependencies": { "@opencode-ai/pty-darwin-arm64": "0.1.12", "@opencode-ai/pty-darwin-x64": "0.1.12", "@opencode-ai/pty-linux-arm64-gnu": "0.1.12", "@opencode-ai/pty-linux-arm64-musl": "0.1.12", "@opencode-ai/pty-linux-x64-gnu": "0.1.12", "@opencode-ai/pty-linux-x64-musl": "0.1.12" }, "bin": { "opencode-pty": "bin/opencode-pty.js" } }, "sha512-dl4FyJUhTXThsWYY8txG/8/nwN7dE0M5Sic9r4L9f2pvtJnbR5zrCrPoiPIBIxZle1wVks1dhz4z/CfqLf5sCg=="],
 
-    "@opencode-ai/pty-darwin-arm64": ["@opencode-ai/pty-darwin-arm64@0.1.10", "", { "os": "darwin", "cpu": "arm64" }, "sha512-j7aszDFRwCIazGUT9eIy4PZwh4rltjvRmoicPRTK3kONN3v0MMflstkmAFDYYpqDPTNh3qJ6xkQmB+DugEbhAg=="],
+    "@opencode-ai/pty-darwin-arm64": ["@opencode-ai/pty-darwin-arm64@0.1.12", "", { "os": "darwin", "cpu": "arm64" }, "sha512-tMvoriq3VegVlj1uEglc6qE0M7VXy61nyf9Si7tTO7xa8JiyxuFJSXOZ1pGeErDu+pe24hvTyVOR+gkdew8w9g=="],
 
-    "@opencode-ai/pty-darwin-x64": ["@opencode-ai/pty-darwin-x64@0.1.10", "", { "os": "darwin", "cpu": "x64" }, "sha512-UAMP/E4lo9RGQF7xrfIwpW2ZEemj308rCogJy14ruKYJt5MwHeGNTynGiHE/1JlDLRy+21wV50jpugADgT71ag=="],
+    "@opencode-ai/pty-darwin-x64": ["@opencode-ai/pty-darwin-x64@0.1.12", "", { "os": "darwin", "cpu": "x64" }, "sha512-Sn5vMLL5giHOhx7J5H6zwDp4YjjXorY+QV0IEYY+SCT4wQfRBliokIyj23pRl6P2RK3u9bDLXJHDNMfDVZ2Rxg=="],
 
-    "@opencode-ai/pty-linux-arm64-gnu": ["@opencode-ai/pty-linux-arm64-gnu@0.1.10", "", { "os": "linux", "cpu": "arm64" }, "sha512-lTPlZNQ66koFHZqoPmvvq0SetlepKVQYgnLryhlVfYtcryWDJM7gV4+P66V12RwqWQTjt2u8j12mtg3axSKg2w=="],
+    "@opencode-ai/pty-linux-arm64-gnu": ["@opencode-ai/pty-linux-arm64-gnu@0.1.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-HbnlKZy052l7G527wK0+05EXaUpZ4ykVAmNBEzqWCoi4TeQj2+Nr9kJ9trx9o1KrVcT4Ki58CCvN5QOls6Z0yQ=="],
 
-    "@opencode-ai/pty-linux-arm64-musl": ["@opencode-ai/pty-linux-arm64-musl@0.1.10", "", { "os": "linux", "cpu": "arm64" }, "sha512-IDmWHRylMR/ZfMw9/AAktO/Edi4TITPC+Tq7Xx3JZHsDgSba3QdyE11uNL0zM1myTGdk6Yrt4rpdAzaItPnDjw=="],
+    "@opencode-ai/pty-linux-arm64-musl": ["@opencode-ai/pty-linux-arm64-musl@0.1.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-2nTN7ggu1h9XgjNcoQMYjP5sirfYnAskpdFCOqjokLqhytX/IMMmkRTQs+foaEaPz0dAIQD3DQplR2jZIgxp1w=="],
 
-    "@opencode-ai/pty-linux-x64-gnu": ["@opencode-ai/pty-linux-x64-gnu@0.1.10", "", { "os": "linux", "cpu": "x64" }, "sha512-Q1yob0/8X2JoJZzFmNKUc32XDRAe0avKQ8PLKkpJr30qWXSrGmhltgcDmn94Q70zW9Ght9on84T7cmge9brvdQ=="],
+    "@opencode-ai/pty-linux-x64-gnu": ["@opencode-ai/pty-linux-x64-gnu@0.1.12", "", { "os": "linux", "cpu": "x64" }, "sha512-FnD5ndnObTQKAoaVvxLKi5W+r3/+dsaMsobz6uK0B9hlmffXxY5CQ6HQyUU/h3aIKLWXxho5XYkA2b9yrp8/gA=="],
 
-    "@opencode-ai/pty-linux-x64-musl": ["@opencode-ai/pty-linux-x64-musl@0.1.10", "", { "os": "linux", "cpu": "x64" }, "sha512-7RLHWQxX/wfUKJJP2ZMMtkXaPsrgoMNKzE6PL/LbnYbMBtkqfld9EDcMv1RFZ0CqjNFgI0Hg4eRk6x+ZNc/wyQ=="],
+    "@opencode-ai/pty-linux-x64-musl": ["@opencode-ai/pty-linux-x64-musl@0.1.12", "", { "os": "linux", "cpu": "x64" }, "sha512-prkrNu6uvjqoffxdGiDHSU5C0Y+kCSfv+lslu7dfRPgPKenVELNpRTAbOduyrWPac2vGt8j5NM61icJyodbJmA=="],
 
     "@opencode-ai/schema": ["@opencode-ai/schema@workspace:packages/schema"],
 

nix/hashes.json

@@ -1,8 +1,8 @@
 {
   "nodeModules": {
-    "x86_64-linux": "sha256-XHUy+Hk+RHUHREt4x0HfSzr3RlEvgBd4H/fV0rlXw2M=",
-    "aarch64-linux": "sha256-/gIaM62uV2X6KCnkSi6QjyT7t7uJ2L7u7CxXxPQYK+w=",
-    "aarch64-darwin": "sha256-PWG6ALh6kG7mnC6AzEuIAU54BEQ4IB+SyyQFGb/s+Dc=",
-    "x86_64-darwin": "sha256-TcgRDHG4CAT+XoCi0JNansONjg06ZPRmeLRsqmdJ4B4="
+    "x86_64-linux": "sha256-iYdVrLtyKmjlyypisF9SqzgyriWT90kSCh3crxw9AKU=",
+    "aarch64-linux": "sha256-BV2t4w5ujArbtSC/Qfm3gLzevQW9A6hMgOyPVp94g/o=",
+    "aarch64-darwin": "sha256-EwMq7zaxzzcsmH0Pjqu4ftGdcM8Lna8mvHgKzRcVI8g=",
+    "x86_64-darwin": "sha256-PokzxlkQy6JvHADF2ZMIIDI1u9ZjSNNedpmR9gvHS5c="
   }
 }

packages/ai/src/protocols/anthropic-messages.ts

@@ -406,6 +406,7 @@ const AnthropicEvent = Schema.Struct({
 type AnthropicEvent = Schema.Schema.Type<typeof AnthropicEvent>
 
 interface ParserState {
+  readonly providerMetadataKey: string
   readonly tools: ToolStream.State<number>
   readonly reasoningSignatures: Readonly<Record<number, string>>
   readonly usage?: Usage
@@ -440,18 +441,18 @@ const cacheControl = (breakpoints: Cache.Breakpoints, cache: CacheHint | undefin
   return Cache.ttlBucket(cache.ttlSeconds) === "1h" ? EPHEMERAL_1H : EPHEMERAL_5M
 }
 
-const anthropicMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ anthropic: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
 
-const signatureFromMetadata = (metadata: ProviderMetadata | undefined): string | undefined => {
-  const anthropic = metadata?.anthropic
-  if (!ProviderShared.isRecord(anthropic)) return undefined
-  return typeof anthropic.signature === "string" ? anthropic.signature : undefined
+const signatureFromMetadata = (metadata: ProviderMetadata | undefined, key: string): string | undefined => {
+  const provider = metadata?.[key]
+  if (!ProviderShared.isRecord(provider)) return undefined
+  return typeof provider.signature === "string" ? provider.signature : undefined
 }
 
-const redactedDataFromMetadata = (metadata: ProviderMetadata | undefined): string | undefined => {
-  const anthropic = metadata?.anthropic
-  if (!ProviderShared.isRecord(anthropic)) return undefined
-  return typeof anthropic.redactedData === "string" ? anthropic.redactedData : undefined
+const redactedDataFromMetadata = (metadata: ProviderMetadata | undefined, key: string): string | undefined => {
+  const provider = metadata?.[key]
+  if (!ProviderShared.isRecord(provider)) return undefined
+  return typeof provider.redactedData === "string" ? provider.redactedData : undefined
 }
 
 const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition, inputSchema: JsonSchema): AnthropicTool => ({
@@ -511,13 +512,16 @@ const serverToolResultType = (name: string): AnthropicServerToolResultType | und
   return undefined
 }
 
-const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult")(function* (part: ToolResultPart) {
+const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult")(function* (
+  part: ToolResultPart,
+  providerMetadataKey: string,
+) {
   const wireType = serverToolResultType(part.name)
   if (!wireType)
     return yield* invalid(`Anthropic Messages does not know how to round-trip server tool result for ${part.name}`)
   // Prefer the provider-owned replay payload; fall back to the result value for
   // histories constructed directly from provider events.
-  const payload = part.providerMetadata?.anthropic?.["result"] ?? part.result.value
+  const payload = part.providerMetadata?.[providerMetadataKey]?.["result"] ?? part.result.value
   return {
     type: wireType,
     tool_use_id: scrubToolCallID(part.id),
@@ -804,6 +808,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
   breakpoints: Cache.Breakpoints,
 ) {
   const messages: AnthropicMessage[] = []
+  const providerMetadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
 
   for (const [index, message] of request.messages.entries()) {
     if (message.role === "system") {
@@ -849,8 +854,8 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
         if (part.type === "reasoning") {
           // A signature marks visible thinking; only signature-less parts carrying
           // redactedData round-trip as opaque redacted_thinking blocks.
-          const signature = part.encrypted ?? signatureFromMetadata(part.providerMetadata)
-          const redactedData = redactedDataFromMetadata(part.providerMetadata)
+          const signature = part.encrypted ?? signatureFromMetadata(part.providerMetadata, providerMetadataKey)
+          const redactedData = redactedDataFromMetadata(part.providerMetadata, providerMetadataKey)
           if (signature === undefined && redactedData !== undefined) {
             content.push({ type: "redacted_thinking", data: redactedData })
             continue
@@ -879,7 +884,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
           continue
         }
         if (part.type === "tool-result" && part.providerExecuted) {
-          content.push(yield* lowerServerToolResult(part))
+          content.push(yield* lowerServerToolResult(part, providerMetadataKey))
           continue
         }
         return yield* invalid(
@@ -1069,7 +1074,7 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // inclusive `inputTokens` the rest of the contract expects. Extended
 // thinking tokens are included in `output_tokens`; newer responses also
 // expose that subset through `output_tokens_details.thinking_tokens`.
-const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
+const mapUsage = (usage: AnthropicUsage | undefined, providerMetadataKey: string): Usage | undefined => {
   if (!usage) return undefined
   const nonCached = usage.input_tokens ?? undefined
   const cacheRead = usage.cache_read_input_tokens ?? undefined
@@ -1083,7 +1088,7 @@ const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
     cacheWriteInputTokens: cacheWrite,
     reasoningTokens: usage.output_tokens_details?.thinking_tokens,
     totalTokens: ProviderShared.totalTokens(inputTokens, usage.output_tokens, undefined),
-    providerMetadata: { anthropic: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
   })
 }
 
@@ -1092,7 +1097,7 @@ const mapUsage = (usage: AnthropicUsage | undefined): Usage | undefined => {
 // field prefers `right` when defined, falls back to `left`. `inputTokens` is
 // recomputed from the merged breakdown so the inclusive total stays
 // consistent with `nonCached + cacheRead + cacheWrite`.
-const mergeUsage = (left: Usage | undefined, right: Usage | undefined) => {
+const mergeUsage = (left: Usage | undefined, right: Usage | undefined, providerMetadataKey: string) => {
   if (!left) return right
   if (!right) return left
   const nonCachedInputTokens = right.nonCachedInputTokens ?? left.nonCachedInputTokens
@@ -1110,7 +1115,9 @@ const mergeUsage = (left: Usage | undefined, right: Usage | undefined) => {
     reasoningTokens,
     totalTokens: ProviderShared.totalTokens(inputTokens, outputTokens, undefined),
     providerMetadata: {
-      anthropic: mergeJsonRecords(left.providerMetadata?.["anthropic"], right.providerMetadata?.["anthropic"]) ?? {},
+      [providerMetadataKey]:
+        mergeJsonRecords(left.providerMetadata?.[providerMetadataKey], right.providerMetadata?.[providerMetadataKey]) ??
+        {},
     },
   })
 }
@@ -1128,7 +1135,7 @@ const SERVER_TOOL_RESULT_NAMES: Record<AnthropicServerToolResultType, string> =
 
 const isServerToolResultType = (type: string): type is AnthropicServerToolResultType => type in SERVER_TOOL_RESULT_NAMES
 
-const serverToolResultEvent = (block: AnthropicStreamBlock): LLMEvent | undefined => {
+const serverToolResultEvent = (block: AnthropicStreamBlock, providerMetadataKey: string): LLMEvent | undefined => {
   if (!block.type || !isServerToolResultType(block.type)) return undefined
   const errorPayload =
     typeof block.content === "object" && block.content !== null && "type" in block.content
@@ -1142,7 +1149,7 @@ const serverToolResultEvent = (block: AnthropicStreamBlock): LLMEvent | undefine
     providerExecuted: true,
     // The complete payload is irreducible provider replay state: subsequent
     // stateless requests must round-trip the typed result block verbatim.
-    providerMetadata: anthropicMetadata({ blockType: block.type, result: block.content }),
+    providerMetadata: providerMetadata(providerMetadataKey, { blockType: block.type, result: block.content }),
   })
 }
 
@@ -1151,8 +1158,8 @@ type StepResult = readonly [ParserState, ReadonlyArray<LLMEvent>]
 const NO_EVENTS: StepResult["1"] = []
 
 const onMessageStart = (state: ParserState, event: AnthropicEvent): StepResult => {
-  const usage = mapUsage(event.message?.usage)
-  return [usage ? { ...state, usage: mergeUsage(state.usage, usage) } : state, NO_EVENTS]
+  const usage = mapUsage(event.message?.usage, state.providerMetadataKey)
+  return [usage ? { ...state, usage: mergeUsage(state.usage, usage, state.providerMetadataKey) } : state, NO_EVENTS]
 }
 
 const onContentBlockStart = (
@@ -1204,14 +1211,16 @@ const onContentBlockStart = (
   if (block.type === "thinking" && block.thinking !== undefined) {
     const events: LLMEvent[] = []
     const id = `reasoning-${event.index ?? 0}`
-    const providerMetadata =
-      block.signature === undefined ? undefined : anthropicMetadata({ signature: block.signature })
-    const lifecycle = Lifecycle.reasoningStart(state.lifecycle, events, id, providerMetadata)
+    const metadata =
+      block.signature === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { signature: block.signature })
+    const lifecycle = Lifecycle.reasoningStart(state.lifecycle, events, id, metadata)
     return [
       {
         ...state,
         lifecycle: block.thinking
-          ? Lifecycle.reasoningDelta(lifecycle, events, id, block.thinking, providerMetadata)
+          ? Lifecycle.reasoningDelta(lifecycle, events, id, block.thinking, metadata)
           : lifecycle,
         reasoningSignatures:
           event.index === undefined || block.signature === undefined
@@ -1234,14 +1243,14 @@ const onContentBlockStart = (
           state.lifecycle,
           events,
           `reasoning-${event.index ?? 0}`,
-          anthropicMetadata({ redactedData: block.data }),
+          providerMetadata(state.providerMetadataKey, { redactedData: block.data }),
         ),
       },
       events,
     ]
   }
 
-  const result = serverToolResultEvent(block)
+  const result = serverToolResultEvent(block, state.providerMetadataKey)
   if (!result) return [state, NO_EVENTS]
   const events: LLMEvent[] = []
   return [{ ...state, lifecycle: Lifecycle.stepStart(state.lifecycle, events) }, [...events, result]]
@@ -1321,7 +1330,7 @@ const onContentBlockStop = Effect.fn("AnthropicMessages.onContentBlockStop")(fun
         Lifecycle.textEnd(state.lifecycle, events, `text-${event.index}`),
         events,
         `reasoning-${event.index}`,
-        signature === undefined ? undefined : anthropicMetadata({ signature }),
+        signature === undefined ? undefined : providerMetadata(state.providerMetadataKey, { signature }),
       )
   events.push(...resultEvents)
   const reasoningSignatures = { ...state.reasoningSignatures }
@@ -1333,7 +1342,7 @@ const onMessageDelta = (
   state: ParserState,
   event: AnthropicEvent & { readonly delta?: AnthropicStreamDelta },
 ): StepResult => {
-  const usage = mergeUsage(state.usage, mapUsage(event.usage))
+  const usage = mergeUsage(state.usage, mapUsage(event.usage, state.providerMetadataKey), state.providerMetadataKey)
   return [
     {
       ...state,
@@ -1346,7 +1355,7 @@ const onMessageDelta = (
         providerMetadata:
           event.delta?.stop_sequence === null || event.delta?.stop_sequence === undefined
             ? undefined
-            : anthropicMetadata({ stopSequence: event.delta.stop_sequence }),
+            : providerMetadata(state.providerMetadataKey, { stopSequence: event.delta.stop_sequence }),
       },
     },
     NO_EVENTS,
@@ -1472,7 +1481,8 @@ export const protocol = Protocol.make({
   },
   stream: {
     event: Protocol.jsonEvent(AnthropicEvent),
-    initial: () => ({
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
       tools: ToolStream.empty<number>(),
       reasoningSignatures: {},
       lifecycle: Lifecycle.initial(),

packages/ai/src/protocols/bedrock-converse.ts

@@ -258,19 +258,21 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
     tool: (name) => ({ tool: { name } }) as const,
   })
 
-const bedrockMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ bedrock: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
 
-const reasoningSignature = (part: ReasoningPart) => {
-  const bedrock = part.providerMetadata?.bedrock
+const reasoningSignature = (part: ReasoningPart, providerMetadataKey: string) => {
+  const metadata = part.providerMetadata?.[providerMetadataKey]
   return (
     part.encrypted ??
-    (ProviderShared.isRecord(bedrock) && typeof bedrock.signature === "string" ? bedrock.signature : undefined)
+    (ProviderShared.isRecord(metadata) && typeof metadata.signature === "string" ? metadata.signature : undefined)
   )
 }
 
-const reasoningRedactedData = (part: ReasoningPart) => {
-  const bedrock = part.providerMetadata?.bedrock
-  return ProviderShared.isRecord(bedrock) && typeof bedrock.redactedData === "string" ? bedrock.redactedData : undefined
+const reasoningRedactedData = (part: ReasoningPart, providerMetadataKey: string) => {
+  const metadata = part.providerMetadata?.[providerMetadataKey]
+  return ProviderShared.isRecord(metadata) && typeof metadata.redactedData === "string"
+    ? metadata.redactedData
+    : undefined
 }
 
 const lowerToolCall = (part: ToolCallPart): BedrockToolUseBlock => ({
@@ -318,6 +320,7 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
   breakpoints: BedrockCache.Breakpoints,
 ) {
   const messages: BedrockMessage[] = []
+  const providerMetadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
 
   for (const message of request.messages) {
     if (message.role === "system") {
@@ -365,8 +368,8 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
           continue
         }
         if (part.type === "reasoning") {
-          const signature = reasoningSignature(part)
-          const redactedData = reasoningRedactedData(part)
+          const signature = reasoningSignature(part, providerMetadataKey)
+          const redactedData = reasoningRedactedData(part, providerMetadataKey)
           if (signature === undefined && redactedData !== undefined) {
             content.push({ reasoningContent: { redactedContent: redactedData } })
             continue
@@ -466,7 +469,7 @@ const mapFinishReason = (reason: string): FinishReason => {
 
 // AWS reports inputTokens separately from cache reads and writes.
 // Bedrock does not break reasoning out of outputTokens for current models.
-const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
+const mapUsage = (usage: BedrockUsageSchema | undefined, providerMetadataKey: string): Usage | undefined => {
   if (!usage) return undefined
   const inputTokens = ProviderShared.sumTokens(
     usage.inputTokens,
@@ -480,11 +483,12 @@ const mapUsage = (usage: BedrockUsageSchema | undefined): Usage | undefined => {
     cacheReadInputTokens: usage.cacheReadInputTokens,
     cacheWriteInputTokens: usage.cacheWriteInputTokens,
     totalTokens: ProviderShared.totalTokens(inputTokens, usage.outputTokens, usage.totalTokens),
-    providerMetadata: { bedrock: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
   })
 }
 
 interface ParserState {
+  readonly providerMetadataKey: string
   readonly tools: ToolStream.State<number>
   // Bedrock splits the finish into `messageStop` (carries `stopReason`) and
   // `metadata` (carries usage). Hold the terminal event in state so `onHalt`
@@ -541,20 +545,14 @@ const step = (state: ParserState, event: BedrockEvent) =>
       const reasoning = event.contentBlockDelta.delta.reasoningContent
       const events: LLMEvent[] = []
       const redactedData = reasoning.redactedContent ?? reasoning.data
-      const providerMetadata = reasoning.signature
-        ? bedrockMetadata({ signature: reasoning.signature })
+      const metadata = reasoning.signature
+        ? providerMetadata(state.providerMetadataKey, { signature: reasoning.signature })
         : redactedData !== undefined
-          ? bedrockMetadata({ redactedData })
+          ? providerMetadata(state.providerMetadataKey, { redactedData })
           : undefined
       const lifecycle =
-        reasoning.text !== undefined || providerMetadata !== undefined
-          ? Lifecycle.reasoningDelta(
-              state.lifecycle,
-              events,
-              `reasoning-${index}`,
-              reasoning.text ?? "",
-              providerMetadata,
-            )
+        reasoning.text !== undefined || metadata !== undefined
+          ? Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text ?? "", metadata)
           : state.lifecycle
       return [
         {
@@ -596,7 +594,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
             events,
             `reasoning-${index}`,
             state.reasoningSignatures[index]
-              ? bedrockMetadata({ signature: state.reasoningSignatures[index] })
+              ? providerMetadata(state.providerMetadataKey, { signature: state.reasoningSignatures[index] })
               : undefined,
           )
       events.push(...resultEvents)
@@ -633,7 +631,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
     }
 
     if (event.metadata) {
-      const usage = mapUsage(event.metadata.usage) ?? state.pendingFinish?.usage
+      const usage = mapUsage(event.metadata.usage, state.providerMetadataKey) ?? state.pendingFinish?.usage
       return [
         {
           ...state,
@@ -698,7 +696,8 @@ export const protocol = Protocol.make({
   },
   stream: {
     event: BedrockEvent,
-    initial: () => ({
+    initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
       tools: ToolStream.empty<number>(),
       pendingFinish: undefined,
       hasToolCalls: false,

packages/ai/src/protocols/gemini.ts

@@ -229,6 +229,7 @@ type GeminiEvent = Schema.Schema.Type<typeof GeminiEvent>
 
 interface ParserState {
   readonly route: string
+  readonly providerMetadataKey: string
   readonly finishReason?: string
   readonly hasToolCalls: boolean
   readonly promptFeedback?: GeminiPromptFeedback
@@ -285,22 +286,23 @@ const lowerUserPart = Effect.fn("Gemini.lowerUserPart")(function* (part: TextPar
   return { inlineData: { mimeType: media.mime, data: media.base64 } }
 })
 
-const googleMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ google: metadata })
+const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
 
-const thoughtSignature = (providerMetadata: ProviderMetadata | undefined) => {
-  const google = providerMetadata?.google
-  return ProviderShared.isRecord(google) && typeof google.thoughtSignature === "string"
-    ? google.thoughtSignature
+const thoughtSignature = (metadata: ProviderMetadata | undefined, key: string) => {
+  const value = metadata?.[key]
+  return ProviderShared.isRecord(value) && typeof value.thoughtSignature === "string"
+    ? value.thoughtSignature
     : undefined
 }
 
-const lowerToolCall = (part: ToolCallPart, omitIds: boolean) => ({
+const lowerToolCall = (part: ToolCallPart, omitIds: boolean, metadataKey: string) => ({
   functionCall: { ...(omitIds ? {} : { id: part.id }), name: part.name, args: part.input },
-  thoughtSignature: thoughtSignature(part.providerMetadata),
+  thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
 })
 
 const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMRequest) {
   const contents: GeminiContent[] = []
+  const metadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
   const omitCallIds = omitsFunctionCallIds(request.model.id)
   const legacyToolMedia = routesLegacyToolMedia(request.model.id)
   let pendingMedia: GeminiInlineDataPart[] | undefined
@@ -342,15 +344,19 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
         if (!ProviderShared.supportsContent(part, ["text", "reasoning", "tool-call"]))
           return yield* ProviderShared.unsupportedContent("Gemini", "assistant", ["text", "reasoning", "tool-call"])
         if (part.type === "text") {
-          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata) })
+          parts.push({ text: part.text, thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey) })
           continue
         }
         if (part.type === "reasoning") {
-          parts.push({ text: part.text, thought: true, thoughtSignature: thoughtSignature(part.providerMetadata) })
+          parts.push({
+            text: part.text,
+            thought: true,
+            thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
+          })
           continue
         }
         if (part.type === "tool-call") {
-          const lowered = lowerToolCall(part, omitCallIds)
+          const lowered = lowerToolCall(part, omitCallIds, metadataKey)
           const signature = lowered.thoughtSignature
           parts.push({
             ...lowered,
@@ -498,7 +504,7 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
 // `cachedContentTokenCount` subset. `candidatesTokenCount` is *exclusive*
 // of `thoughtsTokenCount` — visible-only, not a total — so we sum the two
 // to produce the inclusive `outputTokens` the rest of the contract expects.
-const mapUsage = (usage: GeminiUsage | undefined) => {
+const mapUsage = (usage: GeminiUsage | undefined, metadataKey: string) => {
   if (!usage) return undefined
   // Explicit provider nulls decode as `null`; normalize to `undefined` so the
   // token arithmetic below treats them like absent counts.
@@ -519,7 +525,7 @@ const mapUsage = (usage: GeminiUsage | undefined) => {
     cacheReadInputTokens: cached,
     reasoningTokens: thoughts,
     totalTokens: ProviderShared.totalTokens(promptTokens, outputTokens, usage.totalTokenCount ?? undefined),
-    providerMetadata: { google: usage },
+    providerMetadata: providerMetadata(metadataKey, usage),
   })
 }
 
@@ -567,14 +573,14 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
       lifecycle,
       events,
       "reasoning-0",
-      googleMetadata({ thoughtSignature: state.reasoningSignature }),
+      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.reasoningSignature }),
     )
   if (state.textSignature !== undefined)
     lifecycle = Lifecycle.textEnd(
       lifecycle,
       events,
       "text-0",
-      googleMetadata({ thoughtSignature: state.textSignature }),
+      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.textSignature }),
     )
   Lifecycle.finish(lifecycle, events, {
     reason: {
@@ -584,7 +590,9 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {
     },
     usage: state.usage,
     providerMetadata:
-      state.promptFeedback === undefined ? undefined : googleMetadata({ promptFeedback: state.promptFeedback }),
+      state.promptFeedback === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { promptFeedback: state.promptFeedback }),
   })
   return events
 }
@@ -593,7 +601,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
   const nextState = {
     ...state,
     promptFeedback: event.promptFeedback ?? state.promptFeedback,
-    usage: event.usageMetadata ? (mapUsage(event.usageMetadata) ?? state.usage) : state.usage,
+    usage: event.usageMetadata
+      ? (mapUsage(event.usageMetadata, state.providerMetadataKey) ?? state.usage)
+      : state.usage,
   }
   const candidate = event.candidates?.[0]
   if (!candidate?.content)
@@ -637,7 +647,7 @@ const step = (state: ParserState, event: GeminiEvent) => {
           events,
           "reasoning-0",
           part.text,
-          signature ? googleMetadata({ thoughtSignature: signature }) : undefined,
+          signature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: signature }) : undefined,
         )
         continue
       }
@@ -645,14 +655,16 @@ const step = (state: ParserState, event: GeminiEvent) => {
         lifecycle,
         events,
         "reasoning-0",
-        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
+        reasoningSignature
+          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+          : undefined,
       )
       lifecycle = Lifecycle.textDelta(
         lifecycle,
         events,
         "text-0",
         part.text,
-        textSignature ? googleMetadata({ thoughtSignature: textSignature }) : undefined,
+        textSignature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature }) : undefined,
       )
       textSignature = undefined
       continue
@@ -672,7 +684,9 @@ const step = (state: ParserState, event: GeminiEvent) => {
         lifecycle,
         events,
         "reasoning-0",
-        reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
+        reasoningSignature
+          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+          : undefined,
       )
       lifecycle = Lifecycle.stepStart(lifecycle, events)
       events.push(
@@ -681,7 +695,7 @@ const step = (state: ParserState, event: GeminiEvent) => {
           name: part.functionCall.name,
           input,
           providerMetadata: part.thoughtSignature
-            ? googleMetadata({ thoughtSignature: part.thoughtSignature })
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: part.thoughtSignature })
             : undefined,
         }),
       )
@@ -720,6 +734,7 @@ export const protocol = Protocol.make({
     event: Protocol.jsonEvent(GeminiEvent),
     initial: (request) => ({
       route: `${request.model.provider}/${request.model.route.id}`,
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
       hasToolCalls: false,
       lifecycle: Lifecycle.initial(),
     }),

packages/ai/src/protocols/openai-chat.ts

@@ -253,6 +253,7 @@ interface PendingToolDelta {
 }
 
 export interface ParserState {
+  readonly providerMetadataKey: string
   readonly tools: ToolStream.State<number>
   readonly pendingTools: Partial<Record<number, PendingToolDelta>>
   readonly toolCallEvents: ReadonlyArray<LLMEvent>
@@ -324,17 +325,18 @@ const lowerMedia = Effect.fn("OpenAIChat.lowerMedia")(function* (part: MediaPart
 const openAICompatibleReasoningContent = (native: unknown) =>
   isRecord(native) && typeof native.reasoning_content === "string" ? native.reasoning_content : undefined
 
-const reasoningField = (part: ReasoningPart) => {
-  const field = part.providerMetadata?.openai?.reasoningField
+const reasoningField = (part: ReasoningPart, providerMetadataKey: string) => {
+  const field = part.providerMetadata?.[providerMetadataKey]?.reasoningField
   return typeof field === "string" ? field : undefined
 }
 
-const reasoningDetails = (parts: ReadonlyArray<ReasoningPart>, native: unknown) => {
+const reasoningDetails = (parts: ReadonlyArray<ReasoningPart>, native: unknown, providerMetadataKey: string) => {
   const observed = parts.flatMap((part) => {
-    const details = part.providerMetadata?.openai?.reasoningDetails
+    const details = part.providerMetadata?.[providerMetadataKey]?.reasoningDetails
     return Array.isArray(details) ? details : []
   })
-  if (parts.some((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))) return observed
+  if (parts.some((part) => Array.isArray(part.providerMetadata?.[providerMetadataKey]?.reasoningDetails)))
+    return observed
   if (isRecord(native) && Array.isArray(native.reasoning_details)) return native.reasoning_details
 }
 
@@ -366,7 +368,7 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
   message: OpenAIChatRequestMessage,
   configuredField: string | undefined,
   requireReasoning: boolean,
-  options: LoweringOptions,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
 ) {
   const content: TextPart[] = []
   const reasoning: ReasoningPart[] = []
@@ -388,10 +390,14 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
     }
   }
   const text = reasoning.map((part) => part.text).join("")
-  const details = reasoningDetails(reasoning, message.native?.openaiCompatible)
-  const observedField = reasoning.map(reasoningField).find((value) => value !== undefined)
+  const details = reasoningDetails(reasoning, message.native?.openaiCompatible, options.providerMetadataKey)
+  const observedField = reasoning
+    .map((part) => reasoningField(part, options.providerMetadataKey))
+    .find((value) => value !== undefined)
   const nativeReasoning = openAICompatibleReasoningContent(message.native?.openaiCompatible)
-  const fullyStructured = reasoning.every((part) => Array.isArray(part.providerMetadata?.openai?.reasoningDetails))
+  const fullyStructured = reasoning.every((part) =>
+    Array.isArray(part.providerMetadata?.[options.providerMetadataKey]?.reasoningDetails),
+  )
   const field = (() => {
     if (configuredField !== undefined && (requireReasoning || reasoning.length > 0 || nativeReasoning !== undefined))
       return configuredField
@@ -459,7 +465,7 @@ const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
   message: OpenAIChatRequestMessage,
   reasoningField: string | undefined,
   requireReasoning: boolean,
-  options: LoweringOptions,
+  options: LoweringOptions & { readonly providerMetadataKey: string },
 ) {
   if (message.role === "user") return [yield* lowerUserMessage(message, options)]
   if (message.role === "assistant")
@@ -495,6 +501,7 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
   const mistral = ["mistral", "devstral", "codestral", "pixtral", "mixtral"].some((family) => modelID.includes(family))
   const lowering = {
     ...options,
+    providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
     toolCallID: (id: string) => {
       if (mistral)
         return id
@@ -820,7 +827,7 @@ const mapFinishReason = Effect.fn("OpenAIChat.mapFinishReason")(function* (event
 // Providers differ on cache-hit location: OpenAI uses
 // `prompt_tokens_details.cached_tokens`, DeepSeek uses
 // `prompt_cache_hit_tokens`, and Zai uses top-level `cached_tokens`.
-const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
+const mapUsage = (usage: OpenAIChatEvent["usage"], providerMetadataKey: string): Usage | undefined => {
   if (!usage) return undefined
   const input = usage.prompt_tokens ?? undefined
   const output = usage.completion_tokens ?? undefined
@@ -839,7 +846,7 @@ const mapUsage = (usage: OpenAIChatEvent["usage"]): Usage | undefined => {
     cacheWriteInputTokens: cacheWrite,
     reasoningTokens: reasoning,
     totalTokens: ProviderShared.totalTokens(input, output, usage.total_tokens ?? undefined),
-    providerMetadata: { openai: usage },
+    providerMetadata: { [providerMetadataKey]: usage },
   })
 }
 
@@ -913,8 +920,12 @@ const conflictingReasoningTextDetails = (previous: Record<string, unknown>, curr
 const conflictingDetailValue = (previous: unknown, current: unknown) =>
   previous !== undefined && previous !== null && current !== undefined && current !== null && previous !== current
 
-const reasoningMetadata = (field: ParserState["reasoningField"], details?: ReadonlyArray<unknown>) => ({
-  openai: {
+const reasoningMetadata = (
+  providerMetadataKey: string,
+  field: ParserState["reasoningField"],
+  details?: ReadonlyArray<unknown>,
+) => ({
+  [providerMetadataKey]: {
     ...(field ? { reasoningField: field } : {}),
     ...(details ? { reasoningDetails: details } : {}),
   },
@@ -941,7 +952,10 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
     // Moonshot (and a few other OpenAI-compatible providers) attach usage to
     // `choice.usage` instead of the top-level `usage` field.
     const choiceUsage = (choice as unknown as { usage?: OpenAIChatEvent["usage"] })?.usage
-    const usage = mapUsage(event.usage) ?? (choiceUsage ? mapUsage(choiceUsage) : undefined) ?? state.usage
+    const usage =
+      mapUsage(event.usage, state.providerMetadataKey) ??
+      (choiceUsage ? mapUsage(choiceUsage, state.providerMetadataKey) : undefined) ??
+      state.usage
     const rawFinishReason = choice?.finish_reason
     const finishReason = rawFinishReason
       ? {
@@ -979,7 +993,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
     const detailDelta = Array.isArray(delta?.reasoning_details) ? delta.reasoning_details : undefined
     if (detailDelta !== undefined) appendReasoningDetails(state.reasoningDetails, detailDelta)
     const reasoningDetailsObserved = state.reasoningDetailsObserved || detailDelta !== undefined
-    const deltaMetadata = reasoningMetadata(reasoningField)
+    const deltaMetadata = reasoningMetadata(state.providerMetadataKey, reasoningField)
     const text = detailDelta?.length ? (detailText(detailDelta) ?? reasoning?.text) : reasoning?.text
     if (text !== undefined) lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", text, deltaMetadata)
     else if (
@@ -995,7 +1009,11 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
         lifecycle,
         events,
         "reasoning-0",
-        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
+        reasoningMetadata(
+          state.providerMetadataKey,
+          reasoningField,
+          reasoningDetailsObserved ? state.reasoningDetails : undefined,
+        ),
       )
       lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)
     }
@@ -1005,7 +1023,11 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
         lifecycle,
         events,
         "reasoning-0",
-        reasoningMetadata(reasoningField, reasoningDetailsObserved ? state.reasoningDetails : undefined),
+        reasoningMetadata(
+          state.providerMetadataKey,
+          reasoningField,
+          reasoningDetailsObserved ? state.reasoningDetails : undefined,
+        ),
       )
       lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.refusal)
     }
@@ -1066,6 +1088,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
 
     return [
       {
+        providerMetadataKey: state.providerMetadataKey,
         tools: finished?.tools ?? tools,
         pendingTools,
         toolCallEvents: finished?.events ?? state.toolCallEvents,
@@ -1109,12 +1132,18 @@ const finishEvents = Effect.fn("OpenAIChat.finishEvents")(function* (state: Pars
       }
     : { normalized: hasToolCalls ? ("tool-calls" as const) : ("stop" as const) }
   const metadata = reasoningMetadata(
+    state.providerMetadataKey,
     state.reasoningField,
     state.reasoningDetailsObserved ? state.reasoningDetails : undefined,
   )
   const started =
     state.reasoningDetailsObserved && !state.reasoningEmitted
-      ? Lifecycle.reasoningStart(state.lifecycle, events, "reasoning-0", reasoningMetadata(state.reasoningField))
+      ? Lifecycle.reasoningStart(
+          state.lifecycle,
+          events,
+          "reasoning-0",
+          reasoningMetadata(state.providerMetadataKey, state.reasoningField),
+        )
       : state.lifecycle
   const ended = Lifecycle.reasoningEnd(started, events, "reasoning-0", metadata)
   const lifecycle = toolCallEvents.length ? Lifecycle.stepStart(ended, events) : ended
@@ -1141,6 +1170,7 @@ export const protocol = Protocol.make({
   stream: {
     event: Protocol.jsonEvent(OpenAIChatEvent),
     initial: (request) => ({
+      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
       tools: ToolStream.empty<number>(),
       pendingTools: {},
       toolCallEvents: [],

packages/ai/src/providers/amazon-bedrock-mantle.ts

@@ -23,13 +23,14 @@ export interface Settings extends ProviderPackage.Settings {
   readonly baseURL?: string
   readonly credentials?: Credentials
   readonly region?: string
+  readonly topP?: number
   readonly providerOptions?: OpenAIProviderOptionsInput
 }
 
 const responsesRoute = Route.make({
   id: "bedrock-mantle-responses",
   provider: id,
-  providerMetadataKey: OpenAIResponses.route.providerMetadataKey,
+  providerMetadataKey: "mantle",
   protocol: OpenAIResponses.protocol,
   endpoint: OpenAIResponses.route.endpoint,
   auth: OpenAIResponses.route.auth,
@@ -40,6 +41,7 @@ const responsesRoute = Route.make({
 const chatRoute = OpenAIChat.route.with({
   id: "bedrock-mantle-chat",
   provider: id,
+  providerMetadataKey: "mantle",
 })
 
 export const routes = [responsesRoute, chatRoute]
@@ -76,7 +78,7 @@ export const configure = (input: Config = {}) => {
 
   return {
     id,
-    model: chat,
+    model: responses,
     chat,
     responses,
     configure,
@@ -94,6 +96,7 @@ const config = (settings: Settings): Config => {
     apiKey: settings.auth === "sigv4" ? undefined : settings.apiKey,
     baseURL: settings.baseURL,
     credentials: settings.credentials,
+    generation: settings.topP === undefined ? undefined : { topP: settings.topP },
     headers: settings.headers === undefined ? undefined : { ...settings.headers },
     http: settings.body === undefined ? undefined : { body: { ...settings.body } },
     providerOptions: settings.providerOptions,
@@ -109,4 +112,4 @@ export const responsesModel: ProviderPackage.Definition<Settings, OpenAIProvider
   modelID,
   settings,
 ) => configure(config(settings)).responses(modelID)
-export const model = chatModel
+export const model = responsesModel

packages/ai/src/providers/amazon-bedrock.ts

@@ -35,6 +35,7 @@ const configuredRoute = (input: Config) => {
   return BedrockConverse.route.with({
     ...rest,
     provider: id,
+    providerMetadataKey: "bedrock",
     endpoint: { baseURL: baseURL ?? bedrockBaseURL(resolvedRegion) },
     auth: apiKey === undefined ? BedrockConverse.sigV4Auth(credentials) : Auth.bearer(apiKey),
   })

packages/ai/src/providers/amazon-bedrock/mantle.ts

@@ -1,2 +1,2 @@
-export { chatModel as model } from "../amazon-bedrock-mantle.js"
+export { responsesModel as model } from "../amazon-bedrock-mantle.js"
 export type { Settings } from "../amazon-bedrock-mantle.js"

packages/ai/src/providers/google-vertex-chat.ts

@@ -27,6 +27,7 @@ export interface Settings extends ProviderPackage.Settings {
 const route = OpenAICompatibleChat.route.with({
   id: "google-vertex-chat",
   provider: id,
+  providerMetadataKey: "vertex",
 })
 
 export const routes = [route]

packages/ai/src/providers/google-vertex-responses.ts

@@ -27,6 +27,7 @@ export interface Settings extends ProviderPackage.Settings {
 const route = OpenAICompatibleResponses.route.with({
   id: "google-vertex-responses",
   provider: id,
+  providerMetadataKey: "vertex",
   providerOptions: { store: false },
 })
 

packages/ai/src/providers/google-vertex.ts

@@ -68,7 +68,7 @@ const protocol = {
 const route = Route.make({
   id: "google-vertex-gemini",
   provider: id,
-  providerMetadataKey: "google",
+  providerMetadataKey: "vertex",
   protocol,
   endpoint: Endpoint.path(({ request }) => {
     const model = String(request.model.id)

packages/ai/src/providers/openrouter.ts

@@ -164,6 +164,7 @@ const bodyOptions = (input: unknown) => {
 export const route = Route.make({
   id: ADAPTER,
   provider: profile.provider,
+  providerMetadataKey: "openrouter",
   protocol,
   endpoint: Endpoint.path("/chat/completions", { baseURL: profile.baseURL }),
   framing: Framing.sse,

packages/ai/src/route/client.ts

@@ -89,6 +89,7 @@ export interface RouteDefaultsInput {
 export interface RoutePatch<Body, Prepared> extends RouteDefaultsInput {
   readonly id?: string
   readonly provider?: string | ProviderID
+  readonly providerMetadataKey?: string
   readonly auth?: Auth.Definition
   readonly transport?: Transport<Body, Prepared, unknown>
   readonly endpoint?: EndpointPatch<Body>
@@ -289,11 +290,16 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
       defaults: routeInput.defaults ?? {},
       body: protocol.body,
       with: (patch: RoutePatch<Body, Prepared>) => {
-        const { id, provider, auth, transport, endpoint, ...defaults } = patch
+        const { id, provider, providerMetadataKey, auth, transport, endpoint, ...defaults } = patch
         return build({
           ...routeInput,
           id: id ?? routeInput.id,
           provider: provider ?? routeInput.provider,
+          providerMetadataKey:
+            providerMetadataKey ??
+            (provider !== undefined && String(provider) !== String(routeInput.provider)
+              ? String(provider)
+              : routeInput.providerMetadataKey),
           auth: auth ?? routeInput.auth,
           endpoint: endpoint ? Endpoint.merge(routeInput.endpoint, endpoint) : routeInput.endpoint,
           transport: (transport as Transport<Body, Prepared, Frame> | undefined) ?? routeInput.transport,

packages/ai/src/route/executor.ts

@@ -40,17 +40,6 @@ const headerDetails = (headers: Headers.Headers) =>
 const normalizedHeaders = (headers: Headers.Headers) =>
   Object.fromEntries(Object.entries(headers).map(([key, value]) => [key.toLowerCase(), value]))
 
-const requestId = (headers: Record<string, string>) => {
-  return (
-    headers["x-request-id"] ??
-    headers["request-id"] ??
-    headers["x-amzn-requestid"] ??
-    headers["x-amz-request-id"] ??
-    headers["x-goog-request-id"] ??
-    headers["cf-ray"]
-  )
-}
-
 const retryAfterMs = (headers: Record<string, string>) => {
   const millis = Number(headers["retry-after-ms"])
   if (Number.isFinite(millis)) return Math.max(0, millis)
@@ -147,14 +136,12 @@ const responseHttp = (input: {
   readonly request: HttpClientRequest.HttpClientRequest
   readonly response: HttpClientResponse.HttpClientResponse
   readonly body: ReturnType<typeof responseBody>
-  readonly requestId?: string | undefined
   readonly rateLimit?: HttpRateLimitDetails | undefined
 }) =>
   new HttpContext({
     request: requestDetails(input.request),
     response: responseDetails(input.response),
     ...input.body,
-    requestId: input.requestId,
     rateLimit: input.rateLimit,
   })
 
@@ -179,7 +166,6 @@ const statusError =
             request,
             response,
             body: details,
-            requestId: requestId(headers),
             rateLimit,
           }),
         }),
@@ -216,7 +202,6 @@ export const classifyHttpFailure = (input: {
           ? undefined
           : new HttpResponseDetails({ status: input.status, headers: headerDetails(Headers.fromInput(headers)) }),
       ...details,
-      requestId: requestId(headers),
       rateLimit,
     }),
   })

packages/ai/src/schema/errors.ts

@@ -29,7 +29,6 @@ export class HttpContext extends Schema.Class<HttpContext>("AI.HttpContext")({
   response: Schema.optional(HttpResponseDetails),
   body: Schema.optional(Schema.String),
   bodyTruncated: Schema.optional(Schema.Boolean),
-  requestId: Schema.optional(Schema.String),
   rateLimit: Schema.optional(HttpRateLimitDetails),
 }) {}
 

packages/ai/test/executor.test.ts

@@ -327,7 +327,6 @@ describe("RequestExecutor", () => {
           retryAfterMs: 0,
           rateLimit: { retryAfterMs: 0 },
           http: {
-            requestId: "req_123",
             request: {
               method: "POST",
               url: "https://provider.test/v1/chat?api_key=secret&key=secret&debug=1",

packages/ai/test/provider/anthropic-messages.test.ts

@@ -2,7 +2,7 @@ import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { CacheHint, LLM, AIError, LLMRequest, Message, ToolCallPart, ToolDefinition, Usage } from "../../src/index.js"
-import { Auth, LLMClient } from "../../src/route.js"
+import { Auth, Endpoint, LLMClient, Route } from "../../src/route.js"
 import { compileRequest } from "../../src/route/client.js"
 import * as AnthropicMessages from "../../src/protocols/anthropic-messages.js"
 import { GoogleVertexMessages } from "../../src/providers.js"
@@ -810,6 +810,99 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("round-trips compatible provider metadata in its own namespace", () =>
+    Effect.gen(function* () {
+      const compatible = Route.make({
+        id: "custom-anthropic-messages",
+        provider: "custom-anthropic",
+        protocol: AnthropicMessages.protocol,
+        endpoint: Endpoint.path("/messages", { baseURL: "https://compatible.test/v1" }),
+        auth: Auth.header("x-api-key", "test"),
+        framing: AnthropicMessages.framing,
+      }).model({ id: "custom-model" })
+      const result = [
+        {
+          type: "web_search_result",
+          url: "https://example.com",
+          citations: [{ type: "web_search_result_location", cited_text: "Example" }],
+        },
+      ]
+      const response = yield* LLMClient.generate(LLM.request({ model: compatible, prompt: "Search." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5, custom_start: true } } },
+              { type: "content_block_start", index: 0, content_block: { type: "thinking", thinking: "Thinking." } },
+              { type: "content_block_delta", index: 0, delta: { type: "signature_delta", signature: "custom_sig" } },
+              { type: "content_block_stop", index: 0 },
+              {
+                type: "content_block_start",
+                index: 1,
+                content_block: { type: "redacted_thinking", data: "custom_redacted" },
+              },
+              { type: "content_block_stop", index: 1 },
+              {
+                type: "content_block_start",
+                index: 2,
+                content_block: {
+                  type: "server_tool_use",
+                  id: "custom_tool",
+                  name: "web_search",
+                  input: { query: "example" },
+                },
+              },
+              { type: "content_block_stop", index: 2 },
+              {
+                type: "content_block_start",
+                index: 3,
+                content_block: { type: "web_search_tool_result", tool_use_id: "custom_tool", content: result },
+              },
+              { type: "content_block_stop", index: 3 },
+              {
+                type: "message_delta",
+                delta: { stop_reason: "end_turn", stop_sequence: "custom_stop" },
+                usage: { output_tokens: 2, custom_terminal: true },
+              },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toMatchObject([
+        { type: "reasoning", text: "Thinking.", providerMetadata: { "custom-anthropic": { signature: "custom_sig" } } },
+        { type: "reasoning", text: "", providerMetadata: { "custom-anthropic": { redactedData: "custom_redacted" } } },
+        { type: "tool-call", id: "custom_tool", providerExecuted: true },
+        {
+          type: "tool-result",
+          providerExecuted: true,
+          providerMetadata: { "custom-anthropic": { blockType: "web_search_tool_result", result } },
+        },
+      ])
+      expect(response.usage?.providerMetadata).toEqual({
+        "custom-anthropic": { input_tokens: 5, custom_start: true, output_tokens: 2, custom_terminal: true },
+      })
+      expect(response.events.at(-1)).toMatchObject({
+        providerMetadata: { "custom-anthropic": { stopSequence: "custom_stop" } },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model: compatible, messages: [response.message], cache: "none" }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            { type: "thinking", thinking: "Thinking.", signature: "custom_sig" },
+            { type: "redacted_thinking", data: "custom_redacted" },
+            { type: "server_tool_use", id: "custom_tool", name: "web_search", input: { query: "example" } },
+            { type: "web_search_tool_result", tool_use_id: "custom_tool", content: result },
+          ],
+        },
+      ])
+    }),
+  )
+
   it.effect("parses text, reasoning, and usage stream fixtures", () =>
     Effect.gen(function* () {
       const body = sseEvents(

packages/ai/test/provider/bedrock-converse.test.ts

@@ -569,6 +569,57 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("round-trips reassigned provider reasoning and usage metadata in its own namespace", () =>
+    Effect.gen(function* () {
+      const compatible = model.route.with({ provider: "custom-bedrock" }).model({ id: model.id })
+      const redactedData = "cmVkYWN0ZWQtdGhpbmtpbmc="
+      const response = yield* LLMClient.generate(LLMRequest.update(baseRequest, { model: compatible })).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["messageStart", { role: "assistant" }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { text: "Let me think." } } }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { signature: "custom_sig" } } }],
+              ["contentBlockStop", { contentBlockIndex: 0 }],
+              [
+                "contentBlockDelta",
+                { contentBlockIndex: 1, delta: { reasoningContent: { redactedContent: redactedData } } },
+              ],
+              ["contentBlockStop", { contentBlockIndex: 1 }],
+              ["messageStop", { stopReason: "end_turn" }],
+              ["metadata", { usage: { inputTokens: 5, outputTokens: 2, totalTokens: 7 } }],
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Let me think.",
+          providerMetadata: { "custom-bedrock": { signature: "custom_sig" } },
+        },
+        { type: "reasoning", text: "", providerMetadata: { "custom-bedrock": { redactedData } } },
+      ])
+      expect(response.usage?.providerMetadata).toEqual({
+        "custom-bedrock": { inputTokens: 5, outputTokens: 2, totalTokens: 7 },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model: compatible, messages: [response.message], cache: "none" }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            { reasoningContent: { reasoningText: { text: "Let me think.", signature: "custom_sig" } } },
+            { reasoningContent: { redactedContent: redactedData } },
+          ],
+        },
+      ])
+    }),
+  )
+
   it.effect("preserves reasoning signatures when contentBlockStop is missing", () =>
     Effect.gen(function* () {
       const response = yield* LLMClient.generate(baseRequest).pipe(

packages/ai/test/provider/bedrock-mantle.test.ts

@@ -3,6 +3,7 @@ import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, Message } from "../../src/index.js"
 import { AmazonBedrockMantle } from "../../src/providers.js"
+import { model } from "../../src/providers/amazon-bedrock/mantle.js"
 import { OpenAIResponses } from "../../src/protocols/openai-responses.js"
 import { compileRequest, LLMClient } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
@@ -17,13 +18,16 @@ const credentials = {
 }
 
 describe("Amazon Bedrock Mantle provider", () => {
-  it.effect("uses Chat by default and exposes Responses", () =>
+  it.effect("uses Responses by default and exposes Chat explicitly", () =>
     Effect.gen(function* () {
       const provider = AmazonBedrockMantle.configure({ credentials })
-      expect(provider.responses("openai.gpt-oss-120b").route.transport).toBe(OpenAIResponses.httpTransport)
-      const chat = yield* compileRequest(LLM.request({ model: provider.model("openai.gpt-oss-120b"), prompt: "Hi" }))
+      expect(provider.model).toBe(provider.responses)
+      expect(AmazonBedrockMantle.model).toBe(AmazonBedrockMantle.responsesModel)
+      expect(model).toBe(AmazonBedrockMantle.responsesModel)
+      expect(provider.model("openai.gpt-oss-120b").route.transport).toBe(OpenAIResponses.httpTransport)
+      const chat = yield* compileRequest(LLM.request({ model: provider.chat("openai.gpt-oss-120b"), prompt: "Hi" }))
       const responses = yield* compileRequest(
-        LLM.request({ model: provider.responses("openai.gpt-oss-120b"), prompt: "Hi" }),
+        LLM.request({ model: provider.model("openai.gpt-oss-120b"), prompt: "Hi" }),
       )
 
       expect(chat).toMatchObject({
@@ -36,6 +40,23 @@ describe("Amazon Bedrock Mantle provider", () => {
         protocol: "openai-responses",
         body: { model: "openai.gpt-oss-120b", store: false },
       })
+      expect(provider.model("openai.gpt-oss-120b").route.providerMetadataKey).toBe("mantle")
+      expect(provider.chat("openai.gpt-oss-120b").route.providerMetadataKey).toBe("mantle")
+    }),
+  )
+
+  it.effect("preserves configured top-p generation defaults for Chat and Responses", () =>
+    Effect.gen(function* () {
+      const settings = { apiKey: "test-key", topP: 0.8 }
+      const chat = yield* compileRequest(
+        LLM.request({ model: AmazonBedrockMantle.chatModel("openai.gpt-oss-safeguard-20b", settings), prompt: "Hi" }),
+      )
+      const responses = yield* compileRequest(
+        LLM.request({ model: AmazonBedrockMantle.responsesModel("openai.gpt-oss-120b", settings), prompt: "Hi" }),
+      )
+
+      expect(chat.body.top_p).toBe(0.8)
+      expect(responses.body.top_p).toBe(0.8)
     }),
   )
 
@@ -107,6 +128,9 @@ describe("Amazon Bedrock Mantle provider", () => {
         LLM.request({ model, messages: [response.message, Message.user("Continue.")] }),
       )
 
+      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
+        mantle: { itemId: "msg_95d4d0af4350432a", reasoningEncryptedContent: "mantle-state" },
+      })
       expect(prepared.body.input).toEqual([
         {
           type: "reasoning",

packages/ai/test/provider/cloudflare.test.ts

@@ -126,7 +126,7 @@ describe("Cloudflare", () => {
       expect(response.reasoning).toBe("Thinking")
       expect(response.events.filter(LLMEvent.is.reasoningDelta)).toHaveLength(2)
       expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
-        openai: { reasoningField: "reasoning", reasoningDetails: merged },
+        "cloudflare-ai-gateway": { reasoningField: "reasoning", reasoningDetails: merged },
       })
 
       const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))

packages/ai/test/provider/google-vertex.test.ts

@@ -6,7 +6,7 @@ import { GoogleVertex, GoogleVertexChat, GoogleVertexMessages, GoogleVertexRespo
 import { LLMClient } from "../../src/route.js"
 import { compileRequest } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
-import { dynamicResponse } from "../lib/http.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
 import { deltaChunk, finishChunk } from "../lib/openai-chunks.js"
 import { sseEvents } from "../lib/sse.js"
 
@@ -89,7 +89,7 @@ describe("Google Vertex providers", () => {
                 id: "call_1",
                 name: "lookup",
                 input: { query: "weather" },
-                providerMetadata: { google: { functionCallId: "provider_call_1" } },
+                providerMetadata: { vertex: { functionCallId: "provider_call_1" } },
               }),
             ]),
             Message.tool({
@@ -97,7 +97,7 @@ describe("Google Vertex providers", () => {
               name: "lookup",
               result: "sunny",
               resultType: "text",
-              providerMetadata: { google: { functionCallId: "provider_call_1" } },
+              providerMetadata: { vertex: { functionCallId: "provider_call_1" } },
             }),
           ],
         }),
@@ -122,6 +122,91 @@ describe("Google Vertex providers", () => {
     }),
   )
 
+  it.effect("round-trips Vertex Gemini metadata through signed content, tool calls, and usage", () =>
+    Effect.gen(function* () {
+      const model = GoogleVertex.configure({
+        accessToken: "vertex-token",
+        project: "vertex-project",
+      }).model("gemini-3.5-flash")
+      const response = yield* LLMClient.generate(LLM.request({ model, prompt: "Check the weather." })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents({
+              candidates: [
+                {
+                  content: {
+                    role: "model",
+                    parts: [
+                      { text: "Thinking.", thought: true, thoughtSignature: "reasoning_sig" },
+                      { text: "Checking.", thoughtSignature: "text_sig" },
+                      {
+                        functionCall: { id: "provider_call_1", name: "lookup", args: { query: "weather" } },
+                        thoughtSignature: "tool_sig",
+                      },
+                    ],
+                  },
+                  finishReason: "STOP",
+                },
+              ],
+              promptFeedback: { blockReasonMessage: "Reviewed" },
+              usageMetadata: { promptTokenCount: 5, candidatesTokenCount: 2, thoughtsTokenCount: 1 },
+            }),
+          ),
+        ),
+      )
+      const reasoning = response.events.find((event) => event.type === "reasoning-end")
+      const text = response.events.find((event) => event.type === "text-delta")
+      const toolCall = response.toolCalls[0]
+
+      expect(reasoning?.providerMetadata).toEqual({ vertex: { thoughtSignature: "reasoning_sig" } })
+      expect(text?.providerMetadata).toEqual({ vertex: { thoughtSignature: "text_sig" } })
+      expect(toolCall).toMatchObject({
+        id: "provider_call_1",
+        providerMetadata: { vertex: { thoughtSignature: "tool_sig" } },
+      })
+      expect(response.usage?.providerMetadata).toEqual({
+        vertex: { promptTokenCount: 5, candidatesTokenCount: 2, thoughtsTokenCount: 1 },
+      })
+      expect(response.events.at(-1)?.providerMetadata).toEqual({
+        vertex: { promptFeedback: { blockReasonMessage: "Reviewed" } },
+      })
+
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              { type: "reasoning", text: "Thinking.", providerMetadata: reasoning?.providerMetadata },
+              { type: "text", text: "Checking.", providerMetadata: text?.providerMetadata },
+              ToolCallPart.make({
+                id: toolCall.id,
+                name: toolCall.name,
+                input: toolCall.input,
+                providerMetadata: toolCall.providerMetadata,
+              }),
+            ]),
+            Message.tool({ id: toolCall.id, name: toolCall.name, result: "sunny", resultType: "text" }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        {
+          role: "model",
+          parts: [
+            { text: "Thinking.", thought: true, thoughtSignature: "reasoning_sig" },
+            { text: "Checking.", thoughtSignature: "text_sig" },
+            { functionCall: { name: "lookup", args: { query: "weather" } }, thoughtSignature: "tool_sig" },
+          ],
+        },
+        {
+          role: "user",
+          parts: [{ functionResponse: { name: "lookup", response: { name: "lookup", content: "sunny" } } }],
+        },
+      ])
+    }),
+  )
+
   it.effect("projects Anthropic Messages onto the Vertex raw-predict API", () =>
     Effect.gen(function* () {
       const model = GoogleVertexMessages.configure({

packages/ai/test/provider/native-providers.test.ts

@@ -2,13 +2,80 @@ import { describe, expect } from "bun:test"
 import { ConfigProvider, Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, Message, ToolDefinition } from "../../src/index.js"
-import { Cerebras, DeepInfra, Groq, TogetherAI } from "../../src/providers/index.js"
+import {
+  AmazonBedrock,
+  AmazonBedrockMantle,
+  Anthropic,
+  AnthropicCompatible,
+  Azure,
+  Cerebras,
+  CloudflareAIGateway,
+  CloudflareWorkersAI,
+  DeepInfra,
+  Google,
+  GoogleVertex,
+  GoogleVertexChat,
+  GoogleVertexMessages,
+  GoogleVertexResponses,
+  Groq,
+  OpenAI,
+  OpenAICompatible,
+  OpenAICompatibleResponses,
+  OpenRouter,
+  TogetherAI,
+  XAI,
+} from "../../src/providers/index.js"
 import { compileRequest } from "../../src/route/client.js"
 import { it } from "../lib/effect.js"
 import { dynamicResponse } from "../lib/http.js"
 import { sseEvents } from "../lib/sse.js"
 
 describe("native OpenAI-compatible providers", () => {
+  it.effect("assigns provider-owned metadata namespaces across native routes", () =>
+    Effect.gen(function* () {
+      const vertex = { project: "project", accessToken: "token" }
+      const providers = [
+        [OpenAI.configure({ apiKey: "test" }).chat("model"), "openai"],
+        [OpenAI.configure({ apiKey: "test" }).responses("model"), "openai"],
+        [Azure.configure({ resourceName: "resource", apiKey: "test" }).chat("model"), "azure"],
+        [Azure.configure({ resourceName: "resource", apiKey: "test" }).responses("model"), "azure"],
+        [AmazonBedrock.configure({ apiKey: "test" }).model("model"), "bedrock"],
+        [AmazonBedrockMantle.configure({ apiKey: "test" }).chat("model"), "mantle"],
+        [AmazonBedrockMantle.configure({ apiKey: "test" }).responses("model"), "mantle"],
+        [Google.configure({ apiKey: "test" }).model("model"), "google"],
+        [GoogleVertex.configure(vertex).model("model"), "vertex"],
+        [GoogleVertexChat.configure(vertex).model("model"), "vertex"],
+        [GoogleVertexResponses.configure(vertex).model("model"), "vertex"],
+        [GoogleVertexMessages.configure(vertex).model("model"), "anthropic"],
+        [Anthropic.configure({ apiKey: "test" }).model("model"), "anthropic"],
+        [
+          AnthropicCompatible.configure({ baseURL: "https://example.test/v1", provider: "minimax" }).model("model"),
+          "minimax",
+        ],
+        [
+          OpenAICompatible.configure({ baseURL: "https://example.test/v1", provider: "custom" }).model("model"),
+          "custom",
+        ],
+        [
+          OpenAICompatibleResponses.configure({ baseURL: "https://example.test/v1", provider: "custom" }).model(
+            "model",
+          ),
+          "custom",
+        ],
+        [Cerebras.configure({ apiKey: "test" }).model("model"), "cerebras"],
+        [DeepInfra.configure({ apiKey: "test" }).model("model"), "deepinfra"],
+        [TogetherAI.configure({ apiKey: "test" }).model("model"), "togetherai"],
+        [CloudflareAIGateway.configure({ accountId: "account" }).model("model"), "cloudflare-ai-gateway"],
+        [CloudflareWorkersAI.configure({ accountId: "account" }).model("model"), "cloudflare-workers-ai"],
+        [OpenRouter.configure({ apiKey: "test" }).model("model"), "openrouter"],
+        [XAI.configure({ apiKey: "test" }).chat("model"), "xai"],
+        [XAI.configure({ apiKey: "test" }).responses("model"), "xai"],
+      ] as const
+
+      for (const [model, key] of providers) expect(model.route.providerMetadataKey).toBe(key)
+    }),
+  )
+
   it.effect("preserves native Together AI and Cerebras provider and route identities", () =>
     Effect.gen(function* () {
       const together = TogetherAI.configure({ apiKey: "fixture" }).model("meta-llama/Llama-3.3-70B")

packages/ai/test/provider/openai-chat-reasoning.recorded.test.ts

@@ -68,11 +68,13 @@ for (const item of cases) {
           expect(response.text.replaceAll(",", "").trim()).toBe("37887")
           expect(response.reasoning.length).toBeGreaterThan(0)
           expect(response.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
-          const metadata = response.message.content.find((part) => part.type === "reasoning")?.providerMetadata
-          expect(metadata?.openai?.reasoningField).toBe(item.structured ? "reasoning" : "reasoning_content")
-          expect(Array.isArray(metadata?.openai?.reasoningDetails)).toBe(item.structured)
+          const metadata = response.message.content.find((part) => part.type === "reasoning")?.providerMetadata?.[
+            item.model.route.providerMetadataKey ?? String(item.model.provider)
+          ]
+          expect(metadata?.reasoningField).toBe(item.structured ? "reasoning" : "reasoning_content")
+          expect(Array.isArray(metadata?.reasoningDetails)).toBe(item.structured)
           if (!item.structured) return
-          const details = metadata?.openai?.reasoningDetails
+          const details = metadata?.reasoningDetails
           if (!Array.isArray(details)) return
           expect(
             details.some(
@@ -126,7 +128,11 @@ for (const item of cases) {
           ).toMatch(/^Paris is sunny\.?$/)
           const details = events
             .filter(LLMEvent.is.reasoningEnd)
-            .map((event) => event.providerMetadata?.openai?.reasoningDetails)
+            .map(
+              (event) =>
+                event.providerMetadata?.[item.model.route.providerMetadataKey ?? String(item.model.provider)]
+                  ?.reasoningDetails,
+            )
             .find(Array.isArray)
           expect(Array.isArray(details)).toBe(item.structured)
           if (!item.structured || !Array.isArray(details)) return

packages/ai/test/provider/openai-chat.test.ts

@@ -903,6 +903,70 @@ describe("OpenAI Chat route", () => {
     }),
   )
 
+  it.effect("uses the configured provider metadata namespace for reasoning and usage", () =>
+    Effect.gen(function* () {
+      const selected = LanguageModel.update(model, {
+        route: { ...model.route, providerMetadataKey: "vendor" },
+      })
+      const details = [{ type: "reasoning.text", text: "thinking", signature: "signed" }]
+      const response = yield* LLMClient.generate(LLMRequest.update(request, { model: selected })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { choices: [{ delta: { reasoning: "thinking", reasoning_details: details } }] },
+              deltaChunk({ content: "Hello" }),
+              deltaChunk({}, "stop"),
+              usageChunk({ prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 }),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
+        vendor: { reasoningField: "reasoning", reasoningDetails: details },
+      })
+      expect(response.usage?.providerMetadata).toEqual({
+        vendor: { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 },
+      })
+
+      const replay = yield* compileRequest(LLM.request({ model: selected, messages: [response.message] }))
+      expect(replay.body.messages).toEqual([
+        { role: "assistant", content: "Hello", reasoning: "thinking", reasoning_details: details },
+      ])
+    }),
+  )
+
+  it.effect("falls back to the selected provider for the metadata namespace", () =>
+    Effect.gen(function* () {
+      const compatible = model.route.with({ provider: "deepseek" }).model({ id: "deepseek-chat" })
+      const selected = LanguageModel.update(compatible, {
+        route: { ...compatible.route, providerMetadataKey: undefined },
+      })
+      const response = yield* LLMClient.generate(LLMRequest.update(request, { model: selected })).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              deltaChunk({ reasoning_content: "thinking" }),
+              deltaChunk({ content: "Hello" }),
+              deltaChunk({}, "stop"),
+              usageChunk({ prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 }),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
+        deepseek: { reasoningField: "reasoning_content" },
+      })
+      expect(response.usage?.providerMetadata).toEqual({
+        deepseek: { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 },
+      })
+
+      const replay = yield* compileRequest(LLM.request({ model: selected, messages: [response.message] }))
+      expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", reasoning_content: "thinking" }])
+    }),
+  )
+
   it.effect("parses and replays a configured custom reasoning field", () =>
     Effect.gen(function* () {
       const custom = LanguageModel.update(model, { compatibility: { reasoningField: "vendor_reasoning" } })

packages/ai/test/provider/openai-compatible-chat.test.ts

@@ -437,7 +437,7 @@ describe("OpenAI-compatible Chat route", () => {
         outputTokens: undefined,
         totalTokens: undefined,
         providerMetadata: {
-          openai: {
+          deepseek: {
             prompt_tokens: null,
             completion_tokens: null,
             total_tokens: null,

packages/ai/test/provider/openai-compatible-responses.test.ts

@@ -195,19 +195,19 @@ describe("Open Responses-compatible route", () => {
           model,
           messages: [
             Message.assistant([
-              { type: "text", text: "Kept.", providerMetadata: { openresponses: { itemId: "history_1" } } },
+              { type: "text", text: "Kept.", providerMetadata: { "openai-compatible": { itemId: "history_1" } } },
               {
                 type: "text",
                 text: "Long.",
-                providerMetadata: { openresponses: { itemId: `history_${"a".repeat(64)}` } },
+                providerMetadata: { "openai-compatible": { itemId: `history_${"a".repeat(64)}` } },
               },
               {
                 type: "text",
                 text: "Opaque.",
-                providerMetadata: { openresponses: { itemId: "provider_value/with+symbols" } },
+                providerMetadata: { "openai-compatible": { itemId: "provider_value/with+symbols" } },
               },
-              { type: "text", text: "No suffix.", providerMetadata: { openresponses: { itemId: "msg_" } } },
-              { type: "text", text: "No prefix.", providerMetadata: { openresponses: { itemId: "_item" } } },
+              { type: "text", text: "No suffix.", providerMetadata: { "openai-compatible": { itemId: "msg_" } } },
+              { type: "text", text: "No prefix.", providerMetadata: { "openai-compatible": { itemId: "_item" } } },
             ]),
           ],
         }),
@@ -267,7 +267,7 @@ describe("Open Responses-compatible route", () => {
               name: item.type,
               result: { type: "json", value: item },
               providerExecuted: true,
-              providerMetadata: { openresponses: { itemId: item.id } },
+              providerMetadata: { example: { itemId: item.id } },
             }),
           ),
         }),
@@ -302,7 +302,7 @@ describe("Open Responses-compatible route", () => {
       )
 
       expect(response.message.content).toEqual([
-        { type: "text", text: "Indexed", providerMetadata: { openresponses: { itemId: "msg_1" } } },
+        { type: "text", text: "Indexed", providerMetadata: { "openai-compatible": { itemId: "msg_1" } } },
       ])
     }),
   )
@@ -368,7 +368,7 @@ describe("Open Responses-compatible route", () => {
                 ),
               )
 
-              const metadata = { openresponses: { itemId: routing.id } }
+              const metadata = { "openai-compatible": { itemId: routing.id } }
               if (fixture.item.type === "function_call") {
                 expect(response.toolCalls).toEqual([
                   expect.objectContaining({
@@ -386,7 +386,7 @@ describe("Open Responses-compatible route", () => {
                     type: "reasoning",
                     text: "Preserved",
                     providerMetadata: {
-                      openresponses: { itemId: routing.id, reasoningEncryptedContent: "encrypted-state" },
+                      "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: "encrypted-state" },
                     },
                   },
                 ])
@@ -438,22 +438,26 @@ describe("Open Responses-compatible route", () => {
             {
               type: "reasoning",
               text: "First.",
-              providerMetadata: { openresponses: { itemId: routing.id } },
+              providerMetadata: { "openai-compatible": { itemId: routing.id } },
             },
             {
               type: "reasoning",
               text: "Second.",
-              providerMetadata: { openresponses: { itemId: routing.id, reasoningEncryptedContent: "final-state" } },
+              providerMetadata: {
+                "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: "final-state" },
+              },
             },
           ])
           expect(response.events.filter(LLMEvent.is.reasoningEnd)).toEqual([
             expect.objectContaining({
               id: `${routing.id}:0`,
-              providerMetadata: { openresponses: { itemId: routing.id } },
+              providerMetadata: { "openai-compatible": { itemId: routing.id } },
             }),
             expect.objectContaining({
               id: `${routing.id}:1`,
-              providerMetadata: { openresponses: { itemId: routing.id, reasoningEncryptedContent: "final-state" } },
+              providerMetadata: {
+                "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: "final-state" },
+              },
             }),
           ])
         }),
@@ -483,7 +487,7 @@ describe("Open Responses-compatible route", () => {
             id: "call_1",
             name: "lookup",
             input: { query: "complete" },
-            providerMetadata: { openresponses: { itemId: "" } },
+            providerMetadata: { "openai-compatible": { itemId: "" } },
           }),
         ])
       }),
@@ -510,7 +514,7 @@ describe("Open Responses-compatible route", () => {
         )
 
         expect(response.message.content).toEqual([
-          { type: "text", text: "Before after", providerMetadata: { openresponses: { itemId: "msg_1" } } },
+          { type: "text", text: "Before after", providerMetadata: { "openai-compatible": { itemId: "msg_1" } } },
         ])
         expect(response.events.map((event) => event.type)).toEqual([
           "step-start",
@@ -662,7 +666,7 @@ describe("Open Responses-compatible route", () => {
 
       expect(response.events.find(LLMEvent.is.toolCall)).toMatchObject({
         input: { query: "complete" },
-        providerMetadata: { openresponses: { itemId: "item_1" } },
+        providerMetadata: { example: { itemId: "item_1" } },
       })
     }),
   )
@@ -694,7 +698,7 @@ describe("Open Responses-compatible route", () => {
       )
 
       expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
-        providerMetadata: { openresponses: { itemId: "rs_raw", reasoningEncryptedContent: "raw-state" } },
+        providerMetadata: { "openai-compatible": { itemId: "rs_raw", reasoningEncryptedContent: "raw-state" } },
       })
     }),
   )
@@ -743,7 +747,7 @@ describe("Open Responses-compatible route", () => {
             Message.assistant({
               type: "text",
               text: "Unclassified.",
-              providerMetadata: { openresponses: { phase: null } },
+              providerMetadata: { "openai-compatible": { phase: null } },
             }),
           ],
         }),
@@ -802,7 +806,7 @@ describe("Open Responses-compatible route", () => {
         {
           type: "text",
           text: "I can't help with that.",
-          providerMetadata: { openresponses: { itemId: "msg_refusal" } },
+          providerMetadata: { example: { itemId: "msg_refusal" } },
         },
       ])
 
@@ -891,7 +895,7 @@ describe("Open Responses-compatible route", () => {
 
       expect(response.toolCalls).toEqual([])
       expect(response.events.find(LLMEvent.is.finish)).toMatchObject({
-        providerMetadata: { openresponses: { responseId: "resp_1" } },
+        providerMetadata: { example: { responseId: "resp_1" } },
       })
     }),
   )

packages/ai/test/provider/openrouter.test.ts

@@ -295,7 +295,7 @@ describe("OpenRouter", () => {
               {
                 type: "reasoning",
                 text: "Thinking",
-                providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
+                providerMetadata: { openrouter: { reasoningField: "reasoning", reasoningDetails: details } },
               },
             ]),
           ],
@@ -328,7 +328,7 @@ describe("OpenRouter", () => {
             Message.assistant({
               type: "reasoning",
               text: "Thinking",
-              providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
+              providerMetadata: { openrouter: { reasoningField: "reasoning", reasoningDetails: details } },
             }),
           ],
         }),
@@ -354,7 +354,7 @@ describe("OpenRouter", () => {
             Message.assistant({
               type: "reasoning",
               text: "AB",
-              providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
+              providerMetadata: { openrouter: { reasoningField: "reasoning", reasoningDetails: details } },
             }),
           ],
         }),

packages/ai/test/route.test.ts

@@ -40,4 +40,13 @@ describe("Route.with", () => {
       "x-patch": "patch",
     })
   })
+
+  test("assigns metadata ownership to a replacement provider and preserves explicit overrides", () => {
+    const route = OpenAIChat.route.with({ provider: "azure" })
+    const overridden = route.with({ providerMetadataKey: "custom-azure" }).with({ headers: { "x-test": "value" } })
+
+    expect(route.providerMetadataKey).toBe("azure")
+    expect(overridden.providerMetadataKey).toBe("custom-azure")
+    expect(overridden.defaults).not.toHaveProperty("providerMetadataKey")
+  })
 })

packages/app/README.md

@@ -31,7 +31,7 @@ Your app is ready to be deployed!
 
 ## E2E Testing
 
-Playwright starts the Vite dev server automatically via `webServer`, and UI tests expect an opencode backend at `localhost:4096` by default.
+Locally, Playwright starts the Vite dev server automatically via `webServer`, or reuses one already running at the configured address. The browser suite uses isolated API fixtures rather than a live opencode backend.
 
 ```bash
 bunx playwright install chromium
@@ -39,11 +39,35 @@ bun run test:e2e:local
 bun run test:e2e:local -- --grep "settings"
 ```
 
+CI builds the app once and runs the same suite against Vite preview, serving production assets from `dist`. Managed built runs never reuse an existing server, so a running dev server cannot silently replace the production build. To run this mode locally:
+
+```bash
+bun run test:e2e:built
+bun run test:e2e:built -- --grep "settings"
+```
+
+To test an already-running dev server without starting or building a server:
+
+```bash
+PLAYWRIGHT_BASE_URL=http://127.0.0.1:4444 bun run test:e2e
+```
+
+For an already-running production build, also set `PLAYWRIGHT_BUILD=1` so the fixture API uses the app's origin:
+
+```bash
+PLAYWRIGHT_BUILD=1 PLAYWRIGHT_BASE_URL=http://127.0.0.1:4444 bun run test:e2e
+```
+
+External targets must use HTTP because fixture URLs use HTTP. `PLAYWRIGHT_BASE_URL` skips server startup and building in either mode.
+
+Compiled CLI startup and service lifecycle coverage runs separately in CI via `packages/cli/script/service-smoke.ts`.
+
 Environment options:
 
-- `PLAYWRIGHT_SERVER_HOST` / `PLAYWRIGHT_SERVER_PORT` (backend address, default: `localhost:4096`)
-- `PLAYWRIGHT_PORT` (Vite dev server port, default: `3000`)
-- `PLAYWRIGHT_BASE_URL` (override base URL, default: `http://localhost:<PLAYWRIGHT_PORT>`)
+- `PLAYWRIGHT_BUILD=1` (build and preview locally; always enabled when `CI` is set)
+- `PLAYWRIGHT_SERVER_HOST` / `PLAYWRIGHT_SERVER_PORT` (dev fixture API address, default: `127.0.0.1:4096`; built runs use the app's origin, matching production)
+- `PLAYWRIGHT_PORT` (managed dev or preview server port, default: `3000`)
+- `PLAYWRIGHT_BASE_URL` (use an externally managed app instead of starting a server; otherwise defaults to `http://127.0.0.1:<PLAYWRIGHT_PORT>`)
 
 ## Deployment
 

packages/app/e2e/performance/README.md

@@ -44,6 +44,7 @@ The suite contains:
 - single-session tab close timing through stable home restoration
 - cached session repaint and mutation tracing
 - streaming timeline throughput, RAF-gap, long-task, geometry, and remount diagnostics
+- retained renderer heap with a large model catalog across repeated session navigation
 
 All benchmarks import the shared `benchmark` fixture. Pages created through Playwright's `page` fixture automatically capture main-frame navigation history and emit a Chrome trace when `OPENCODE_PERFORMANCE_TRACE_DIR` is set. Benchmarks that need isolated browser contexts use `withBenchmarkPage`, which owns the context and the same diagnostics lifecycle.
 
@@ -78,6 +79,26 @@ Benchmarks do not assert machine-dependent performance budgets. Streaming proces
 
 Committed smoke and regression tests continue to own correctness coverage for pagination, tab paint, context resize, collapse state, and composer spacing.
 
+Tab-switch timing starts at `mousedown`, when mouse-selected tabs actually navigate, with a `click` fallback for keyboard activation. The probe excludes hidden/transparent content and intersects answers with their virtual-row clip and viewport. The tab workload requires the destination's final answer to be visible with Markdown ready. These results are not directly comparable to older click-start, geometry-only measurements. `stableObservedMs` includes confirmation across three correct samples; `firstCorrectObservedMs` is the first sample meeting all content and geometry checks. Neither is a compositor presentation timestamp.
+
+Each tab scenario reports one sample, including its raw observations. Use Playwright's `--repeat-each=5` for repeated measurements. Cached scenarios warm the destination at the same panel width before leaving it; a separate resized scenario validates reuse after opening the review pane changes that width.
+
+```sh
+bunx playwright test --config e2e/performance/playwright.config.ts \
+  timeline/session-tab-switch-benchmark.spec.ts --repeat-each=5
+```
+
+## Retained renderer memory
+
+Run the catalog workload against the production app bundle:
+
+```sh
+bunx playwright test --config e2e/performance/playwright.config.ts \
+  timeline/provider-memory-benchmark.spec.ts --repeat-each=3
+```
+
+`PROVIDER_MEMORY_MODELS` defaults to 1,200 and `PROVIDER_MEMORY_SWITCHES` defaults to 10. Each sample records Chromium's `Runtime.getHeapUsage` and `Memory.getDOMCounters` after an explicit garbage collection. This measures retained state, not allocation peaks or normal GC timing. It does not include worker heaps, the Electron main/GPU processes, or the OpenCode server, and must not be reported as total desktop RAM. Use identical model counts and navigation sequences for before/after comparisons.
+
 ## Chrome traces
 
 Set `OPENCODE_PERFORMANCE_TRACE_DIR` to emit a standard Chrome DevTools trace for every benchmark page automatically:

packages/app/e2e/performance/playwright.config.ts

@@ -1,7 +1,7 @@
 import config from "../../playwright.config"
 
 const port = Number(process.env.PLAYWRIGHT_PORT ?? 3000)
-process.env.PLAYWRIGHT_SERVER_PORT = String(port)
+process.env.PLAYWRIGHT_SERVER_PORT = new URL(process.env.PLAYWRIGHT_BASE_URL ?? `http://127.0.0.1:${port}`).port || "80"
 process.env.OPENCODE_PERFORMANCE_RUN_ID ??= `${new Date().toISOString().replace(/[:.]/g, "-")}-${process.pid}`
 
 export default {
@@ -12,9 +12,11 @@ export default {
   fullyParallel: false,
   workers: 1,
   reporter: [["html", { outputFolder: "../playwright-report/performance", open: "never" }], ["line"]],
-  webServer: {
-    ...config.webServer,
-    command: `bun run build && bun run serve -- --host 0.0.0.0 --port ${port} --strictPort`,
-    reuseExistingServer: false,
-  },
+  webServer: config.webServer
+    ? {
+        ...config.webServer,
+        command: `bun run build && bun run serve -- --host 0.0.0.0 --port ${port} --strictPort`,
+        reuseExistingServer: false,
+      }
+    : undefined,
 }

packages/app/e2e/performance/timeline-stability/scroll-interaction.spec.ts

@@ -126,13 +126,19 @@ test("keeps moving upward while drag-selecting above the timeline", async ({ pag
       )
     })
   })
-  const textBox = await text.boundingBox()
+  const textBox = await text.evaluate((element) => {
+    const range = document.createRange()
+    range.selectNodeContents(element)
+    const rect = range.getClientRects()[0]
+    return rect ? { x: rect.x, y: rect.y, width: rect.width, height: rect.height } : null
+  })
   const scrollBox = await scroller.boundingBox()
   expect(textBox).not.toBeNull()
   expect(scrollBox).not.toBeNull()
   if (!textBox || !scrollBox) return
 
-  await page.mouse.move(textBox.x + textBox.width - 10, textBox.y + textBox.height / 2)
+  // Start on a text line, not the empty right edge or gap between wrapped lines.
+  await page.mouse.move(textBox.x + Math.min(20, textBox.width / 2), textBox.y + textBox.height / 2)
   await page.mouse.down()
   await page.mouse.move(textBox.x + 20, scrollBox.y - 120, { steps: 30 })
 
@@ -195,6 +201,45 @@ test("does not pull a keyboard-scrolled user during shell remeasurement", async
   await reportVisualStability(testInfo, "keyboard-during-resize", trace, anchorPlan(regions))
 })
 
+test("keeps an older answer selected while scrolling within the interaction buffer", async ({ page }) => {
+  await setupTimeline(page, {
+    messages: history(80),
+    viewport: { width: 1400, height: 700 },
+    reducedMotion: true,
+  })
+  const scroller = page.locator(".scroll-view__viewport", { has: page.locator("[data-timeline-row]") })
+  const answer = page.getByText("History 78.", { exact: false })
+  await expect(answer).toBeVisible()
+  await expect
+    .poll(() =>
+      answer.evaluate((element) => element.closest('[data-component="markdown"]')?.hasAttribute("data-markdown-ready")),
+    )
+    .toBe(true)
+  const textBox = await answer.evaluate((element) => {
+    const range = document.createRange()
+    range.selectNodeContents(element)
+    const rect = range.getClientRects()[0]
+    return { x: rect.x, y: rect.y, width: rect.width, height: rect.height }
+  })
+  const scrollBox = await scroller.boundingBox()
+  expect(scrollBox).not.toBeNull()
+  if (!scrollBox) return
+  await page.mouse.move(textBox.x + Math.min(180, textBox.width - 2), textBox.y + textBox.height / 2)
+  await page.mouse.down()
+  await page.mouse.move(textBox.x + 2, textBox.y + textBox.height / 2, { steps: 30 })
+  await page.mouse.up()
+  await expect.poll(() => page.evaluate(() => window.getSelection()?.toString())).toContain("History 78.")
+  await page.mouse.move(scrollBox.x + scrollBox.width / 2, scrollBox.y + scrollBox.height / 2)
+  await page.mouse.wheel(0, -450)
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeGreaterThan(400)
+  await expect(answer).toHaveCount(1)
+  await expect.poll(() => page.evaluate(() => window.getSelection()?.toString())).toContain("History 78.")
+  await page.getByRole("heading", { name: "Timeline visual stability" }).click()
+  await expect.poll(() => page.evaluate(() => window.getSelection()?.isCollapsed)).toBe(true)
+})
+
 test("tracks keyboard scrolling from a focused timeline descendant", async ({ page }, testInfo) => {
   const shellID = "prt_descendant_keyboard_01_shell"
   const timeline = await setupTimeline(page, {
@@ -259,12 +304,16 @@ test("does not claim keyboard scrolling owned by a nested scrollable", async ({
   const before = await scroller.evaluate((element) => element.scrollTop)
   const nestedBefore = await nested.evaluate((element) => element.scrollTop)
   await nested.press("PageUp")
-  await page.waitForTimeout(300)
+  await expect.poll(() => nested.evaluate((element) => element.scrollTop)).toBeLessThan(nestedBefore)
   expect(await scroller.evaluate((element) => element.scrollTop)).toBe(before)
-  expect(await nested.evaluate((element) => element.scrollTop)).toBeLessThan(nestedBefore)
 
-  await nested.evaluate((element) => (element.scrollTop = 0))
-  await scroller.evaluate((element) => (element.scrollTop = Math.min(300, element.scrollHeight - element.clientHeight)))
+  await nested.evaluate((element) => element.scrollTo({ top: 0, behavior: "instant" }))
+  await expect.poll(() => nested.evaluate((element) => element.scrollTop)).toBe(0)
+  await scroller.evaluate((element) => {
+    element.dispatchEvent(new WheelEvent("wheel", { bubbles: true, cancelable: true, deltaY: -1 }))
+    element.scrollTo({ top: Math.min(300, element.scrollHeight - element.clientHeight), behavior: "instant" })
+  })
+  await expect.poll(() => scroller.evaluate((element) => element.scrollTop)).toBeLessThan(500)
   const boundaryBefore = await scroller.evaluate((element) => element.scrollTop)
   expect(boundaryBefore).toBeGreaterThan(0)
   await nested.press("PageUp")

packages/app/e2e/performance/timeline/provider-memory-benchmark.spec.ts

@@ -0,0 +1,98 @@
+import { benchmark, expect } from "../benchmark"
+import { mockOpenCodeServer } from "../../utils/mock-server"
+import { expectSessionTitle } from "../../utils/waits"
+import { fixture, pageMessages } from "./session-timeline-stress.fixture"
+import { installStressSessionTabs, installTimelineSettings, stressSessionHref } from "./timeline-test-helpers"
+import { waitForStableTimeline } from "./session-tab-switch-probe"
+import type { CatalogUpdated } from "@opencode-ai/client/promise"
+
+benchmark("measures retained renderer memory with a large model catalog", async ({ page, report }) => {
+  benchmark.setTimeout(120_000)
+  const count = Number(process.env.PROVIDER_MEMORY_MODELS ?? 1200)
+  const switches = Number(process.env.PROVIDER_MEMORY_SWITCHES ?? 10)
+  const provider = fixture.provider.all[0]
+  const selected = { ...provider.models["claude-opus-4-6"] }
+  await mockOpenCodeServer(page, {
+    directory: fixture.directory,
+    project: fixture.project,
+    sessions: fixture.sessions,
+    pageMessages,
+    provider: {
+      ...fixture.provider,
+      all: [
+        {
+          ...provider,
+          models: {
+            [selected.id]: selected,
+            ...Object.fromEntries(
+              Array.from({ length: count - 1 }, (_, index) => {
+                const id = `catalog-model-${index}`
+                return [
+                  id,
+                  {
+                    id,
+                    name: `Catalog model ${index}`,
+                    cost: { input: 1, output: 2 },
+                    limit: { context: 200_000, output: 8192 },
+                    variants: { high: { reasoningEffort: "high" } },
+                  },
+                ]
+              }),
+            ),
+          },
+        },
+      ],
+    },
+  })
+  await installTimelineSettings(page)
+  await installStressSessionTabs(page)
+  await page.goto(stressSessionHref(fixture.sourceID))
+  await expectSessionTitle(page, fixture.expected.sourceTitle)
+  await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
+  await expect(page.locator('[data-action="composer-model"]')).toContainText("Claude Opus 4.6")
+  const cdp = await page.context().newCDPSession(page)
+  const samples = []
+  for (let index = 0; index <= switches; index++) {
+    if (index > 0) {
+      const target = index % 2 === 1
+      const id = target ? fixture.targetID : fixture.sourceID
+      await page.locator(`[data-slot="titlebar-tabs"] a[href="${stressSessionHref(id)}"]`).click()
+      await expectSessionTitle(page, target ? fixture.expected.targetTitle : fixture.expected.sourceTitle)
+      await waitForStableTimeline(
+        page,
+        (target ? fixture.expected.targetMessageIDs : fixture.expected.sourceMessageIDs).at(-1)!,
+      )
+      await expect(page.locator('[data-action="composer-model"]')).toContainText("Claude Opus 4.6")
+    }
+    // GC is an explicit retained-heap measurement, not an application optimization or readiness wait.
+    await cdp.send("HeapProfiler.collectGarbage")
+    samples.push({
+      switches: index,
+      heap: await cdp.send("Runtime.getHeapUsage"),
+      dom: await cdp.send("Memory.getDOMCounters"),
+    })
+  }
+  expect(samples).toHaveLength(switches + 1)
+  expect(samples.every((sample) => sample.heap.usedSize > 0)).toBe(true)
+  selected.name = "Updated catalog model"
+  await page.evaluate(
+    (event) => {
+      const host = window as Window & { __mockServerStream?: { push: (events: CatalogUpdated[]) => void } }
+      if (!host.__mockServerStream) throw new Error("Missing fixture event stream")
+      host.__mockServerStream.push([event])
+    },
+    {
+      id: "evt_catalog_refresh",
+      created: Date.now(),
+      type: "catalog.updated",
+      location: { directory: fixture.directory },
+      data: {},
+    } satisfies CatalogUpdated,
+  )
+  await expect(page.locator('[data-action="composer-model"]')).toContainText(selected.name)
+  report(
+    { samples },
+    { models: count, switches, gc: "explicit", scope: "renderer main isolate; not total desktop RAM" },
+  )
+  await cdp.detach()
+})

packages/app/e2e/performance/timeline/session-tab-switch-benchmark.spec.ts

@@ -11,115 +11,73 @@ import {
 } from "./timeline-test-helpers"
 import { measureSessionSwitch, waitForStableTimeline } from "./session-tab-switch-probe"
 
-type Result = Awaited<ReturnType<typeof measureSessionSwitch>>
+const scenarios = [
+  { cached: false, review: false, resized: false },
+  { cached: false, review: true, resized: false },
+  { cached: true, review: false, resized: false },
+  { cached: true, review: true, resized: false },
+  { cached: true, review: true, resized: true },
+]
 
-benchmark(
-  "benchmarks session tab switching with and without the review pane",
-  async ({ browser, report }, testInfo) => {
-    benchmark.setTimeout(360_000)
-    const runs = Number(process.env.SESSION_TAB_SWITCH_RUNS ?? 5)
-    const results = {
-      closed: { cold: [] as Result[], hot: [] as Result[] },
-      open: { cold: [] as Result[], hot: [] as Result[] },
-    }
-    for (const reviewPane of ["closed", "open"] as const) {
-      for (const mode of ["cold", "hot"] as const) {
-        for (let run = 0; run < runs; run++) {
-          results[reviewPane][mode].push(
-            await withBenchmarkPage(
-              browser,
-              `session-tab-switch-${reviewPane}-${mode}-${run}`,
-              (page) => trial(page, mode, reviewPane),
-              testInfo,
-            ),
-          )
+scenarios.forEach((scenario) => {
+  const name = `tab switch: ${scenario.cached ? "cached" : "unmounted"}, review ${scenario.review ? "open" : "closed"}${scenario.resized ? ", resized" : ""}`
+  benchmark(name, async ({ browser, report }, testInfo) => {
+    const result = await withBenchmarkPage(
+      browser,
+      name,
+      async (page) => {
+        await mockStressTimeline(page, { vcsDiff: createReviewDiffs() })
+        await installTimelineSettings(page)
+        await installStressSessionTabs(page)
+        await page.goto(stressSessionHref(fixture.sourceID))
+        await expectSessionTitle(page, fixture.expected.sourceTitle)
+        await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
+        if (scenario.review && !scenario.resized) await openReviewPane(page)
+        if (scenario.cached) {
+          await switchSession(page, fixture.targetID, fixture.expected.targetTitle)
+          const answer = page.locator(`[data-timeline-part-id="${fixture.expected.targetPartIDs.at(-1)}"]`)
+          await expect(answer.locator('[data-component="markdown"]')).toHaveAttribute("data-markdown-ready", "")
+          await expect
+            .poll(() =>
+              answer.evaluate((element) => element.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })),
+            )
+            .toBe(true)
+          await waitForStableTimeline(page, fixture.expected.targetMessageIDs.at(-1)!)
+          await switchSession(page, fixture.sourceID, fixture.expected.sourceTitle)
         }
-      }
-    }
-    report({ results, summary: summarizeReviewPane(results) }, { runs, reviewDiffs: createReviewDiffs().length })
-  },
-)
+        if (scenario.resized) await openReviewPane(page)
+        await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
 
-async function trial(page: Page, mode: "cold" | "hot", reviewPane: "closed" | "open") {
-  const reviewDiffs = createReviewDiffs()
-  await mockStressTimeline(page, { vcsDiff: reviewDiffs })
-  await installTimelineSettings(page)
-  await installStressSessionTabs(page)
-  if (mode === "hot") {
-    await page.goto(stressSessionHref(fixture.targetID))
-    await expectSessionTitle(page, fixture.expected.targetTitle)
-    await waitForStableTimeline(page, fixture.expected.targetMessageIDs.at(-1)!)
-    await switchSession(page, fixture.sourceID, fixture.expected.sourceTitle)
-  } else {
-    await page.goto(stressSessionHref(fixture.sourceID))
-    await expectSessionTitle(page, fixture.expected.sourceTitle)
-  }
-  await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
-  if (reviewPane === "open") {
-    await openReviewPane(page)
-    await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
-  }
-
-  const destinationIDs = fixture.messages[fixture.targetID].map((message) => message.id)
-  const sourceIDs = fixture.messages[fixture.sourceID].map((message) => message.id)
-  const lastID = fixture.expected.targetMessageIDs.at(-1)!
-  const href = stressSessionHref(fixture.targetID)
-  const result = await measureSessionSwitch(page, {
-    destinationIDs,
-    sourceIDs,
-    lastID,
-    href,
-    switch: () => switchSession(page, fixture.targetID, fixture.expected.targetTitle),
-  })
-  return result
-}
-
-function summarize(results: Record<"cold" | "hot", Result[]>) {
-  const stats = (values: (number | null)[]) => {
-    const sorted = values.filter((value): value is number => value !== null).sort((a, b) => a - b)
-    return {
-      min: sorted[0] ?? null,
-      median: sorted[Math.floor(sorted.length / 2)] ?? null,
-      max: sorted.at(-1) ?? null,
-      missing: values.length - sorted.length,
-    }
-  }
-  return Object.fromEntries(
-    Object.entries(results).map(([mode, values]) => [
-      mode,
-      {
-        firstDestinationObservedMs: stats(values.map((value) => value.firstDestinationObservedMs)),
-        firstCorrectObservedMs: stats(values.map((value) => value.firstCorrectObservedMs)),
-        stableObservedMs: stats(values.map((value) => value.stableObservedMs)),
+        return measureSessionSwitch(page, {
+          destinationIDs: fixture.messages[fixture.targetID].map((message) => message.id),
+          sourceIDs: fixture.messages[fixture.sourceID].map((message) => message.id),
+          lastID: fixture.expected.targetMessageIDs.at(-1)!,
+          requiredPartID: fixture.expected.targetPartIDs.at(-1),
+          href: stressSessionHref(fixture.targetID),
+          switch: () => switchSession(page, fixture.targetID, fixture.expected.targetTitle),
+        })
       },
-    ]),
-  )
-}
-
-function summarizeReviewPane(results: Record<"closed" | "open", Record<"cold" | "hot", Result[]>>) {
-  return Object.fromEntries(
-    Object.entries(results).map(([reviewPane, values]) => [
-      reviewPane,
-      summarize(values as Record<"cold" | "hot", Result[]>),
-    ]),
-  )
-}
+      testInfo,
+    )
+    expect(result.unknownSamples).toBe(0)
+    expect(result.wrongDestinationSamples).toBe(0)
+    if (scenario.cached) expect(result.blankSamples).toBe(0)
+    report(result, { ...scenario, inputEvent: "mousedown", requireReadyAnswer: true })
+  })
+})
 
 async function switchSession(page: Page, sessionID: string, title: string) {
-  const href = stressSessionHref(sessionID)
-  const tab = page.locator(`[data-slot="titlebar-tabs"] a[href="${href}"]`).first()
-  await expect(tab).toBeVisible()
+  const tab = page.locator(`[data-slot="titlebar-tabs"] a[href="${stressSessionHref(sessionID)}"]`)
+  await expect(tab).toHaveCount(1)
   await tab.click()
   await expectSessionTitle(page, title)
 }
 
 async function openReviewPane(page: Page) {
   await page.getByRole("button", { name: "Toggle review" }).click()
-  const panel = page.locator("#review-panel")
-  await expect(panel).toBeVisible()
+  await expect(page.locator("#review-panel")).toBeVisible()
   await page.waitForFunction(() => {
-    const panel = document.querySelector<HTMLElement>("#review-panel")
-    const text = panel?.textContent ?? ""
+    const text = document.querySelector("#review-panel")?.textContent ?? ""
     return text.includes("generated-000.ts") && text.includes("+3")
   })
 }

packages/app/e2e/performance/timeline/session-tab-switch-metrics.ts

@@ -20,9 +20,10 @@ export function classifySessionSwitch(samples: SessionSwitchSample[]) {
   const firstCorrect = samples.findIndex(isCorrectDestination)
   const stable = samples.findIndex((_, index) => isStableSessionSwitch(samples.slice(index, index + 3)))
   return {
+    samples,
     firstDestinationObservedMs: samples[firstDestination]?.observedAtMs ?? null,
     firstCorrectObservedMs: samples[firstCorrect]?.observedAtMs ?? null,
-    stableObservedMs: samples[stable + 2]?.observedAtMs ?? null,
+    stableObservedMs: stable < 0 ? null : samples[stable + 2].observedAtMs,
     wrongDestinationSamples: samples
       .slice(firstDestination)
       .filter((sample) => sample.destination.length > 0 && !sample.last).length,

packages/app/e2e/performance/timeline/session-tab-switch-probe.spec.ts

@@ -0,0 +1,67 @@
+import { benchmark, expect } from "../benchmark"
+import { measureSessionSwitch } from "./session-tab-switch-probe"
+import type { SessionSwitchSample } from "./session-tab-switch-metrics"
+
+benchmark("starts at mousedown and excludes hidden or unfinished destination content", async ({ page, report }) => {
+  await page.setContent(`
+    <a href="/session/destination">Destination</a>
+    <div class="scroll-view__viewport" style="height:200px;overflow:auto">
+      <div data-timeline-row="message" data-timeline-key="row" data-message-id="source">
+        <div data-timeline-part-id="answer"><div data-component="markdown">Destination answer</div></div>
+      </div>
+    </div>
+  `)
+  await page.evaluate(() => {
+    document.querySelector("a")!.addEventListener("mousedown", () => {
+      const row = document.querySelector<HTMLElement>("[data-message-id]")!
+      row.dataset.messageId = "destination"
+      row.style.visibility = "hidden"
+    })
+  })
+  const result = await measureSessionSwitch(page, {
+    destinationIDs: ["destination"],
+    sourceIDs: ["source"],
+    lastID: "destination",
+    requiredPartID: "answer",
+    requireBottomAnchor: false,
+    href: "/session/destination",
+    switch: async () => {
+      // No click is dispatched: the probe must observe the event that activates tabs.
+      await page.getByRole("link", { name: "Destination" }).dispatchEvent("mousedown", { button: 0 })
+      await page.waitForFunction(() => {
+        const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
+        return host.__sessionSwitchProbe?.samples.some((sample) => !sample.hasVisibleRows)
+      })
+      await page.locator("[data-message-id]").evaluate((row) => row.style.removeProperty("visibility"))
+      await page.waitForFunction(() => {
+        const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
+        return host.__sessionSwitchProbe?.samples.some(
+          (sample) => sample.destination.length > 0 && sample.requiredPartVisible === false,
+        )
+      })
+      const beforeClip = await page.evaluate(() => {
+        const row = document.querySelector<HTMLElement>("[data-timeline-key]")!
+        row.style.cssText = "height:10px;position:relative;overflow:clip"
+        const answer = row.querySelector<HTMLElement>("[data-timeline-part-id]")!
+        answer.style.cssText = "position:absolute;top:30px;width:150px"
+        answer.querySelector('[data-component="markdown"]')!.setAttribute("data-markdown-ready", "")
+        return (
+          (window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }).__sessionSwitchProbe
+            ?.samples.length ?? 0
+        )
+      })
+      await page.waitForFunction((count) => {
+        const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
+        return host.__sessionSwitchProbe?.samples.slice(count).some((sample) => sample.requiredPartVisible === false)
+      }, beforeClip)
+      await page.locator("[data-timeline-key]").evaluate((row) => {
+        row.style.height = "100px"
+      })
+    },
+  })
+  expect(result.blankSamples).toBeGreaterThan(0)
+  expect(result.firstCorrectObservedMs).not.toBeNull()
+  expect(result.stableObservedMs).not.toBeNull()
+  expect(result.firstCorrectObservedMs).toBeGreaterThan(result.firstDestinationObservedMs!)
+  report(result)
+})

packages/app/e2e/performance/timeline/session-tab-switch-probe.ts

@@ -25,7 +25,7 @@ async function installSessionSwitchProbe(
     let running = true
     const reviewLevels: Record<string, string> = {
       panel: "#review-panel",
-    tabs: '#review-panel [data-component="tabs"]',
+      tabs: '#review-panel [data-component="tabs"]',
       body: '#review-panel [data-slot="session-review-v2-body"]',
       review: '#review-panel [data-component="session-review-v2"]',
       preview: '#review-panel [data-slot="session-review-v2-preview"]',
@@ -37,7 +37,6 @@ async function installSessionSwitchProbe(
       if (!running || started === undefined) return
       setTimeout(() => {
         if (!running || started === undefined) return
-        const observedAtMs = performance.now() - started
         const reviewPanel = document.querySelector<HTMLElement>("#review-panel")
         const reviewFile = reviewPanel?.querySelector('[data-component="file"][data-mode="diff"]')
         const initialReviewFile = initialReviewNodes.file
@@ -63,26 +62,30 @@ async function installSessionSwitchProbe(
         )
         if (root) {
           const view = root.getBoundingClientRect()
-          const visible = [...root.querySelectorAll<HTMLElement>("[data-message-id]")]
-            .filter((element) => {
-              const rect = element.getBoundingClientRect()
-              return rect.bottom > view.top && rect.top < view.bottom
-            })
-            .map((element) => element.dataset.messageId!)
-          const hasVisibleRows = [...root.querySelectorAll<HTMLElement>("[data-timeline-key]")].some((element) => {
+          const inViewport = (element: HTMLElement) => {
+            if (!element.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })) return false
             const rect = element.getBoundingClientRect()
-            return rect.bottom > view.top && rect.top < view.bottom
-          })
+            const clip = element.closest<HTMLElement>("[data-timeline-key]")?.getBoundingClientRect() ?? view
+            return (
+              Math.min(rect.bottom, clip.bottom, view.bottom) > Math.max(rect.top, clip.top, view.top) &&
+              Math.min(rect.right, clip.right, view.right) > Math.max(rect.left, clip.left, view.left)
+            )
+          }
+          const visible = [...root.querySelectorAll<HTMLElement>("[data-message-id]")]
+            .filter(inViewport)
+            .map((element) => element.dataset.messageId!)
+          const hasVisibleRows = [...root.querySelectorAll<HTMLElement>("[data-timeline-key]")].some(inViewport)
           const requiredPartVisible = requiredPartID
             ? [...root.querySelectorAll<HTMLElement>("[data-timeline-part-id]")].some((element) => {
                 if (element.dataset.timelinePartId !== requiredPartID) return false
-                const rect = element.getBoundingClientRect()
-                return rect.width > 0 && rect.height > 0 && rect.bottom > view.top && rect.top < view.bottom
+                if (!element.textContent?.trim()) return false
+                if (element.querySelector('[data-component="markdown"]:not([data-markdown-ready])')) return false
+                return inViewport(element)
               })
             : undefined
           const spacer = root.querySelector<HTMLElement>('[data-timeline-row="bottom-spacer"]')?.getBoundingClientRect()
           samples.push({
-            observedAtMs,
+            observedAtMs: performance.now() - started,
             destination: visible.filter((id) => destination.has(id)),
             source: visible.filter((id) => source.has(id)),
             hasVisibleRows,
@@ -94,7 +97,7 @@ async function installSessionSwitchProbe(
           })
         } else {
           samples.push({
-            observedAtMs,
+            observedAtMs: performance.now() - started,
             destination: [],
             source: [],
             hasVisibleRows: false,
@@ -107,23 +110,25 @@ async function installSessionSwitchProbe(
         requestAnimationFrame(sample)
       }, 0)
     }
-    document.addEventListener(
-      "click",
-      (event) => {
-        const link = event.target instanceof Element ? event.target.closest("a") : undefined
-        if (link?.getAttribute("href") !== href) return
-        started = performance.now()
-        for (const [name, selector] of Object.entries(reviewLevels)) {
-          initialReviewNodes[name] = document.querySelector(selector)
-        }
-        requestAnimationFrame(sample)
-      },
-      { capture: true, once: true },
-    )
+    const start = (event: MouseEvent) => {
+      if (started !== undefined || event.button !== 0) return
+      const link = event.target instanceof Element ? event.target.closest("a") : undefined
+      if (link?.getAttribute("href") !== href) return
+      started = performance.now()
+      for (const [name, selector] of Object.entries(reviewLevels)) {
+        initialReviewNodes[name] = document.querySelector(selector)
+      }
+      requestAnimationFrame(sample)
+    }
+    // Tabs activate on mousedown; click alone misses the synchronous navigation work.
+    document.addEventListener("mousedown", start, true)
+    document.addEventListener("click", start, true)
     ;(window as Window & { __sessionSwitchProbe?: SessionSwitchProbe }).__sessionSwitchProbe = {
       samples,
       stop: () => {
         running = false
+        document.removeEventListener("mousedown", start, true)
+        document.removeEventListener("click", start, true)
       },
     }
   }, input)

packages/app/e2e/performance/unit/session-tab-switch-metrics.test.ts

@@ -53,6 +53,15 @@ test("reports missing correctness without throwing", () => {
   expect(result.stableObservedMs).toBeNull()
 })
 
+test("does not report stability for only two correct samples", () => {
+  const result = classifySessionSwitch([
+    { observedAtMs: 16, destination: ["destination"], source: [], hasVisibleRows: true, last: true, bottomErrorPx: 0 },
+    { observedAtMs: 32, destination: ["destination"], source: [], hasVisibleRows: true, last: true, bottomErrorPx: 0 },
+  ])
+  expect(result.firstCorrectObservedMs).toBe(16)
+  expect(result.stableObservedMs).toBeNull()
+})
+
 test("requires an explicitly tracked part to be visible", () => {
   const result = classifySessionSwitch([
     {

packages/app/e2e/regression/cross-server-tab-close.spec.ts

@@ -3,7 +3,7 @@ import { base64Encode } from "@opencode-ai/util/encode"
 import { currentSession } from "../utils/mock-server"
 import { installSseTransport } from "../utils/sse-transport"
 
-const serverA = "http://127.0.0.1:4096"
+const serverA = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
 const serverB = "http://127.0.0.1:4097"
 const sessionA = session("ses_server_a", "C:/server-a", "Server A session")
 const sessionB = session("ses_server_b", "/home/server-b", "Server B session")
@@ -12,17 +12,17 @@ test("closing the active server's last tab opens the remaining server tab", asyn
   const requests: string[] = []
   await mockServers(page, requests)
   await page.addInitScript(
-    ({ serverB, sessionA, sessionB }) => {
+    ({ serverA, serverB, sessionA, sessionB }) => {
       localStorage.setItem("opencode.global.dat:server", JSON.stringify({ list: [serverB] }))
       localStorage.setItem(
         "opencode.window.browser.dat:tabs",
         JSON.stringify([
-          { type: "session", server: "http://127.0.0.1:4096", sessionId: sessionA },
+          { type: "session", server: serverA, sessionId: sessionA },
           { type: "session", server: serverB, sessionId: sessionB },
         ]),
       )
     },
-    { serverB, sessionA: sessionA.id, sessionB: sessionB.id },
+    { serverA, serverB, sessionA: sessionA.id, sessionB: sessionB.id },
   )
 
   const hrefA = `/server/${base64Encode(serverA)}/session/${sessionA.id}`
@@ -55,7 +55,7 @@ function session(id: string, directory: string, title: string) {
 async function mockServers(page: Page, requests: string[]) {
   await installSseTransport(page, { server: serverA })
   await installSseTransport(page, { server: serverB })
-  await page.route("**/*", async (route) => {
+  await page.route("**/api/**", async (route) => {
     const url = new URL(route.request().url())
     if (url.origin !== serverA && url.origin !== serverB) return route.fallback()
     requests.push(url.toString())

packages/app/e2e/regression/project-extensions.spec.ts

@@ -0,0 +1,103 @@
+import { expect, test } from "@playwright/test"
+import { base64Encode } from "@opencode-ai/util/encode"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+const directory = "C:/Projects/extensions-demo"
+const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
+const session = {
+  id: "ses_project_extensions",
+  title: "Existing session",
+  directory,
+  projectID: "proj_extensions_demo",
+  time: { created: 1700000000000, updated: 1700000000000 },
+}
+
+test.use({ viewport: { width: 1440, height: 1000 }, colorScheme: "dark" })
+
+test("project Extensions stays inside settings while plugins load", async ({ page }) => {
+  await mockOpenCodeServer(page, {
+    directory,
+    project: {
+      id: session.projectID,
+      canonical: directory,
+      name: "Extensions demo",
+      vcs: "git",
+      time: session.time,
+      sandboxes: [],
+    },
+    provider: { all: [], connected: [], default: {} },
+    sessions: [session],
+    pageMessages: () => ({ items: [] }),
+  })
+  await page.addInitScript(
+    ({ server, sessionID, directory }) => {
+      localStorage.setItem(
+        "opencode.global.dat:server",
+        JSON.stringify({ projects: { local: [{ worktree: directory, expanded: true }] } }),
+      )
+      localStorage.setItem(
+        "opencode.window.browser.dat:tabs",
+        JSON.stringify([{ type: "session", server, sessionId: sessionID }]),
+      )
+    },
+    { server, sessionID: session.id, directory },
+  )
+  const href = `/server/${base64Encode(server)}/session/${session.id}`
+  await page.goto(href)
+  await expect(page.getByRole("heading", { name: session.title, exact: true })).toBeVisible()
+  await page.keyboard.press("Control+,")
+  const settings = page.getByTestId("settings-screen")
+  await settings.getByRole("tab", { name: "Projects", exact: true }).click()
+  await settings.getByText("Extensions demo", { exact: true }).click()
+  const dialog = page.getByRole("dialog")
+  await expect(dialog.getByRole("textbox", { name: "Name", exact: true })).toBeFocused()
+
+  const globalPlugins = Promise.withResolvers<void>()
+  const projectPlugins = Promise.withResolvers<void>()
+  await page.route(
+    (url) => url.pathname === "/api/plugin",
+    async (route) => {
+      const project = new URL(route.request().url()).searchParams.get("location[directory]")
+      await (project ? projectPlugins : globalPlugins).promise
+      await route.fulfill({
+        json: {
+          location: project ? { directory: project } : {},
+          data: (project ? ["shared-plugin", "project-plugin"] : ["shared-plugin"]).map((id) => ({
+            id,
+            source: { type: "package", package: id },
+            status: "active",
+            tui: false,
+          })),
+        },
+      })
+    },
+  )
+  const requested = page.waitForRequest((request) => {
+    const url = new URL(request.url())
+    return url.pathname === "/api/plugin" && url.searchParams.get("location[directory]") === directory
+  })
+  await dialog.getByRole("tab", { name: "Extensions", exact: true }).click()
+  await requested
+  await expect(page).toHaveURL(href)
+  await expect(dialog.getByRole("heading", { name: "Extensions", exact: true })).toBeVisible()
+  await expect(settings).toBeVisible()
+  await expect(page.getByRole("heading", { name: session.title, exact: true, includeHidden: true })).toBeHidden()
+  await dialog.getByRole("tab", { name: "Plugins", exact: true }).click()
+  await expect(dialog.getByRole("tab", { name: "Plugins", exact: true })).toHaveAttribute("aria-selected", "true")
+
+  globalPlugins.resolve()
+  await dialog.getByRole("tab", { name: "Scripts", exact: true }).click()
+  await expect(dialog.getByRole("heading", { name: "Scripts", exact: true })).toBeVisible()
+  await dialog.getByRole("tab", { name: "Extensions", exact: true }).click()
+  projectPlugins.resolve()
+  await dialog.getByRole("tab", { name: "Plugins", exact: true }).click()
+  await expect(dialog.getByText("project-plugin", { exact: true })).toBeVisible()
+  await dialog.getByRole("button", { name: "Shared with all projects 1", exact: true }).click()
+  await expect(dialog.getByText("shared-plugin", { exact: true })).toBeVisible()
+  await expect(page).toHaveURL(href)
+
+  await page.keyboard.press("Escape")
+  await expect(dialog).toBeHidden()
+  await expect(settings.getByRole("tab", { name: "Projects", exact: true })).toHaveAttribute("aria-selected", "true")
+  await expect(page.getByRole("heading", { name: session.title, exact: true, includeHidden: true })).toBeHidden()
+})

packages/app/e2e/regression/remote-session-settings.spec.ts

@@ -3,7 +3,7 @@ import { expect, test, type Page, type Route } from "@playwright/test"
 import { installSseTransport } from "../utils/sse-transport"
 import { currentSession } from "../utils/mock-server"
 
-const serverA = `http://127.0.0.1:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
+const serverA = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
 const serverB = "http://127.0.0.1:4097"
 const directoryA = "C:/server-a"
 const directoryB = "/home/server-b"
@@ -308,7 +308,7 @@ async function mockServers(
   permissionResponses: PermissionResponse[] = [],
   options: MockServerOptions = {},
 ) {
-  await page.route("**/*", async (route) => {
+  await page.route("**/api/**", async (route) => {
     const url = new URL(route.request().url())
     if (url.origin !== serverA && url.origin !== serverB) return route.fallback()
     const remote = url.origin === serverB

packages/app/e2e/regression/remote-tab-busy.spec.ts

@@ -2,7 +2,7 @@ import { expect, test, type Page, type Route } from "@playwright/test"
 import { base64Encode } from "@opencode-ai/util/encode"
 import { currentSession } from "../utils/mock-server"
 
-const serverA = "http://127.0.0.1:4096"
+const serverA = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
 const serverB = "http://127.0.0.1:4097"
 const sessionA = session("ses_server_a", "C:/server-a", "Server A session")
 const sessionB = session("ses_server_b", "/home/server-b", "Server B session")
@@ -51,7 +51,7 @@ function session(id: string, directory: string, title: string) {
 }
 
 async function mockServers(page: Page) {
-  await page.route("**/*", async (route) => {
+  await page.route("**/api/**", async (route) => {
     const url = new URL(route.request().url())
     if (url.origin !== serverA && url.origin !== serverB) return route.fallback()
     const current = url.origin === serverA ? sessionA : sessionB

packages/app/e2e/regression/session-queue.spec.ts

@@ -1,5 +1,5 @@
 import { expect, test, type Page } from "@playwright/test"
-import type { OpenCodeEvent } from "@opencode-ai/client/promise"
+import type { OpenCodeEvent, SessionMessageInfo } from "@opencode-ai/client/promise"
 import { base64Encode } from "@opencode-ai/util/encode"
 import { mockOpenCodeServer } from "../utils/mock-server"
 import { expectAppVisible } from "../utils/waits"
@@ -18,7 +18,7 @@ type InboxRow = {
   delivery: "steer" | "queue"
 }
 
-function createQueueMock(seed: string[]) {
+function createQueueMock(seed: string[], messages: SessionMessageInfo[] = []) {
   const rows: InboxRow[] = seed.map((text, index) => ({
     id: `inb_seed_${index + 1}`,
     sessionID,
@@ -32,13 +32,16 @@ function createQueueMock(seed: string[]) {
   const changes: { inboxID: string; action: "cancel" | "steer" }[] = []
   const log: string[] = []
   let sequence = 0
-  const emit = (type: OpenCodeEvent["type"], data: OpenCodeEvent["data"]) => {
+  const emit = <Type extends OpenCodeEvent["type"]>(
+    type: Type,
+    data: Extract<OpenCodeEvent, { type: Type }>["data"],
+  ) => {
     sequence += 1
     events.push({
       id: `evt_queue_${sequence}`,
       type,
       created: Date.now(),
-      durable: { aggregateID: sessionID, seq: sequence, version: 1 },
+      durable: { aggregateID: sessionID, seq: sequence, version: type === "session.tool.success" ? 2 : 1 },
       data,
     } as OpenCodeEvent)
   }
@@ -47,6 +50,8 @@ function createQueueMock(seed: string[]) {
     prompts,
     changes,
     log,
+    messages,
+    emit,
     events: () => events.splice(0),
     onPrompt: (input: { sessionID: string; body: Record<string, unknown> }) => {
       prompts.push(input.body)
@@ -126,10 +131,11 @@ async function openSession(page: Page, mock: ReturnType<typeof createQueueMock>,
         directory,
         title: "Session queue regression",
         version: "dev",
+        model: { id: "queue-model", providerID: "opencode" },
         time: { created: 1700000000000, updated: 1700000000000 },
       },
     ],
-    pageMessages: () => ({ items: [] }),
+    pageMessages: () => ({ items: mock.messages }),
     sessionStatus: () => ({ [sessionID]: { type: "running" } }),
     inbox: () => mock.rows.map((row) => ({ ...row, payload: { ...row.payload } })),
     onPrompt: mock.onPrompt,
@@ -227,3 +233,130 @@ test("editing restores the existing draft and replaces only the original queue p
   expect(mock.changes.map((change) => change.action)).toEqual(["cancel", "cancel", "cancel"])
   expect(mock.log[0]).toBe("prompt:queue")
 })
+
+for (const delivery of ["steer", "queue"] as const) {
+  test(`keeps finished tools above a pending ${delivery === "queue" ? "queue-to-steer" : "steer"} follow-up`, async ({
+    page,
+  }, testInfo) => {
+    const model = { id: "queue-model", providerID: "opencode" }
+    const userID = "msg_queue_initial_user"
+    const assistantID = "msg_queue_continued_assistant"
+    const followUp = "U2: Also check the retry path."
+    const mock = createQueueMock(
+      [],
+      [
+        { id: userID, type: "user", text: "U1: Inspect the queue ordering.", time: { created: 1700000000000 } },
+        {
+          id: "msg_queue_initial_assistant",
+          type: "assistant",
+          agent: "build",
+          model,
+          content: [{ type: "text", text: "A1: I will inspect the current implementation." }],
+          finish: "tool-calls",
+          time: { created: 1700000000001, completed: 1700000000002 },
+        },
+      ],
+    )
+    const view = await openSession(page, mock, delivery)
+    const transcript = page.locator("[data-timeline-virtual-content]")
+    const thinking = transcript.locator('[data-timeline-row="Thinking"]')
+    await expect(transcript.getByText("A1: I will inspect the current implementation.", { exact: true })).toBeVisible()
+    await expect(thinking).toBeVisible()
+    await expect(view.input).toBeEditable()
+    await view.input.fill(followUp)
+    await view.input.press("Enter")
+    await expect.poll(() => mock.rows.map((row) => row.delivery)).toEqual([delivery])
+    await expect(view.input).toHaveText("")
+
+    const inboxID = mock.rows[0].id
+    const pending = transcript.locator(`[data-timeline-row="UserMessage"][data-message-id="${inboxID}"]`)
+    if (delivery === "queue") {
+      const queued = view.rows.filter({ hasText: followUp })
+      await expect(queued).toBeVisible()
+      await expect(pending).toHaveCount(0)
+      await queued.hover()
+      await queued.getByRole("button", { name: "Steer", exact: true }).click()
+      await expect.poll(() => mock.changes).toEqual([{ inboxID, action: "steer" }])
+    }
+    await expect(view.rows).toHaveCount(0)
+    await expect(pending).toContainText(followUp)
+
+    // The next assistant step still belongs to U1: U2 has been admitted, not delivered.
+    mock.emit("session.step.started", { sessionID, assistantMessageID: assistantID, agent: "build", model })
+    for (const tool of [
+      { id: "tool_queue_read", name: "read", input: { path: "src/queue.ts" } },
+      { id: "tool_queue_grep", name: "grep", input: { pattern: "retry", path: "src" } },
+    ]) {
+      const ref = { sessionID, assistantMessageID: assistantID, id: tool.id }
+      mock.emit("session.tool.input.started", { ...ref, name: tool.name })
+      mock.emit("session.tool.input.ended", { ...ref, text: JSON.stringify(tool.input) })
+      mock.emit("session.tool.called", { ...ref, input: tool.input, executed: true })
+      mock.emit("session.tool.success", {
+        ...ref,
+        content: [{ type: "text", text: "Inspection complete." }],
+        executed: true,
+      })
+    }
+    mock.emit("session.step.ended", {
+      sessionID,
+      assistantMessageID: assistantID,
+      finish: "tool-calls",
+      cost: 0,
+      tokens: { input: 100, output: 20, reasoning: 0, cache: { read: 0, write: 0 } },
+    })
+    const tools = page.locator('[data-timeline-part-ids="tool_queue_read,tool_queue_grep"]')
+    await expect(tools).toBeVisible()
+    await expect(tools).toContainText(/Used\s*Read, Grep/)
+    await expect(tools.locator('[data-component="tag"]')).toHaveText("2")
+    await expect(thinking).toBeVisible()
+    await expect(pending).toBeVisible()
+    expect(mock.rows.map((row) => ({ id: row.id, delivery: row.delivery }))).toEqual([
+      { id: inboxID, delivery: "steer" },
+    ])
+    await transcript.screenshot({ path: testInfo.outputPath("pending-steer.png") })
+
+    // Soft assertions let delivery run too, even when the pending ordering regresses.
+    await expect
+      .soft(tools.or(thinking).or(pending))
+      .toHaveText([/Used\s*Read, Grep/, /Thinking/, /U2: Also check the retry path\./])
+    await expect
+      .soft(transcript.locator('[data-timeline-row="AssistantPart"]').filter({ has: tools }))
+      .toHaveAttribute("data-message-id", userID)
+    await expect
+      .configure({ soft: true })
+      .poll(async () => {
+        const boxes = await Promise.all([tools.boundingBox(), thinking.boundingBox(), pending.boundingBox()])
+        return (
+          boxes.every((box) => box !== null) &&
+          boxes[0]!.y + boxes[0]!.height <= boxes[1]!.y &&
+          boxes[1]!.y + boxes[1]!.height <= boxes[2]!.y
+        )
+      })
+      .toBe(true)
+
+    mock.rows.splice(0, 1)
+    mock.emit("session.inbox.delivered", { sessionID, inboxID })
+    await expect(thinking).toHaveAttribute("data-message-id", inboxID)
+    await expect(pending).toHaveCount(1)
+    await expect(transcript.locator('[data-timeline-row="UserMessage"]')).toHaveCount(2)
+    await expect(transcript.locator('[data-timeline-row="AssistantPart"]').filter({ has: tools })).toHaveAttribute(
+      "data-message-id",
+      userID,
+    )
+
+    const later = { sessionID, assistantMessageID: "msg_queue_follow_up_assistant" }
+    mock.emit("session.step.started", { ...later, agent: "build", model })
+    mock.emit("session.text.started", { ...later, ordinal: 0 })
+    mock.emit("session.text.ended", { ...later, ordinal: 0, text: "A3: Now checking the retry path for U2." })
+    const response = transcript
+      .locator('[data-timeline-row="AssistantPart"]')
+      .filter({ hasText: "A3: Now checking the retry path for U2." })
+    await expect(response).toHaveAttribute("data-message-id", inboxID)
+    await expect(tools.or(pending).or(response).or(thinking)).toHaveText([
+      /Used\s*Read, Grep/,
+      /U2: Also check the retry path\./,
+      /A3: Now checking the retry path for U2\./,
+      /Thinking/,
+    ])
+  })
+}

packages/app/e2e/regression/settings-loading.spec.ts

@@ -0,0 +1,121 @@
+import { expect, test } from "@playwright/test"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+const directory = "C:/Projects/settings-demo"
+const sandboxes = Array.from({ length: 12 }, (_, index) => `${directory}/workspace-${index + 1}`)
+
+test.use({ viewport: { width: 1440, height: 1000 }, colorScheme: "dark" })
+
+test.beforeEach(async ({ page }) => {
+  await mockOpenCodeServer(page, {
+    directory,
+    project: {
+      id: "proj_settings_demo",
+      canonical: directory,
+      name: "Settings demo",
+      vcs: "git",
+      time: { created: 1700000000000, updated: 1700000000000 },
+      sandboxes,
+    },
+    provider: { all: [], connected: [], default: {} },
+    sessions: sandboxes.map((directory, index) => ({
+      id: `ses_settings_${index + 1}`,
+      title: `Workspace ${index + 1} session`,
+      directory,
+      projectID: "proj_settings_demo",
+      time: { created: 1700000000000, updated: 1700000000000 },
+    })),
+    pageMessages: () => ({ items: [] }),
+  })
+  await page.goto("/")
+  await page.getByRole("button", { name: "Settings", exact: true }).click()
+  await expect(page.getByTestId("settings-screen").getByRole("tab", { name: "Preferences" })).toBeVisible()
+})
+
+test("workspaces opens without waiting for inventory or sessions", async ({ page }) => {
+  const inventory = Promise.withResolvers<void>()
+  const sessions = Promise.withResolvers<void>()
+  await page.route("**/api/worktree/*", async (route) => {
+    await inventory.promise
+    await route.fallback()
+  })
+  await page.route("**/api/session?*", async (route) => {
+    if (new URL(route.request().url()).searchParams.has("directory")) await sessions.promise
+    await route.fallback()
+  })
+  const settings = page.getByTestId("settings-screen")
+  const requested = page.waitForRequest((request) => new URL(request.url()).pathname.startsWith("/api/worktree/"))
+  await settings.getByRole("tab", { name: "Workspaces", exact: true }).click()
+  await requested
+  await expect(settings.getByRole("heading", { name: "Workspaces", exact: true })).toBeVisible()
+  await expect(settings.getByRole("button", { name: "Back to app" })).toBeVisible()
+  await expect(settings.getByText("No workspaces", { exact: true })).toHaveCount(0)
+
+  inventory.resolve()
+  await expect(settings.getByText(sandboxes[0], { exact: true })).toBeVisible()
+  await expect(settings.getByText("12 workspaces", { exact: true })).toBeVisible()
+  sessions.resolve()
+  await expect(settings.getByText("Workspace 1 session", { exact: true })).toBeVisible()
+
+  const refresh = Promise.withResolvers<void>()
+  await page.route("**/api/worktree/*", async (route) => {
+    await refresh.promise
+    await route.fallback()
+  })
+  await settings.getByRole("tab", { name: "Preferences", exact: true }).click()
+  await settings.getByRole("tab", { name: "Workspaces", exact: true }).click()
+  await expect(settings.getByText("Workspace 1 session", { exact: true })).toBeVisible()
+  refresh.resolve()
+})
+
+test("extensions opens without waiting for MCPs or plugins", async ({ page }) => {
+  const mcps = Promise.withResolvers<void>()
+  const plugins = Promise.withResolvers<void>()
+  await page.route("**/api/mcp", async (route) => {
+    await mcps.promise
+    await route.fulfill({
+      json: { location: { directory }, data: [{ name: "demo-mcp", status: { status: "connected" } }] },
+    })
+  })
+  await page.route("**/api/plugin", async (route) => {
+    await plugins.promise
+    await route.fulfill({
+      json: {
+        location: { directory },
+        data: [
+          { id: "demo-plugin", source: { type: "package", package: "demo-plugin" }, status: "active", tui: false },
+        ],
+      },
+    })
+  })
+  const settings = page.getByTestId("settings-screen")
+  const requested = page.waitForRequest((request) => new URL(request.url()).pathname === "/api/mcp")
+  await settings.getByRole("tab", { name: "Extensions", exact: true }).click()
+  await requested
+  await expect(settings.getByRole("heading", { name: "Extensions", exact: true })).toBeVisible()
+  await expect(settings.getByRole("button", { name: "Back to app" })).toBeVisible()
+  await settings.getByRole("tab", { name: "Plugins", exact: true }).click()
+  await expect(settings.getByRole("tab", { name: "Plugins", exact: true })).toHaveAttribute("aria-selected", "true")
+  plugins.resolve()
+  await expect(settings.getByText("demo-plugin", { exact: true })).toBeVisible()
+  mcps.resolve()
+  await settings.getByRole("tab", { name: "MCPs", exact: true }).click()
+  await expect(settings.getByRole("switch", { name: "demo-mcp" })).toBeChecked()
+})
+
+test("workspace inventory uses the settings panel scroll area", async ({ page }) => {
+  const settings = page.getByTestId("settings-screen")
+  await settings.getByRole("tab", { name: "Workspaces", exact: true }).click()
+  await expect(settings.getByText("Workspace 1 session", { exact: true })).toBeVisible()
+  const list = settings.locator('[data-component="settings-list"]')
+  await expect(list).toHaveCSS("max-height", "none")
+  await expect(list).toHaveCSS("overflow-y", "visible")
+  await settings.getByText("Workspace 12 session", { exact: true }).scrollIntoViewIfNeeded()
+  await expect(settings.getByText("Workspace 12 session", { exact: true })).toBeInViewport()
+  await expect(settings.getByRole("button", { name: "Back to app" })).toBeInViewport()
+  await page.setViewportSize({ width: 390, height: 844 })
+  await expect(list).toHaveCSS("max-height", "none")
+  await expect(list).toHaveCSS("overflow-y", "visible")
+  await settings.getByText("Workspace 12 session", { exact: true }).scrollIntoViewIfNeeded()
+  await expect(settings.getByText("Workspace 12 session", { exact: true })).toBeInViewport()
+})

packages/app/e2e/regression/tab-navigate-mousedown.spec.ts

@@ -2,7 +2,7 @@ import { expect, test, type Page, type Route } from "@playwright/test"
 import { base64Encode } from "@opencode-ai/util/encode"
 import { currentSession } from "../utils/mock-server"
 
-const server = "http://127.0.0.1:4096"
+const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`
 const sessionA = session("ses_tab_a", "Tab A session")
 const sessionB = session("ses_tab_b", "Tab B session")
 const sessionC = session("ses_tab_c", "Tab C session")
@@ -238,7 +238,7 @@ function session(id: string, title: string) {
 
 async function mockServer(page: Page) {
   const sessions = [sessionA, sessionB, sessionC]
-  await page.route("**/*", async (route) => {
+  await page.route("**/api/**", async (route) => {
     const url = new URL(route.request().url())
     if (url.origin !== server) return route.fallback()
     if (url.pathname === `/api/session/${unresolvedSessionID}`) return new Promise(() => {})

packages/app/e2e/utils/mock-server.ts

@@ -144,12 +144,14 @@ export async function mockOpenCodeServer(page: Page, config: MockServerConfig) {
   )
   page.on("close", () => void transport.dispose())
 
-  await page.route("**/*", async (route) => {
+  await page.route("**/api/**", async (route) => {
     const url = new URL(route.request().url())
     const appPort = new URL(
       process.env.PLAYWRIGHT_BASE_URL ?? `http://127.0.0.1:${process.env.PLAYWRIGHT_PORT ?? "3000"}`,
     ).port
     if (url.origin !== server && url.port !== appPort) return route.fallback()
+    // Production serves the UI and API from one origin; leave app assets to Vite.
+    if (!url.pathname.startsWith("/api/")) return route.fallback()
     if (route.request().method() === "OPTIONS") {
       return route.fulfill({ status: 204, headers: corsHeaders })
     }

packages/app/package.json

@@ -25,6 +25,7 @@
     "test:browser": "bun test --conditions=browser --preload ./happydom.ts ./test-browser",
     "test:unit:watch": "bun test --conditions=solid --watch --preload ./happydom.ts ./src",
     "test:e2e": "playwright test",
+    "test:e2e:built": "PLAYWRIGHT_BUILD=1 playwright test",
     "test:e2e:local": "playwright test",
     "test:components": "playwright test --config playwright.components.config.ts",
     "test:components:ui": "playwright test --config playwright.components.config.ts --ui",

packages/app/playwright.config.ts

@@ -2,10 +2,19 @@ import { defineConfig, devices } from "@playwright/test"
 
 const port = Number(process.env.PLAYWRIGHT_PORT ?? 3000)
 const baseURL = process.env.PLAYWRIGHT_BASE_URL ?? `http://127.0.0.1:${port}`
+const url = new URL(baseURL)
+if (url.protocol !== "http:") throw new Error("E2E fixtures require an http:// app URL")
+const built = !!process.env.CI || process.env.PLAYWRIGHT_BUILD === "1"
+// Production connects to its own origin, so fixture URLs must match the preview server.
+if (built) {
+  process.env.PLAYWRIGHT_SERVER_HOST = url.hostname
+  process.env.PLAYWRIGHT_SERVER_PORT = url.port || "80"
+}
 const serverHost = process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"
 const serverPort = process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"
-const command = `bun run dev -- --host 0.0.0.0 --port ${port}`
-const reuse = !process.env.CI
+const command = built
+  ? `bun run build && bun run serve -- --host 127.0.0.1 --port ${port} --strictPort`
+  : `bun run dev -- --host 127.0.0.1 --port ${port} --strictPort`
 const workers = Number(process.env.PLAYWRIGHT_WORKERS ?? (process.env.CI ? 5 : 0)) || undefined
 export default defineConfig({
   testDir: "./e2e",
@@ -23,16 +32,18 @@ export default defineConfig({
   retries: process.env.CI ? 2 : 0,
   workers,
   reporter: [["html", { outputFolder: "e2e/playwright-report", open: "never" }], ["line"]],
-  webServer: {
-    command,
-    url: baseURL,
-    reuseExistingServer: reuse,
-    timeout: 120_000,
-    env: {
-      VITE_OPENCODE_SERVER_HOST: serverHost,
-      VITE_OPENCODE_SERVER_PORT: serverPort,
-    },
-  },
+  webServer: process.env.PLAYWRIGHT_BASE_URL
+    ? undefined
+    : {
+        command,
+        url: baseURL,
+        reuseExistingServer: !built,
+        timeout: 120_000,
+        env: {
+          VITE_OPENCODE_SERVER_HOST: serverHost,
+          VITE_OPENCODE_SERVER_PORT: serverPort,
+        },
+      },
   use: {
     baseURL,
     trace: "on-first-retry",

packages/app/src/home/route.tsx

@@ -17,7 +17,7 @@ export function Home() {
   return (
     <div
       class={`
-        m-2 min-h-0 flex-1 self-stretch overflow-hidden rounded-[10px]
+        mx-2 mb-2 mt-[var(--shell-top-inset,8px)] min-h-0 flex-1 self-stretch overflow-hidden rounded-[10px]
         bg-v2-background-bg-base shadow-[var(--v2-elevation-raised)]
       `}
     >

packages/app/src/new-session/screen.tsx

@@ -73,7 +73,7 @@ export default function NewSessionPage(props: { draftId: string }) {
     <div class="relative size-full overflow-hidden flex flex-col">
       {suspendUntilPromptReady()}
       <NewSessionStatus visible={settings.visibility.status()} />
-      <div class="flex-1 min-h-0 flex flex-col gap-2 p-2">
+      <div class="flex-1 min-h-0 flex flex-col gap-2 px-2 pb-2 pt-[var(--shell-top-inset,8px)]">
         <NewSessionView composer={model} project={project} workspace={workspace} />
       </div>
     </div>

packages/app/src/runtime/server/global-sync/utils.ts

@@ -1,10 +1,16 @@
 import type { AgentListOutput, ModelListOutput, ProviderListOutput } from "@opencode-ai/client/promise"
 import type { Agent, Project, Provider, ProviderListResponse } from "@/runtime/server/types"
 import type { Project as CurrentProject } from "@opencode-ai/client/promise"
+import { unwrap } from "solid-js/store"
 export { pathKey as directoryKey, type PathKey as DirectoryKey } from "@/workspaces/path-key"
 
 export const cmp = (a: string, b: string) => (a < b ? -1 : a > b ? 1 : 0)
 
+const providerCatalogs = new WeakMap<
+  ProviderListOutput["data"],
+  WeakMap<ModelListOutput["data"], ProviderListResponse>
+>()
+
 export function normalizeAgentList(input: AgentListOutput["data"] | Agent[]): Agent[] {
   if (input.every((agent) => !("request" in agent))) return input as Agent[]
   return (input as AgentListOutput["data"]).map((agent) => ({
@@ -30,12 +36,16 @@ export function normalizeAgentList(input: AgentListOutput["data"] | Agent[]): Ag
 }
 
 export function normalizeProviderList(
-  providers: ProviderListOutput["data"] | ProviderListResponse,
-  models?: ModelListOutput["data"],
+  input: ProviderListOutput["data"] | ProviderListResponse,
+  catalog?: ModelListOutput["data"],
 ): ProviderListResponse {
-  if (!Array.isArray(providers)) {
-    return providers
-  }
+  if (!Array.isArray(input)) return input
+  // Client sync replaces whole catalog lists. Track those reads at the caller,
+  // not every model field, and share conversions without retaining old lists.
+  const providers = unwrap(input)
+  const models = unwrap(catalog)
+  const cached = models && providerCatalogs.get(providers)?.get(models)
+  if (cached) return cached
   const all = new Map<string, Provider>()
 
   for (const provider of providers) {
@@ -101,7 +111,7 @@ export function normalizeProviderList(
     }
   }
 
-  return {
+  const result = {
     all,
     connected: providers.map((provider) => provider.id),
     default: Object.fromEntries(
@@ -111,6 +121,12 @@ export function normalizeProviderList(
       }),
     ),
   }
+  if (models) {
+    const cache = providerCatalogs.get(providers) ?? new WeakMap<ModelListOutput["data"], ProviderListResponse>()
+    cache.set(models, result)
+    providerCatalogs.set(providers, cache)
+  }
+  return result
 }
 
 export function normalizeProjectInfo(project: Project | CurrentProject): Project {

packages/app/src/session/route.tsx

@@ -44,9 +44,7 @@ function TargetSessionSettingsCommand() {
   return null
 }
 
-function SessionRouteErrorBoundary(
-  props: ParentProps<{ sessionID?: string; serverKey?: ServerConnection.Key }>,
-) {
+function SessionRouteErrorBoundary(props: ParentProps<{ sessionID?: string; serverKey?: ServerConnection.Key }>) {
   return (
     <ErrorBoundary
       fallback={(error) => (
@@ -106,7 +104,7 @@ function PendingSessionState(props: { sessionID: string }) {
 
 function SessionStatePanel(props: ParentProps) {
   return (
-    <div class="flex min-h-0 flex-1 p-2">
+    <div class="flex min-h-0 flex-1 px-2 pb-2 pt-[var(--shell-top-inset,8px)]">
       <SessionPanelFrame raised>{props.children}</SessionPanelFrame>
     </div>
   )

packages/app/src/session/screen-layout.ts

@@ -8,7 +8,7 @@ import type { SessionModel } from "./model"
 import { sessionPanelLayout } from "./session-panel-layout"
 import { clampSessionPanelWidth, sessionPanelWidthMax } from "./session-panel-width"
 
-export function createSessionScreenLayout(session: SessionModel, serverScope: string) {
+export function createSessionScreenLayout(session: SessionModel) {
   const layout = useLayout()
   const settings = useSettings()
   const size = createSizing()
@@ -92,7 +92,6 @@ export function createSessionScreenLayout(session: SessionModel, serverScope: st
     centered: createMemo(() => session.isDesktop()),
     files: { open: fileTreeOpen },
     panel: {
-      key: createMemo(() => (session.identity.params.id ? `${serverScope}\0${session.identity.params.id}` : undefined)),
       max: panelMax,
       ref: (element: HTMLDivElement) => {
         row = element

packages/app/src/session/screen.tsx

@@ -4,7 +4,6 @@ import createPresence from "solid-presence"
 import { ResizeHandle } from "@opencode-ai/ui/resize-handle"
 import { SessionHeader } from "@/session/header/session-header"
 import { useLayout } from "@/shell/state/layout"
-import { useServerSDK } from "@/runtime/server/client"
 import { useSettings } from "@/settings/model"
 import { MessageTimeline } from "@/session/timeline/message-timeline"
 import type { SessionModel } from "@/session/model"
@@ -23,10 +22,9 @@ import { SessionIdentityHeader } from "./session-identity-header"
 export function SessionScreen(props: { session: SessionModel }) {
   const session = props.session
   const layout = useLayout()
-  const serverSDK = useServerSDK()
   const settings = useSettings()
   const isDesktop = session.isDesktop
-  const screen = createSessionScreenLayout(session, serverSDK.scope)
+  const screen = createSessionScreenLayout(session)
   const timeline = createSessionTimelineInteraction(session)
   const messagesReady = timeline.ready
   const [store, setStore] = createStore({
@@ -164,7 +162,7 @@ export function SessionScreen(props: { session: SessionModel }) {
   return (
     <>
       <SessionHeader />
-      <div class="flex-1 min-h-0 flex flex-col gap-2 p-2">
+      <div class="flex-1 min-h-0 flex flex-col gap-2 px-2 pb-2 pt-[var(--shell-top-inset,8px)]">
         <div ref={screen.panel.ref} class="relative flex-1 min-h-0 flex flex-col md:flex-row gap-2">
           <div
             classList={{
@@ -177,12 +175,10 @@ export function SessionScreen(props: { session: SessionModel }) {
               width: screen.panel.width(),
             }}
           >
-            <Show when={screen.panel.key()} keyed>
-              {(_) => (
-                <SessionPanelFrame raised={!!session.identity.params.id}>
-                  <ErrorBoundary fallback={sessionErrorFallback}>{sessionPanelContent()}</ErrorBoundary>
-                </SessionPanelFrame>
-              )}
+            <Show when={!!session.identity.params.id}>
+              <SessionPanelFrame raised>
+                <ErrorBoundary fallback={sessionErrorFallback}>{sessionPanelContent()}</ErrorBoundary>
+              </SessionPanelFrame>
             </Show>
 
             <Show when={screen.panel.resizable()}>

packages/app/src/session/session-frame.tsx

@@ -2,7 +2,10 @@ import type { ParentProps } from "solid-js"
 
 export function SessionRouteFrame(props: ParentProps<{ padded?: boolean }>) {
   return (
-    <div class="relative flex size-full flex-col overflow-hidden" classList={{ "p-2": props.padded }}>
+    <div
+      class="relative flex size-full flex-col overflow-hidden"
+      classList={{ "px-2 pb-2 pt-[var(--shell-top-inset,8px)]": props.padded }}
+    >
       {props.children}
     </div>
   )

packages/app/src/session/timeline/controller-projection.test.ts

@@ -1,6 +1,8 @@
 import { describe, expect, test } from "bun:test"
 import type { SessionInboxInfo, SessionMessageInfo } from "@opencode-ai/client/promise"
+import { createRoot } from "solid-js"
 import { applyTimelineMessageHandoff, visibleTimelineMessages } from "./controller-projection"
+import { createTimelineProjection } from "./projection"
 
 const messages = [
   { id: "msg_1", type: "user", text: "first", time: { created: 1 } },
@@ -17,6 +19,105 @@ const messages = [
 ] satisfies SessionMessageInfo[]
 
 describe("visibleTimelineMessages", () => {
+  const steer = {
+    id: "msg_3",
+    sessionID: "ses_1",
+    timeCreated: 3,
+    type: "user",
+    delivery: "steer",
+    payload: { text: "queued" },
+  } satisfies SessionInboxInfo
+  const work = {
+    id: "msg_5",
+    type: "assistant",
+    agent: "build",
+    model: { id: "model", providerID: "provider" },
+    content: [
+      {
+        type: "tool",
+        id: "tool_read",
+        name: "read",
+        state: {
+          status: "completed",
+          input: { filePath: "src/example.ts" },
+          content: [{ type: "text", text: "export const example = true" }],
+          metadata: {},
+        },
+        time: { created: 5, completed: 6 },
+      },
+    ],
+    time: { created: 5, completed: 6 },
+  } satisfies SessionMessageInfo
+
+  test("keeps work and thinking above an undelivered steer", () => {
+    const source = [...messages.slice(0, 3), work]
+    const visible = visibleTimelineMessages(source, [steer])
+    expect(visible.map((message) => message.id)).toEqual(["msg_1", "msg_2", "msg_5", "msg_3"])
+    expect(source.map((message) => message.id)).toEqual(["msg_1", "msg_2", "msg_3", "msg_5"])
+    expect(visible[2]).toBe(work)
+
+    createRoot((dispose) => {
+      const projection = createTimelineProjection({
+        sessionMessages: () => visible,
+        status: () => ({ type: "busy" }),
+        showReasoningSummaries: () => false,
+        shellToolDefaultOpen: () => false,
+        editToolDefaultOpen: () => false,
+        pendingUserMessageIDs: () => new Set([steer.id]),
+      })
+      expect(projection.activeMessageID()).toBe("msg_1")
+      expect(projection.rows().map((row) => [row._tag, row.userMessageID])).toEqual([
+        ["UserMessage", "msg_1"],
+        ["AssistantPart", "msg_1"],
+        ["Thinking", "msg_1"],
+        ["TurnGap", "msg_3"],
+        ["UserMessage", "msg_3"],
+      ])
+      expect(
+        projection
+          .assistantMessagesByParent()
+          .get("msg_1")
+          ?.map((message) => message.id),
+      ).toEqual(["msg_2", "msg_5"])
+      expect(projection.assistantMessagesByParent().has(steer.id)).toBe(false)
+      dispose()
+    })
+  })
+
+  test("moves a queued input after existing work when changed to steer", () => {
+    const source = [...messages.slice(0, 3), work]
+    expect(visibleTimelineMessages(source, [{ ...steer, delivery: "queue" }]).map((message) => message.id)).toEqual([
+      "msg_1",
+      "msg_2",
+      "msg_5",
+    ])
+    expect(visibleTimelineMessages(source, [steer]).map((message) => message.id)).toEqual([
+      "msg_1",
+      "msg_2",
+      "msg_5",
+      "msg_3",
+    ])
+    const delivered = [messages[0], messages[1], work, messages[2]]
+    expect(visibleTimelineMessages(delivered, [])).toBe(delivered)
+  })
+
+  test("preserves steer order and excludes reverted steers", () => {
+    const source = [...messages, work]
+    const pending = [steer, { ...steer, id: "msg_4" }]
+    expect(visibleTimelineMessages(source, pending).map((message) => message.id)).toEqual([
+      "msg_1",
+      "msg_2",
+      "msg_5",
+      "msg_3",
+      "msg_4",
+    ])
+    expect(visibleTimelineMessages(source, pending, "msg_4").map((message) => message.id)).toEqual([
+      "msg_1",
+      "msg_2",
+      "msg_3",
+    ])
+  })
+
   test("hides queued inputs until delivery", () => {
     const pending = [
       {

packages/app/src/session/timeline/controller-projection.ts

@@ -17,8 +17,19 @@ export function visibleTimelineMessages(
   const queued = new Set(
     pending.flatMap((item) => (item.type === "user" && item.delivery === "queue" ? [item.id] : [])),
   )
-  if (queued.size === 0 && !revertMessageID) return messages
-  return messages.filter((message) => !queued.has(message.id) && (!revertMessageID || message.id < revertMessageID))
+  const steers = new Set(
+    pending.flatMap((item) => (item.type === "user" && item.delivery === "steer" ? [item.id] : [])),
+  )
+  if (queued.size === 0 && steers.size === 0 && !revertMessageID) return messages
+  const visible = messages.filter(
+    (message) => !queued.has(message.id) && (!revertMessageID || message.id < revertMessageID),
+  )
+  if (steers.size === 0) return visible
+  // Pending steers do not own assistant work until they are delivered.
+  return [
+    ...visible.filter((message) => !steers.has(message.id)),
+    ...visible.filter((message) => steers.has(message.id)),
+  ]
 }
 
 export function timelineChildTitle(input: {

packages/app/src/session/timeline/message-timeline.tsx

@@ -481,7 +481,6 @@ function MessageTimelineView(
   const backgroundHintPresence = createAnimatedPresence(backgroundHintPartID, () => backgroundHintRef() ?? null)
   return (
     <VirtualizedTimeline
-      workspaceSession={workspaceSession}
       bottomSpacer={
         <Show when={backgroundHintPresence.present()}>
           <div

packages/app/src/session/timeline/virtualizer.tsx

@@ -1,8 +1,15 @@
-import { createVirtualizer, defaultRangeExtractor, elementScroll, type VirtualItem } from "@tanstack/solid-virtual"
+import {
+  createVirtualizer,
+  defaultRangeExtractor,
+  elementScroll,
+  type Range,
+  type VirtualItem,
+} from "@tanstack/solid-virtual"
 import { isScrollKeyTarget, scrollKey, scrollKeyOwner, ScrollView } from "@opencode-ai/ui/scroll-view"
 import { TimelineRow } from "@opencode-ai/session-ui/timeline/projection"
 import { useLanguage } from "@/runtime/i18n/language"
 import {
+  batch,
   createEffect,
   createMemo,
   createSignal,
@@ -55,7 +62,6 @@ type Input = {
 type ViewProps = {
   header: JSX.Element
   bottomSpacer?: JSX.Element
-  workspaceSession: Accessor<boolean>
   deferred: (row: TimelineRow.TimelineRow) => boolean
   renderRow: (row: Accessor<TimelineRow.TimelineRow>, onSizeChange?: () => void) => JSX.Element
 }
@@ -68,11 +74,12 @@ export function createTimelineVirtualizer(input: Input) {
   const coldBottomMount = !initialMeasurements?.length && input.pinned()
   const [listRoot, setListRoot] = createSignal<HTMLDivElement>()
   const [toolOpen, setToolOpen] = createStore<Record<string, boolean | undefined>>(cached?.toolOpen ?? {})
-  const [renderOverscan, setRenderOverscan] = createSignal(initialMeasurements?.length || coldBottomMount ? 6 : 20)
+  const [overscan, setOverscan] = createSignal(2)
   const rows = input.projection.rows
   const rowByKey = input.projection.rowByKey
   const knownKeys = new Set(rows().map(TimelineRow.key))
   const addedKeys = new Set<string>()
+  const measuredElements = new WeakSet<Element>()
   let touchStart: number | undefined
   let pointerHeld = false
   let maxScroll = 0
@@ -92,14 +99,19 @@ export function createTimelineVirtualizer(input: Input) {
     initialOffset: () => (input.pinned() ? Number.MAX_SAFE_INTEGER : 0),
     initialMeasurementsCache: initialMeasurements,
     estimateSize: () => fallbackItemSize,
-    // Do not replace this with TanStack's default measurer: without a ResizeObserver entry,
-    // it returns the cached height instead of reading the element (TanStack/virtual#1183).
-    // Restored sessions, deferred tools, and rewrapped content can then keep stale heights;
-    // our fixed-height, overflow-clipped rows will hide their content. Keep observer entries
-    // on the cheap precomputed path, but make explicit measurements read the real height.
-    measureElement: (element, entry) => {
+    // A newly observed element gets a real ResizeObserver box before paint. Reuse
+    // its snapshot on attachment, but later explicit measurements must read layout
+    // so deferred/rewrapped content cannot keep stale, clipped heights (TanStack/virtual#1183).
+    measureElement: (element, entry, instance) => {
+      const initial = !measuredElements.has(element)
+      measuredElements.add(element)
       const box = entry?.borderBoxSize[0]
-      return box ? Math.round(box.blockSize) : element.offsetHeight
+      if (box) return Math.round(box.blockSize)
+      if (initial) {
+        const size = instance.itemSizeCache.get(instance.options.getItemKey(instance.indexFromElement(element)))
+        if (size !== undefined) return size
+      }
+      return element.offsetHeight
     },
     scrollToFn: (offset, options, instance) => {
       if (virtualContent) virtualContent.style.height = `${instance.getTotalSize()}px`
@@ -131,41 +143,59 @@ export function createTimelineVirtualizer(input: Input) {
       return input.showHeader() ? 64 : 0
     },
     paddingEnd: 64,
-    rangeExtractor: (range) => {
+    get rangeExtractor() {
       const id = input.projection.activeMessageID()
       const active = id ? (input.projection.messageLastRowIndex().get(id) ?? -1) : -1
-      const indexes = defaultRangeExtractor({ ...range, overscan: renderOverscan() })
-      return filterVirtualIndexes(
-        [...new Set([...indexes, ...(active < 0 ? [] : [active])])].sort((a, b) => a - b),
-        range.count,
-      )
+      const buffer = overscan()
+      return (range: Range) => {
+        const indexes = defaultRangeExtractor({ ...range, overscan: buffer })
+        return filterVirtualIndexes(
+          [...new Set([...indexes, ...(active < 0 ? [] : [active])])].sort((a, b) => a - b),
+          range.count,
+        )
+      }
     },
   })
   const resizeItem = virtualizer.resizeItem
-  let resizeAnchorScheduled = false
-  // Rows measure asynchronously, so the last row can still hold its estimate when TanStack
-  // reconciles the end. Coalesce one correction per measurement batch, before paint.
-  const anchorResizedBottom = () => {
-    if (resizeAnchorScheduled) return
-    resizeAnchorScheduled = true
+  const pendingSizes = new Map<number, { key: string; size: number }>()
+  let resizeScheduled = false
+  // Read the whole measurement delivery before committing reactive row sizes.
+  // Otherwise each row can render and force layout before the next is measured.
+  virtualizer.resizeItem = (index, size) => {
+    const row = rows()[index]
+    if (!row) return
+    const key = TimelineRow.key(row)
+    if (virtualizer.itemSizeCache.get(key) === size) {
+      pendingSizes.delete(index)
+      return
+    }
+    pendingSizes.set(index, { key, size })
+    if (resizeScheduled) return
+    resizeScheduled = true
     queueMicrotask(() => {
-      resizeAnchorScheduled = false
+      resizeScheduled = false
+      if (!pendingSizes.size) return
+      const sizes = [...pendingSizes]
+      pendingSizes.clear()
+      batch(() => {
+        sizes.forEach(([index, value]) => {
+          const row = rows()[index]
+          if (row && TimelineRow.key(row) === value.key) resizeItem(index, value.size)
+        })
+      })
       if (!input.pinned()) return
-      virtualizer.scrollToEnd()
+      const root = listRoot()
+      // Reopening a settled scroll-to-end operation can fight subsequent keyboard scrolling.
+      if (root && Math.abs(root.scrollHeight - root.clientHeight - root.scrollTop) > endEpsilon)
+        virtualizer.scrollToEnd()
     })
   }
-  virtualizer.resizeItem = (index, size) => {
-    resizeItem(index, size)
-    if (listRoot() && input.pinned()) anchorResizedBottom()
-  }
+  onCleanup(() => pendingSizes.clear())
   virtualizer.shouldAdjustScrollPositionOnItemSizeChange = (item, _delta, instance) => {
     // Prepended rows can resize more than once as deferred content mounts. Keep
     // compensating while they remain entirely above the visible content fold.
     if (addedKeys.has(String(item.key)))
-      return (
-        item.end <=
-        (instance.scrollOffset ?? 0) + instance.scrollAdjustments + instance.options.scrollMargin
-      )
+      return item.end <= (instance.scrollOffset ?? 0) + instance.scrollAdjustments + instance.options.scrollMargin
     const first = instance.range?.startIndex
     return first !== undefined && item.index < first
   }
@@ -186,50 +216,41 @@ export function createTimelineVirtualizer(input: Input) {
     })
   })
 
+  let settleFrame: number | undefined
   let overscanFrame: number | undefined
+  let overscanTimer: number | undefined
+  const expandOverscan = () => {
+    overscanFrame = requestAnimationFrame(() => {
+      overscanFrame = undefined
+      // Let the visible rows paint before building the normal interaction buffer.
+      overscanTimer = window.setTimeout(() => {
+        overscanTimer = undefined
+        setOverscan(20)
+      }, 0)
+    })
+  }
   const pendingMeasurements = () =>
     virtualizer.getVirtualItems().some((item) => !virtualizer.itemSizeCache.has(item.key))
   const settleColdBottom = () => {
     if (input.pinned()) virtualizer.scrollToEnd()
     if (virtualContent?.querySelector(pendingMarkdown) || pendingMeasurements()) {
-      overscanFrame = requestAnimationFrame(settleColdBottom)
+      settleFrame = requestAnimationFrame(settleColdBottom)
       return
     }
-    overscanFrame = requestAnimationFrame(() => {
+    settleFrame = requestAnimationFrame(() => {
       if (input.pinned()) virtualizer.scrollToEnd()
       if (virtualContent?.querySelector(pendingMarkdown) || pendingMeasurements()) {
         settleColdBottom()
         return
       }
-      overscanFrame = undefined
-      const content = virtualContent
-      if (!content) return
-      if (window.matchMedia("(prefers-reduced-motion: reduce)").matches) {
-        content.style.removeProperty("visibility")
-        return
-      }
-      const animation = ["animate-in", "fade-in", "duration-150"]
-      const clearAnimation = (event: AnimationEvent) => {
-        if (event.target !== content) return
-        content.removeEventListener("animationend", clearAnimation)
-        content.removeEventListener("animationcancel", clearAnimation)
-        content.classList.remove(...animation)
-      }
-      content.addEventListener("animationend", clearAnimation)
-      content.addEventListener("animationcancel", clearAnimation)
-      content.classList.add(...animation)
-      content.style.removeProperty("visibility")
+      settleFrame = undefined
+      virtualContent?.style.removeProperty("visibility")
+      expandOverscan()
     })
   }
   onMount(() => {
-    overscanFrame = requestAnimationFrame(() => {
-      if (renderOverscan() < 20) setRenderOverscan(20)
-      if (!coldBottomMount) {
-        overscanFrame = undefined
-        return
-      }
-      settleColdBottom()
-    })
+    if (coldBottomMount) settleFrame = requestAnimationFrame(settleColdBottom)
+    if (!coldBottomMount) expandOverscan()
   })
 
   let measuredSessionKey = input.sessionKey()
@@ -256,11 +277,13 @@ export function createTimelineVirtualizer(input: Input) {
   const handleListWheel = (event: WheelEvent & { currentTarget: HTMLDivElement }) => {
     input.onUserScroll(event.target)
     if (event.deltaY < 0) input.onUnpin()
+    setOverscan(20)
   }
 
   const handleListTouchStart = (event: TouchEvent) => {
     input.onUserScroll(event.target)
     touchStart = event.touches[0]?.clientY
+    setOverscan(20)
   }
 
   const handleListTouchMove = (event: TouchEvent & { currentTarget: HTMLDivElement }) => {
@@ -277,14 +300,19 @@ export function createTimelineVirtualizer(input: Input) {
   const handleListPointerDown = (event: PointerEvent & { currentTarget: HTMLDivElement }) => {
     input.onUserScroll(event.target)
     pointerHeld = true
-    const release = () => {
-      pointerHeld = false
-      window.removeEventListener("pointerup", release)
-      window.removeEventListener("pointercancel", release)
-    }
-    window.addEventListener("pointerup", release)
-    window.addEventListener("pointercancel", release)
+    setOverscan(20)
   }
+  const releasePointer = () => {
+    pointerHeld = false
+  }
+  onMount(() => {
+    window.addEventListener("pointerup", releasePointer)
+    window.addEventListener("pointercancel", releasePointer)
+  })
+  onCleanup(() => {
+    window.removeEventListener("pointerup", releasePointer)
+    window.removeEventListener("pointercancel", releasePointer)
+  })
 
   const handleListKeyDown = (event: KeyboardEvent & { currentTarget: HTMLDivElement }) => {
     const key = scrollKey(event)
@@ -293,6 +321,7 @@ export function createTimelineVirtualizer(input: Input) {
     if (scrollKeyOwner(event.currentTarget, event.target, key) !== event.currentTarget) return
     input.onUserScroll(event.currentTarget)
     if (upwardKeys.has(key)) input.onUnpin()
+    setOverscan(20)
   }
 
   // Following resumes by arriving at the end, either by scrolling there or by content shrinking
@@ -371,7 +400,7 @@ export function createTimelineVirtualizer(input: Input) {
     }
 
     return (
-      <div class="relative w-full h-full min-w-0" data-workspace-session={props.workspaceSession() ? "" : undefined}>
+      <div class="relative w-full h-full min-w-0">
         <div
           class="absolute left-1/2 -translate-x-1/2 z-[60] pointer-events-none transition-all duration-200 ease-out"
           classList={{
@@ -415,7 +444,6 @@ export function createTimelineVirtualizer(input: Input) {
           <Show when={input.showHeader()}>{props.header}</Show>
           <div
             data-timeline-virtual-content
-            class="motion-reduce:animate-none"
             ref={(element) => {
               virtualContent = element
               input.setContentRef(element)
@@ -447,7 +475,9 @@ export function createTimelineVirtualizer(input: Input) {
     cache.delete(ownerSessionKey)
     cache.set(ownerSessionKey, { measurements: virtualizer.takeSnapshot(), toolOpen: { ...toolOpen } })
     while (cache.size > 16) cache.delete(cache.keys().next().value!)
+    if (settleFrame !== undefined) cancelAnimationFrame(settleFrame)
     if (overscanFrame !== undefined) cancelAnimationFrame(overscanFrame)
+    if (overscanTimer !== undefined) window.clearTimeout(overscanTimer)
     input.setScrollRef(undefined)
     input.setRevealMessage?.(() => {})
     input.setScrollToEnd?.(() => {})

packages/app/src/settings/providers/extensions.tsx

@@ -27,6 +27,7 @@ export const SettingsExtensions: Component = () => {
   const [mcpList, { refetch: refetchMcp }] = createResource(
     () => serverSdk.connection.status() === "connected",
     () => serverSdk.api.mcp.list().then((result) => result.data),
+    { initialValue: [] },
   )
   const toggleMcp = useMcpToggle(() => undefined, refetchMcp)
   const mcps = createMemo<McpRowItem[]>(() => {
@@ -44,6 +45,7 @@ export const SettingsExtensions: Component = () => {
   const [pluginList] = createResource(
     () => serverSdk.connection.status() === "connected",
     () => serverSdk.api.plugin.list().then((result) => result.data),
+    { initialValue: [] },
   )
   const plugins = createMemo<PluginRowItem[]>(() => pluginLabels(pluginList.latest ?? []).map((name) => ({ name })))
 

packages/app/src/settings/settings.css

@@ -885,8 +885,6 @@
   display: flex;
   flex-direction: column;
   gap: 0;
-  max-height: 480px;
-  overflow-y: auto;
   padding: 20px;
   border-radius: 6px;
   background-color: var(--v2-background-bg-base);
@@ -1048,8 +1046,6 @@
   }
 
   .settings-workspaces-inventory [data-component="settings-list"] {
-    max-height: none;
-    overflow-y: visible;
     padding: 14px;
   }
 
@@ -1082,8 +1078,6 @@
   }
 
   .settings-workspaces-inventory [data-component="settings-list"] {
-    max-height: none;
-    overflow-y: visible;
     padding: 14px;
   }
 

packages/app/src/settings/workspaces/project-extensions.tsx

@@ -97,10 +97,12 @@ export const ProjectSettingsExtensions: Component = () => {
   const [globalPluginList] = createResource(
     () => serverSDK.connection.status() === "connected",
     () => serverSDK.api.plugin.list().then((result) => result.data),
+    { initialValue: [] },
   )
   const [projectPluginList] = createResource(
     () => (serverSDK.connection.status() === "connected" ? directorySDK().directory : undefined),
     (directory) => serverSDK.api.plugin.list({ location: { directory } }).then((result) => result.data),
+    { initialValue: [] },
   )
   const globalPlugins = createMemo(() => pluginLabels(globalPluginList.latest ?? []))
   const projectPlugins = createMemo(() => {

packages/app/src/settings/workspaces/workspaces.tsx

@@ -60,6 +60,7 @@ export const SettingsWorkspaces: Component<{ activeDirectory?: string }> = (prop
 
   const projectQuery = useQuery(() => ({
     queryKey: [serverSDK.scope, "settings-workspace-projects"] as const,
+    enabled: serverSDK.connection.status() === "connected",
     queryFn: async () =>
       Promise.all(
         (await serverSDK.api.project.list()).map(async (project) => {
@@ -71,10 +72,9 @@ export const SettingsWorkspaces: Component<{ activeDirectory?: string }> = (prop
       ),
     refetchOnMount: "always",
   }))
-  const workspaces = createMemo(() => workspaceInventory(projectQuery.data ?? []))
-  const projects = createMemo(() =>
-    (projectQuery.data ?? []).filter((project) => managedWorkspaceDirectories(project).length > 0),
-  )
+  const inventory = createMemo(() => (projectQuery.isPending ? [] : (projectQuery.data ?? [])))
+  const workspaces = createMemo(() => workspaceInventory(inventory()))
+  const projects = createMemo(() => inventory().filter((project) => managedWorkspaceDirectories(project).length > 0))
   const projectName = (project: Project) => project.name || getFilename(project.worktree)
   const projectOptions = createMemo(() => [
     { id: "all", label: language.t("settings.workspaces.filter.all") },
@@ -110,18 +110,18 @@ export const SettingsWorkspaces: Component<{ activeDirectory?: string }> = (prop
       workspaceDirectories().map((directory) => String(pathKey(directory))),
     ] as const,
     queryFn: () => loadSessions(workspaceDirectories()),
-    enabled: workspaceDirectories().length > 0,
+    enabled: serverSDK.connection.status() === "connected" && workspaceDirectories().length > 0,
     refetchOnMount: "always",
   }))
-  const sessionsByWorkspace = createMemo(
-    () =>
-      new Map(
-        workspaces().map((workspace) => [
-          pathKey(workspace.directory),
-          sessionQuery.data ? sessionsForWorkspace(sessionQuery.data, workspace.directory) : [],
-        ]),
-      ),
-  )
+  const sessionsByWorkspace = createMemo(() => {
+    const sessions = sessionQuery.isPending ? [] : (sessionQuery.data ?? [])
+    return new Map(
+      workspaces().map((workspace) => [
+        pathKey(workspace.directory),
+        sessionsForWorkspace(sessions, workspace.directory),
+      ]),
+    )
+  })
   const workspaceSessions = (workspace: Workspace) => sessionsByWorkspace().get(pathKey(workspace.directory)) ?? []
   const sessionCount = (workspace: Workspace) => {
     if (sessionQuery.isPending) return language.t("session.messages.loading")
@@ -279,7 +279,9 @@ export const SettingsWorkspaces: Component<{ activeDirectory?: string }> = (prop
       <div class="settings-tab-body settings-workspaces">
         <div class="settings-workspaces-toolbar">
           <span class="settings-workspaces-count">
-            {language.plural("settings.workspaces.count", filtered().length)}
+            <Show when={!projectQuery.isPending && !projectQuery.isError}>
+              {language.plural("settings.workspaces.count", filtered().length)}
+            </Show>
           </span>
           <div class="settings-workspaces-toolbar-actions">
             <Show when={projects().length > 1}>
@@ -332,7 +334,17 @@ export const SettingsWorkspaces: Component<{ activeDirectory?: string }> = (prop
         <div class="settings-workspaces-inventory">
           <Show
             when={filtered().length > 0}
-            fallback={<div class="settings-workspaces-empty">{language.t("settings.workspaces.empty")}</div>}
+            fallback={
+              <div class="settings-workspaces-empty">
+                {language.t(
+                  projectQuery.isPending
+                    ? "common.loading"
+                    : projectQuery.isError
+                      ? "common.requestFailed"
+                      : "settings.workspaces.empty",
+                )}
+              </div>
+            }
           >
             <SettingsList>
               <For each={filtered()}>

packages/app/src/shell/routes/routes.tsx

@@ -36,7 +36,7 @@ export function AppRoutes() {
           <SessionRouteFrame>
             <Suspense
               fallback={
-                <div class="flex min-h-0 flex-1 p-2">
+                <div class="flex min-h-0 flex-1 px-2 pb-2 pt-[var(--shell-top-inset,8px)]">
                   <SessionPanelFrame raised />
                 </div>
               }

packages/app/src/shell/shell.tsx

@@ -43,6 +43,13 @@ export default function Layout(props: ParentProps) {
         style={{
           "padding-top": "env(safe-area-inset-top, 0px)",
           "padding-bottom": "env(safe-area-inset-bottom, 0px)",
+          // The native Windows titlebar already includes the gap above the content panels.
+          "--shell-top-inset":
+            platform.platform === "desktop" &&
+            platform.os === "windows" &&
+            !(mobile() && preferences.general.mobileTitlebarPosition() === "bottom")
+              ? "0px"
+              : "8px",
         }}
       >
         <Titlebar
@@ -59,7 +66,7 @@ export default function Layout(props: ParentProps) {
             <aside
               ref={(element) => setState("tabsMount", element)}
               data-slot="vertical-tabs-sidebar"
-              class="relative flex min-h-0 shrink-0 flex-col bg-v2-background-bg-deep px-2.5 py-2"
+              class="relative flex min-h-0 shrink-0 flex-col bg-v2-background-bg-deep px-2.5 pb-2 pt-[var(--shell-top-inset,8px)]"
               style={{ width: `${state.tabsWidth}px` }}
             >
               <ResizeHandle

packages/app/src/shell/titlebar/titlebar.tsx

@@ -27,6 +27,7 @@ import { newTabTooltipKeybind } from "@/shell/commands/tooltip-keybind"
 import { TitlebarRightMount } from "@/shell/titlebar/right-slot"
 
 const titlebarHeight = 36
+const windowsTitlebarHeight = 44 // Includes the content inset; matches the native Windows overlay.
 const minTitlebarZoom = 0.25
 const windowsControlsBaseWidth = 138 // 3 native Windows caption buttons at 46px each.
 const macTrafficLightsBaseWidth = 84
@@ -59,7 +60,7 @@ export function Titlebar(props: {
   const titlebarZoom = () => (windows() ? Math.max(zoom(), minTitlebarZoom) : zoom())
   const minHeight = () => {
     if (mac()) return `${titlebarHeight / zoom()}px`
-    if (windows()) return `env(titlebar-area-height, ${titlebarHeight / Math.min(titlebarZoom(), 1)}px)`
+    if (windows()) return `env(titlebar-area-height, ${windowsTitlebarHeight / Math.min(titlebarZoom(), 1)}px)`
     return undefined
   }
   const windowsControlsWidth = () => `${windowsControlsBaseWidth / Math.max(titlebarZoom(), 1)}px`

packages/app/test-browser/playwright-config.test.ts

@@ -0,0 +1,70 @@
+import { expect, test } from "bun:test"
+import path from "node:path"
+
+test.each([
+  { name: "local development", ci: "", build: "", built: false },
+  { name: "local production build", ci: "", build: "1", built: true },
+  { name: "CI production build", ci: "true", build: "", built: true },
+  { name: "CI cannot opt into development", ci: "true", build: "0", built: true },
+])("Playwright uses $name", ({ ci, build, built }) => {
+  const result = Bun.spawnSync({
+    cmd: [process.execPath, "-e", 'import config from "./playwright.config.ts"; console.log(JSON.stringify(config))'],
+    cwd: path.resolve(import.meta.dirname, ".."),
+    env: {
+      ...process.env,
+      CI: ci,
+      PLAYWRIGHT_BUILD: build,
+      PLAYWRIGHT_BASE_URL: undefined,
+      PLAYWRIGHT_PORT: "4321",
+      PLAYWRIGHT_SERVER_HOST: "127.0.0.2",
+      PLAYWRIGHT_SERVER_PORT: "4322",
+    },
+  })
+  expect(result.exitCode).toBe(0)
+  const config = JSON.parse(result.stdout.toString())
+  expect(config.use.baseURL).toBe("http://127.0.0.1:4321")
+  expect(config.webServer.url).toBe(config.use.baseURL)
+  expect(config.webServer.command).toBe(
+    built
+      ? "bun run build && bun run serve -- --host 127.0.0.1 --port 4321 --strictPort"
+      : "bun run dev -- --host 127.0.0.1 --port 4321 --strictPort",
+  )
+  expect(config.webServer.reuseExistingServer).toBe(!built)
+  expect(config.webServer.env).toEqual({
+    VITE_OPENCODE_SERVER_HOST: built ? "127.0.0.1" : "127.0.0.2",
+    VITE_OPENCODE_SERVER_PORT: built ? "4321" : "4322",
+  })
+})
+
+test.each([
+  "./playwright.config.ts",
+  "./e2e/performance/playwright.config.ts",
+  "./e2e/performance/playwright.uncapped.config.ts",
+  "./e2e/performance/timeline-stability/playwright.config.ts",
+])("%s leaves an explicit external app unmanaged", (file) => {
+  const result = Bun.spawnSync({
+    cmd: [
+      process.execPath,
+      "-e",
+      `import config from ${JSON.stringify(file)}; console.log(JSON.stringify({ ...config, fixtureHost: process.env.PLAYWRIGHT_SERVER_HOST, fixturePort: process.env.PLAYWRIGHT_SERVER_PORT }))`,
+    ],
+    cwd: path.resolve(import.meta.dirname, ".."),
+    env: { ...process.env, CI: "true", PLAYWRIGHT_BASE_URL: "http://127.0.0.1:4444" },
+  })
+  expect(result.exitCode).toBe(0)
+  const config = JSON.parse(result.stdout.toString())
+  expect(config.webServer).toBeUndefined()
+  expect(config.use.baseURL).toBe("http://127.0.0.1:4444")
+  expect(config.fixtureHost).toBe("127.0.0.1")
+  expect(config.fixturePort).toBe("4444")
+})
+
+test("Playwright rejects HTTPS targets unsupported by the API fixtures", () => {
+  const result = Bun.spawnSync({
+    cmd: [process.execPath, "-e", 'import "./playwright.config.ts"'],
+    cwd: path.resolve(import.meta.dirname, ".."),
+    env: { ...process.env, CI: "true", PLAYWRIGHT_BASE_URL: "https://e2e.example.com" },
+  })
+  expect(result.exitCode).not.toBe(0)
+  expect(result.stderr.toString()).toContain("E2E fixtures require an http:// app URL")
+})

packages/app/test-browser/provider-catalog.test.ts

@@ -0,0 +1,52 @@
+import { expect, test } from "bun:test"
+import { createMemo, createRoot } from "solid-js"
+import { createStore } from "solid-js/store"
+import { normalizeProviderList } from "@/runtime/server/global-sync/utils"
+
+test("preserves an already normalized reactive catalog", () => {
+  const [store] = createStore({ catalog: { all: new Map<string, never>(), connected: [], default: {} } })
+  expect(normalizeProviderList(store.catalog)).toBe(store.catalog)
+})
+
+test("shares catalog snapshots and reacts to replacement lists", () => {
+  createRoot((dispose) => {
+    const provider = { id: "openai", name: "OpenAI", package: "@ai-sdk/openai", activation: "enabled" as const }
+    const model = {
+      id: "gpt-5",
+      modelID: "gpt-5",
+      providerID: "openai",
+      name: "GPT-5",
+      settings: {},
+      headers: {},
+      capabilities: { tools: true, input: ["text" as const], output: ["text" as const] },
+      variants: [],
+      time: { released: 1 },
+      cost: [],
+      status: "active" as const,
+      enabled: true,
+      limit: { context: 128_000, output: 8192 },
+    }
+    const [store, setStore] = createStore({ providers: [provider], models: [model] })
+    const first = createMemo(() => normalizeProviderList(store.providers, store.models))
+    const second = createMemo(() => normalizeProviderList(store.providers, store.models))
+    expect(first()).toBe(second())
+    const initial = first()
+    setStore("models", [{ ...model, name: "Renamed", limit: { context: 256_000, output: 16384 } }])
+    expect(first()).not.toBe(initial)
+    expect(first()).toBe(second())
+    expect(first().all.get("openai")?.models[model.id]).toMatchObject({
+      name: "Renamed",
+      limit: { context: 256_000, output: 16384 },
+    })
+    const renamed = first()
+    setStore("providers", [{ ...provider, name: "Custom OpenAI" }])
+    expect(first()).not.toBe(renamed)
+    expect(first()).toBe(second())
+    expect(first().all.get("openai")?.name).toBe("Custom OpenAI")
+    expect(normalizeProviderList(store.providers).all.get("openai")?.models).toEqual({})
+    setStore("models", [])
+    expect(first().all.get("openai")?.models).toEqual({})
+    expect(first().default).toEqual({})
+    dispose()
+  })
+})

packages/cli/package.json

@@ -27,7 +27,7 @@
     "@effect/platform-node": "catalog:",
     "@opencode-ai/client": "workspace:*",
     "@opencode-ai/plugin": "workspace:*",
-    "@opencode-ai/pty": "0.1.10",
+    "@opencode-ai/pty": "0.1.12",
     "@opencode-ai/schema": "workspace:*",
     "@opencode-ai/server": "workspace:*",
     "@opencode-ai/tui": "workspace:*",

packages/core/package.json

@@ -118,7 +118,7 @@
     "@ff-labs/fff-node": "0.10.5",
     "@opencode-ai/codemode": "workspace:*",
     "@opencode-ai/ai": "workspace:*",
-    "@opencode-ai/pty": "0.1.10",
+    "@opencode-ai/pty": "0.1.12",
     "@opencode-ai/schema": "workspace:*",
     "@opencode-ai/plugin": "workspace:*",
     "@opencode-ai/util": "workspace:*",

packages/core/src/bus.ts

@@ -11,6 +11,9 @@ import { KeyedMutex } from "./effect/keyed-mutex.js"
 import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
 import { isDeepStrictEqual } from "node:util"
 import { Durable } from "@opencode-ai/schema/durable-event-manifest"
+import { SessionEvent } from "@opencode-ai/schema/session-event"
+import type { SessionID } from "@opencode-ai/schema/session-id"
+import { AbsolutePath } from "@opencode-ai/schema/schema"
 
 export type Subscriber<D extends Event.Definition = Event.Definition> = (event: Event.Payload<D>) => Effect.Effect<void>
 export type Unsubscribe = Effect.Effect<void>
@@ -119,6 +122,9 @@ export interface Subscribe {
   /**
    * Volatile live channel: every event published from now on, nothing before or
    * across a disconnect. Consumers that need reliability combine it with `log`.
+   * With an ambient Location, delivery is restricted to that Location and global
+   * events. Unlocated Session events use the Session's owner at publication time.
+   * Session moves reach both the old and new Location, without changing the event.
    */
   (): Stream.Stream<Event.Payload>
   <D extends Event.Definition>(definition: D): Stream.Stream<Event.Payload<D>>
@@ -183,6 +189,7 @@ export function configured(options?: Options) {
         // Deferred import: a static one would close the module cycle
         // bus → location → project → bus and hit the node bindings in TDZ.
         const { Location } = yield* Effect.promise(() => import("./location.js"))
+        const { SessionTable } = yield* Effect.promise(() => import("./session/sql.js"))
         const pubsub = {
           live: yield* PubSub.unbounded<Event.Payload>(),
           durable: new Map<string, Set<PubSub.PubSub<void>>>(),
@@ -194,6 +201,64 @@ export function configured(options?: Options) {
         const { db } = yield* Database.Service
         const logReadPageSize = options?.logReadPageSize ?? 512
         const persist = options?.persist ?? false
+        const sessions = new Map<SessionID, Location.Ref>()
+        // Keep routing separate from the public event, and retain its snapshot
+        // while a slow subscriber drains events queued before a move or deletion.
+        const routes = new WeakMap<Event.Payload, readonly Location.Ref[]>()
+
+        const isSessionEvent = (event: Event.Payload): event is SessionEvent.Event =>
+          Object.hasOwn(SessionEvent.All.cases, event.type)
+
+        const prepareRoutes = Effect.fnUntraced(function* (events: readonly Event.Payload[]) {
+          const updates = new Map<SessionID, Location.Ref | undefined>()
+          const resolved = new Map<Event.Payload, readonly Location.Ref[]>()
+          for (const event of events) {
+            if (!isSessionEvent(event)) continue
+            const id = event.data.sessionID
+            if (event.type === "session.created") {
+              updates.set(id, event.data.location)
+              resolved.set(event, [event.location ?? event.data.location])
+              continue
+            }
+            if (event.location && event.type !== "session.forked" && event.type !== "session.moved") {
+              if (event.type === "session.deleted") updates.set(id, undefined)
+              continue
+            }
+            const owner = event.type === "session.forked" ? event.data.parentID : id
+            let ref = updates.has(owner) ? updates.get(owner) : sessions.get(owner)
+            if (!ref && !updates.has(owner)) {
+              const row = yield* db
+                .select({ directory: SessionTable.directory, workspaceID: SessionTable.workspace_id })
+                .from(SessionTable)
+                .where(eq(SessionTable.id, owner))
+                .get()
+                .pipe(Effect.orDie)
+              ref = row
+                ? { directory: AbsolutePath.make(row.directory), workspaceID: row.workspaceID ?? undefined }
+                : undefined
+              updates.set(owner, ref)
+            }
+            if (event.type === "session.moved") {
+              // Both owners need the transition, even if the producer supplied
+              // an envelope location. Later events use only the destination.
+              updates.set(id, event.data.location)
+              resolved.set(event, ref ? [ref, event.data.location] : [event.data.location])
+              continue
+            }
+            if (event.type === "session.forked") updates.set(id, ref)
+            resolved.set(event, event.location ? [event.location] : ref ? [ref] : [])
+            if (event.type === "session.deleted") updates.set(id, undefined)
+          }
+          // Apply only after the projection transaction commits. A failed move
+          // must not redirect events away from the Session's actual location.
+          return () => {
+            for (const [id, ref] of updates) {
+              if (ref) sessions.set(id, ref)
+              else sessions.delete(id)
+            }
+            for (const [event, ref] of resolved) routes.set(event, ref)
+          }
+        })
 
         const getOrCreate = (definition: Event.Definition) =>
           Effect.gen(function* () {
@@ -335,6 +400,7 @@ export function configured(options?: Options) {
                               ...event,
                               durable: { aggregateID, seq, version: durable.version },
                             } as Event.Payload
+                            const route = yield* prepareRoutes([committed])
                             for (const projector of list) {
                               yield* projector(committed)
                             }
@@ -366,12 +432,13 @@ export function configured(options?: Options) {
                                 ])
                                 .run()
                                 .pipe(Effect.orDie)
-                            return { aggregateID, seq }
+                            return { aggregateID, seq, event: committed, route }
                           }),
                         { behavior: "immediate" },
                       )
                       .pipe(Effect.orDie)
                     if (committed) {
+                      committed.route()
                       yield* Effect.forEach(
                         pubsub.durable.get(committed.aggregateID) ?? [],
                         (wake) => PubSub.publish(wake, undefined),
@@ -409,15 +476,14 @@ export function configured(options?: Options) {
                 Effect.gen(function* () {
                   const committed = yield* commitDurableEvent(definition, event as Event.Payload, undefined, commit)
                   if (!committed) return event
-                  event = {
-                    ...event,
-                    durable: envelope(committed.aggregateID, committed.seq, definition.durable.version),
-                  }
+                  event = committed.event as Event.Payload<D>
                   yield* notify(event as Event.Payload, true)
                   return event
                 }),
               )
             }
+            const route = yield* prepareRoutes([event as Event.Payload])
+            route()
             yield* notify(event as Event.Payload, false)
             return event
           })
@@ -527,7 +593,11 @@ export function configured(options?: Options) {
                             .pipe(Effect.orDie)
                           const firstSeq = (row?.seq ?? -1) + 1
                           const finalSeq = firstSeq + payloads.length - 1
-                          const result = new Array<Event.Payload>()
+                          const queued = payloads.map((item, index) => ({
+                            ...item.event,
+                            durable: envelope(aggregateID, firstSeq + index, item.definition.durable.version),
+                          }))
+                          const route = yield* prepareRoutes(queued)
                           const rows = new Array<typeof EventTable.$inferInsert>()
                           const ids = new Set<Event.ID>()
                           for (const [index, item] of payloads.entries()) {
@@ -559,10 +629,7 @@ export function configured(options?: Options) {
                                   }),
                                 )
                             }
-                            const event = {
-                              ...item.event,
-                              durable: envelope(aggregateID, seq, item.definition.durable.version),
-                            } as Event.Payload
+                            const event = queued[index]
                             for (const projector of projectors.get(
                               versionedType(item.definition.type, item.definition.durable.version),
                             ) ?? []) {
@@ -578,7 +645,6 @@ export function configured(options?: Options) {
                                 type: versionedType(item.definition.type, item.definition.durable.version),
                                 data: encoded,
                               })
-                            result.push(event)
                           }
                           yield* db
                             .insert(EventSequenceTable)
@@ -587,11 +653,12 @@ export function configured(options?: Options) {
                             .run()
                             .pipe(Effect.orDie)
                           if (persist) yield* db.insert(EventTable).values(rows).run().pipe(Effect.orDie)
-                          return result
+                          return { events: queued, route }
                         }),
                       { behavior: "immediate" },
                     )
                     .pipe(Effect.orDie)
+                  committed.route()
                   yield* Effect.forEach(
                     pubsub.durable.get(aggregateID) ?? [],
                     (wake) => PubSub.publish(wake, undefined),
@@ -599,8 +666,8 @@ export function configured(options?: Options) {
                       discard: true,
                     },
                   )
-                  yield* Effect.forEach(committed, (event) => notify(event, true), { discard: true })
-                  return committed as PublishResult<I>
+                  yield* Effect.forEach(committed.events, (event) => notify(event, true), { discard: true })
+                  return committed.events as PublishResult<I>
                 }),
               ),
             )
@@ -632,13 +699,7 @@ export function configured(options?: Options) {
                   strictOwner: options?.strictOwner,
                 })
                 if (committed && options?.publish) {
-                  yield* notify(
-                    {
-                      ...payload,
-                      durable: envelope(committed.aggregateID, committed.seq, definition.durable.version),
-                    },
-                    true,
-                  )
+                  yield* notify(committed.event, true)
                 }
               }),
             )
@@ -653,7 +714,10 @@ export function configured(options?: Options) {
                 yield* db.delete(EventTable).where(eq(EventTable.aggregate_id, aggregateID)).run()
               }),
             )
-            .pipe(Effect.orDie)
+            .pipe(
+              Effect.tap(() => Effect.sync(() => sessions.delete(aggregateID as SessionID))),
+              Effect.orDie,
+            )
         }
 
         function claim(aggregateID: string, ownerID: string) {
@@ -671,15 +735,17 @@ export function configured(options?: Options) {
               Effect.map((location) =>
                 Option.match(location, {
                   onNone: () => stream,
-                  onSome: (location) =>
-                    stream.pipe(
-                      Stream.filter(
-                        (event) =>
-                          !event.location ||
-                          (event.location.directory === location.directory &&
-                            event.location.workspaceID === location.workspaceID),
-                      ),
-                    ),
+                  onSome: (location) => {
+                    const matches = (ref: Location.Ref) =>
+                      ref.directory === location.directory && ref.workspaceID === location.workspaceID
+                    return stream.pipe(
+                      Stream.filter((event) => {
+                        const refs = routes.get(event)
+                        if (refs) return refs.some(matches)
+                        return !event.location || matches(event.location)
+                      }),
+                    )
+                  },
                 }),
               ),
             ),

packages/core/src/git.ts

@@ -507,9 +507,11 @@ const layer = Layer.effect(
       from: TreeID
       to: TreeID
     }) {
+      // Undo needs both paths of a rename, not only its destination.
       return (yield* repositoryOperation("list_files", input.repository, [
         "diff",
         "--name-only",
+        "--no-renames",
         "-z",
         input.from,
         input.to,

packages/core/src/image.ts

@@ -55,7 +55,7 @@ export interface Interface extends State.Transformable<Draft> {
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/Image") {}
 
-const layer = Layer.effect(
+export const layer = Layer.effect(
   Service,
   Effect.gen(function* () {
     const state = State.create<Limits, Draft>({

packages/core/src/job.ts

@@ -402,14 +402,7 @@ export const make = Effect.gen(function* () {
   })
 
   const pendingBackground: Interface["pendingBackground"] = Effect.gen(function* () {
-    const recovered: Background[] = []
-    let after: string | undefined
-    do {
-      const page = yield* kv.scan({ prefix: backgroundPrefix, after })
-      recovered.push(...Array.filterMap(page.entries, (entry) => decodeBackground(entry.value)))
-      after = page.next
-    } while (after)
-    return recovered
+    return Array.filterMap(yield* kv.scanAll(backgroundPrefix), (entry) => decodeBackground(entry.value))
   }).pipe(Effect.withSpan("Job.pendingBackground"))
 
   const completeBackground: Interface["completeBackground"] = Effect.fn("Job.completeBackground")((notificationID) =>

packages/core/src/kv.ts

@@ -29,6 +29,7 @@ export interface Interface {
   readonly set: (key: string, value: Value) => Effect.Effect<void>
   readonly remove: (key: string) => Effect.Effect<void>
   readonly scan: (options: ScanOptions) => Effect.Effect<ScanResult>
+  readonly scanAll: (prefix: string) => Effect.Effect<readonly Entry[]>
 }
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/KV") {}
@@ -37,6 +38,28 @@ const layer = Layer.effect(
   Service,
   Effect.gen(function* () {
     const db = (yield* Database.Service).db
+    const scan: Interface["scan"] = Effect.fn("KV.scan")(function* (options) {
+      const limit = Number.isNaN(options.limit) ? 100 : Math.min(Math.max(Math.floor(options.limit ?? 100), 1), 1000)
+      const end = prefixEnd(options.prefix)
+      const rows = yield* db
+        .select({ key: KVTable.key, value: KVTable.value })
+        .from(KVTable)
+        .where(
+          and(
+            options.prefix === "" ? undefined : gte(KVTable.key, options.prefix),
+            end === undefined ? undefined : lt(KVTable.key, end),
+            options.after === undefined ? undefined : gt(KVTable.key, options.after),
+          ),
+        )
+        .orderBy(asc(KVTable.key))
+        .limit(limit + 1)
+        .all()
+        .pipe(Effect.orDie)
+      const entries = rows.slice(0, limit)
+      if (rows.length <= limit) return { entries }
+      return { entries, next: entries[entries.length - 1].key }
+    })
+
     return Service.of({
       get: Effect.fn("KV.get")(function* (key) {
         return (yield* db
@@ -57,26 +80,16 @@ const layer = Layer.effect(
       remove: Effect.fn("KV.remove")(function* (key) {
         yield* db.delete(KVTable).where(eq(KVTable.key, key)).run().pipe(Effect.orDie)
       }),
-      scan: Effect.fn("KV.scan")(function* (options) {
-        const limit = Number.isNaN(options.limit) ? 100 : Math.min(Math.max(Math.floor(options.limit ?? 100), 1), 1000)
-        const end = prefixEnd(options.prefix)
-        const rows = yield* db
-          .select({ key: KVTable.key, value: KVTable.value })
-          .from(KVTable)
-          .where(
-            and(
-              options.prefix === "" ? undefined : gte(KVTable.key, options.prefix),
-              end === undefined ? undefined : lt(KVTable.key, end),
-              options.after === undefined ? undefined : gt(KVTable.key, options.after),
-            ),
-          )
-          .orderBy(asc(KVTable.key))
-          .limit(limit + 1)
-          .all()
-          .pipe(Effect.orDie)
-        const entries = rows.slice(0, limit)
-        if (rows.length <= limit) return { entries }
-        return { entries, next: entries[entries.length - 1].key }
+      scan,
+      scanAll: Effect.fn("KV.scanAll")(function* (prefix) {
+        const entries: Entry[] = []
+        let after: string | undefined
+        do {
+          const page = yield* scan({ prefix, after, limit: 1000 })
+          entries.push(...page.entries)
+          after = page.next
+        } while (after !== undefined)
+        return entries
       }),
     })
   }),

packages/core/src/mcp/client.ts

@@ -207,7 +207,10 @@ export const connect = Effect.fnUntraced(function* (
     }
     if (!URL.canParse(config.url))
       return yield* new ConnectError({ server, message: `Invalid MCP URL for "${server}"` })
-    return new StreamableHTTPClientTransport(new URL(config.url), {
+    // Prefer raw tools for our Code Mode without changing the configured URL used for OAuth identity.
+    const url = new URL(config.url)
+    if (config.codemode !== false && !url.searchParams.has("codemode")) url.searchParams.set("codemode", "false")
+    return new StreamableHTTPClientTransport(url, {
       requestInit: config.headers ? { headers: config.headers } : undefined,
       authProvider,
     })

packages/core/src/permission.ts

@@ -100,6 +100,10 @@ export function merge(...rulesets: Permission.Ruleset[]): Permission.Ruleset {
   return rulesets.flat()
 }
 
+export function relevant(input: Pick<Request, "action">, rules: Permission.Ruleset) {
+  return rules.filter((rule) => Wildcard.match(input.action, rule.action))
+}
+
 export interface Interface {
   readonly ask: (input: AssertInput) => Effect.Effect<AskResult, SessionErrors.NotFoundError>
   readonly assert: (input: AssertInput) => Effect.Effect<void, Error | SessionErrors.NotFoundError>
@@ -161,10 +165,6 @@ const layer = Layer.effect(
       return input.resources.some((resource) => evaluate(input.action, resource, rules).effect === "deny")
     }
 
-    function relevant(input: AssertInput, rules: Permission.Ruleset) {
-      return rules.filter((rule) => Wildcard.match(input.action, rule.action))
-    }
-
     const evaluateInput = Effect.fnUntraced(function* (input: AssertInput) {
       const rules = yield* configured(input.sessionID, input.agent)
       if (denied(input, rules)) return { effect: "deny" as const, rules }

packages/core/src/permissions.ts

@@ -0,0 +1,38 @@
+export * as Permissions from "./permissions.js"
+
+import { Effect } from "effect"
+import { Permission } from "./permission.js"
+import type { SessionSchema } from "./session/schema.js"
+import { Source } from "./source.js"
+
+export interface Interface {
+  readonly visibility: Source.Interface<Permission.Ruleset>
+  readonly ask: (
+    session: SessionSchema.Info,
+    request: Omit<Permission.AssertInput, "sessionID" | "agent">,
+  ) => Effect.Effect<void, Permission.Error | Permission.DeclinedError>
+}
+
+export const allowAll: Interface = {
+  visibility: Source.constant([{ action: "*", resource: "*", effect: "allow" }]),
+  ask: () => Effect.void,
+}
+
+export function rules(source: Source.Value<Permission.Ruleset>): Interface {
+  const visibility = Source.from(source)
+  return {
+    visibility,
+    ask: Effect.fn("Permissions.ask")(function* (session, request) {
+      const rules = yield* visibility.get(session)
+      if (
+        request.resources.every((resource) => Permission.evaluate(request.action, resource, rules).effect === "allow")
+      )
+        return
+      return yield* new Permission.BlockedError({
+        rules: Permission.relevant(request, rules),
+        permission: request.action,
+        resources: request.resources,
+      })
+    }),
+  }
+}

packages/core/src/plugin/hooks.ts

@@ -60,7 +60,7 @@ export interface Interface {
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/PluginHooks") {}
 
-const layer = Layer.effect(
+export const layer = Layer.effect(
   Service,
   Effect.gen(function* () {
     const callbacks = new Map<string, Entry[]>()

packages/core/src/plugin/host.ts

@@ -73,6 +73,7 @@ export const make = Effect.fn("PluginHost.make")(function* (plugin: Interface, p
 
   return {
     app,
+    location: locationInfo(),
     options: {},
     agent: {
       get: (input) => {
@@ -288,26 +289,6 @@ export const make = Effect.fn("PluginHost.make")(function* (plugin: Interface, p
         if (ref && !isCurrentLocation(ref)) return runtime.location.mcp.list(ref)
         return response(mcp.servers())
       },
-      add: (input) => {
-        const ref = locationRef(input)
-        if (ref && !isCurrentLocation(ref)) return runtime.location.mcp.add(ref, input.server, input.config)
-        return mcp.add(input.server, input.config)
-      },
-      remove: (input) => {
-        const ref = locationRef(input)
-        if (ref && !isCurrentLocation(ref)) return runtime.location.mcp.remove(ref, input.server)
-        return mcp.remove(input.server)
-      },
-      connect: (input) => {
-        const ref = locationRef(input)
-        if (ref && !isCurrentLocation(ref)) return runtime.location.mcp.connect(ref, input.server)
-        return mcp.connect(input.server)
-      },
-      disconnect: (input) => {
-        const ref = locationRef(input)
-        if (ref && !isCurrentLocation(ref)) return runtime.location.mcp.disconnect(ref, input.server)
-        return mcp.disconnect(input.server)
-      },
       reload: mcp.reload,
       transform: (callback) =>
         mcp.transform((draft) => {
@@ -377,14 +358,8 @@ export const make = Effect.fn("PluginHost.make")(function* (plugin: Interface, p
       hook: (name, callback) => hooks.register("shell", name, callback),
     },
     tool: {
-      transform: (callback) =>
-        tools
-          .transform((draft) =>
-            callback({
-              add: (tool) => draft.add(tool),
-            }),
-          )
-          .pipe(Effect.as({ dispose: Effect.void })),
+      transform: tools.transform,
+      reload: tools.reload,
       hook: (name, callback) => hooks.register("tool", name, callback),
     },
     vcs: {

packages/core/src/plugin/mcp-codemode-exclusion.ts

@@ -4,7 +4,7 @@ import { define } from "@opencode-ai/plugin/effect/plugin"
 import { Effect } from "effect"
 
 // These servers provide Code Mode, so expose them directly instead of nesting them inside OpenCode Code Mode.
-const urls = [/^https:\/\/mcp\.cloudflare\.com\/mcp$/, /^https:\/\/executor\.sh\/[^/]+\/mcp$/]
+const urls = [/^https:\/\/executor\.sh\/[^/]+\/mcp$/]
 
 export const Plugin = define({
   id: "opencode.mcp.codemode.exclusion",

packages/core/src/plugin/runtime.ts

@@ -2,7 +2,6 @@ export * as PluginRuntime from "./runtime.js"
 
 import { Context, Effect, Layer } from "effect"
 import { Agent } from "../agent.js"
-import { Mcp } from "@opencode-ai/schema/mcp"
 import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
 import { Job } from "../job.js"
 import { Location } from "../location.js"
@@ -40,10 +39,6 @@ export interface Interface {
       readonly list: (
         ref: Location.Ref,
       ) => Effect.Effect<{ readonly location: Location.Info; readonly data: MCP.ServerInfo[] }, unknown>
-      readonly add: (ref: Location.Ref, server: string, config: Mcp.ServerConfig) => Effect.Effect<void, unknown>
-      readonly remove: (ref: Location.Ref, server: string) => Effect.Effect<void, unknown>
-      readonly connect: (ref: Location.Ref, server: string) => Effect.Effect<void, unknown>
-      readonly disconnect: (ref: Location.Ref, server: string) => Effect.Effect<void, unknown>
     }
   }
 }
@@ -101,10 +96,6 @@ export const layerWithCell = (cell: Cell) =>
         },
         mcp: {
           list: (ref) => require(cell, (runtime) => runtime.location.mcp.list(ref)),
-          add: (ref, server, config) => require(cell, (runtime) => runtime.location.mcp.add(ref, server, config)),
-          remove: (ref, server) => require(cell, (runtime) => runtime.location.mcp.remove(ref, server)),
-          connect: (ref, server) => require(cell, (runtime) => runtime.location.mcp.connect(ref, server)),
-          disconnect: (ref, server) => require(cell, (runtime) => runtime.location.mcp.disconnect(ref, server)),
         },
       },
     }),
@@ -149,14 +140,6 @@ export const providerLayerWithCell = (cell: Cell) =>
                   data: yield* mcp.servers(),
                 }
               }).pipe(Effect.provide(locations.get(ref))),
-            add: (ref, server, config) =>
-              MCP.Service.use((mcp) => mcp.add(server, config)).pipe(Effect.provide(locations.get(ref))),
-            remove: (ref, server) =>
-              MCP.Service.use((mcp) => mcp.remove(server)).pipe(Effect.provide(locations.get(ref))),
-            connect: (ref, server) =>
-              MCP.Service.use((mcp) => mcp.connect(server)).pipe(Effect.provide(locations.get(ref))),
-            disconnect: (ref, server) =>
-              MCP.Service.use((mcp) => mcp.disconnect(server)).pipe(Effect.provide(locations.get(ref))),
           },
         },
       }

packages/core/src/project.ts

@@ -46,6 +46,15 @@ export interface Resolved {
   readonly vcsBackend?: string
 }
 
+export function markerless(directory: AbsolutePath): Resolved {
+  return {
+    id: ID.make(Hash.fast(`directory:${directory}`)),
+    directory,
+    canonical: directory,
+    vcs: undefined,
+  }
+}
+
 // Keep this filesystem-only; permission checks use it and should not execute VCS commands.
 export const root = Effect.fn("Project.root")(function* (
   fs: FSUtil.Interface,
@@ -361,12 +370,7 @@ const layer = Layer.effect(
         })
       }
 
-      return yield* persist({
-        id: ID.make(Hash.fast(`directory:${directory}`)),
-        directory,
-        canonical: directory,
-        vcs: undefined,
-      })
+      return yield* persist(markerless(directory))
     })
 
     return Service.of({ list, update, resolve })

packages/core/src/session.ts

@@ -1,5 +1,6 @@
 export * as Session from "./session.js"
 export * from "./session/schema.js"
+export type { OpenInput, Handle } from "./session/capabilities.js"
 
 import { Cause, Effect, Layer, Schema, Context, RcMap, Stream, Scope } from "effect"
 import { ListAnchor } from "@opencode-ai/schema/session"
@@ -55,6 +56,8 @@ import { fileURLToPath } from "url"
 import { SessionEnvironment } from "./session/environment.js"
 import { SessionHistory } from "./session/history.js"
 import { InstructionEntry } from "./session/instruction-entry.js"
+import { SessionResolve } from "./session/resolve.js"
+import type { SessionCapabilities } from "./session/capabilities.js"
 
 // get project -> project.locations
 //
@@ -187,6 +190,7 @@ export const MessageNotFoundError = SessionRevert.MessageNotFoundError
 export type MessageNotFoundError = SessionRevert.MessageNotFoundError
 
 export interface Interface {
+  readonly open: (input: SessionCapabilities.OpenInput) => Effect.Effect<SessionCapabilities.Handle>
   readonly list: (input?: ListInput) => Effect.Effect<{
     readonly data: SessionSchema.Info[]
   }>
@@ -341,6 +345,7 @@ const layer = Layer.effect(
     const execution = yield* SessionExecution.Service
     const store = yield* SessionStore.Service
     const locations = yield* LocationServiceMap.Service
+    const resolve = yield* SessionResolve.Service
     const fs = yield* FSUtil.Service
     const jobs = yield* Job.Service
     const environments = yield* SessionEnvironment.Service
@@ -348,6 +353,9 @@ const layer = Layer.effect(
     const activeShells = new Set<SessionSchema.ID>()
     const shellLocks = KeyedMutex.makeUnsafe<SessionSchema.ID>()
     const closeTransport = Effect.fn("Session.closeTransport")(function* (session: SessionSchema.Info) {
+      const resolved = yield* resolve.resolve(session)
+      if (resolved.status === "attached") return yield* resolved.capabilities.transport.close(session.id)
+      if (resolved.status === "owned-detached") return
       const location = Location.Ref.make({
         directory: session.location.directory,
         workspaceID: session.location.workspaceID,
@@ -356,7 +364,7 @@ const layer = Layer.effect(
       yield* SessionModelTransport.Service.use((transport) => transport.close(session.id)).pipe(
         Effect.provide(locations.get(location)),
       )
-    })
+    }, Effect.scoped)
     const isDurableSessionEvent = Schema.is(SessionEvent.Durable)
     const persistProject = (project: Project.Resolved) => upsertProject(db, project).pipe(Effect.orDie)
 
@@ -383,61 +391,90 @@ const layer = Layer.effect(
         }),
       )
 
+    const create = Effect.fn("Session.create")(function* (input: CreateInput, resolved?: Project.Resolved) {
+      const sessionID = input.id ?? SessionSchema.ID.create()
+      const recorded = yield* store.get(sessionID)
+      if (recorded) return recorded
+      const parent = input.parentID ? yield* store.get(input.parentID) : undefined
+      if (input.parentID && parent === undefined) return yield* new NotFoundError({ sessionID: input.parentID })
+      const location = parent?.location ?? input.location
+      if (location === undefined)
+        return yield* Effect.die(new Error("Session.create requires either location or an existing parentID"))
+      const project = resolved ?? (yield* projects.resolve(location.directory))
+      yield* persistProject(project)
+      const projected = yield* bus
+        .publish(
+          SessionEvent.Created,
+          {
+            sessionID,
+            slug: Slug.create(),
+            version: app.version,
+            projectID: project.id,
+            parentID: input.parentID,
+            location,
+            subpath: RelativePath.make(path.relative(project.directory, location.directory).replaceAll("\\", "/")),
+            title: input.title,
+            agent: input.agent,
+            model: input.model
+              ? {
+                  id: Model.ID.make(input.model.id),
+                  providerID: input.model.providerID,
+                  variant: input.model.variant,
+                }
+              : undefined,
+          },
+          { location },
+        )
+        .pipe(
+          Effect.as({ type: "created" } as const),
+          Effect.catchDefect((defect) => {
+            if (!(defect instanceof SessionProjector.SessionAlreadyProjected)) {
+              return Effect.die(defect)
+            }
+            // Concurrent creation lost the projection race. The existing Session identity wins.
+            return store
+              .get(sessionID)
+              .pipe(
+                Effect.flatMap((session) =>
+                  session ? Effect.succeed({ type: "existing", session } as const) : Effect.die(defect),
+                ),
+              )
+          }),
+        )
+      if (projected.type === "existing") return projected.session
+      // TODO: Restore recorded sessions onto replacement synchronized workspaces in a future API slice.
+      return yield* result.get(sessionID).pipe(Effect.orDie)
+    })
     const result = Service.of({
-      create: Effect.fn("Session.create")(function* (input) {
-        const sessionID = input.id ?? SessionSchema.ID.create()
-        const recorded = yield* store.get(sessionID)
-        if (recorded) return recorded
-        const parent = input.parentID ? yield* store.get(input.parentID) : undefined
-        if (input.parentID && parent === undefined) return yield* new NotFoundError({ sessionID: input.parentID })
-        const location = parent?.location ?? input.location
-        if (location === undefined)
-          return yield* Effect.die(new Error("Session.create requires either location or an existing parentID"))
-        const project = yield* projects.resolve(location.directory)
-        yield* persistProject(project)
-        const projected = yield* bus
-          .publish(
-            SessionEvent.Created,
-            {
-              sessionID,
-              slug: Slug.create(),
-              version: app.version,
-              projectID: project.id,
-              parentID: input.parentID,
-              location,
-              subpath: RelativePath.make(path.relative(project.directory, location.directory).replaceAll("\\", "/")),
-              title: input.title,
-              agent: input.agent,
-              model: input.model
-                ? {
-                    id: Model.ID.make(input.model.id),
-                    providerID: input.model.providerID,
-                    variant: input.model.variant,
-                  }
-                : undefined,
-            },
-            { location },
-          )
-          .pipe(
-            Effect.as({ type: "created" } as const),
-            Effect.catchDefect((defect) => {
-              if (!(defect instanceof SessionProjector.SessionAlreadyProjected)) {
-                return Effect.die(defect)
-              }
-              // Concurrent creation lost the projection race. The existing Session identity wins.
-              return store
-                .get(sessionID)
-                .pipe(
-                  Effect.flatMap((session) =>
-                    session ? Effect.succeed({ type: "existing", session } as const) : Effect.die(defect),
-                  ),
-                )
-            }),
-          )
-        if (projected.type === "existing") return projected.session
-        // TODO: Restore recorded sessions onto replacement synchronized workspaces in a future API slice.
-        return yield* result.get(sessionID).pipe(Effect.orDie)
-      }),
+      create,
+      open: Effect.fn("Session.open")((input) =>
+        Effect.uninterruptible(
+          Effect.gen(function* () {
+            const directory = AbsolutePath.make(process.cwd())
+            // Transitional placement is unused by supplied capabilities, but listings group
+            // these Sessions under the deterministic cwd-derived project. Adoption never moves it.
+            const session = yield* resolve.own(
+              create(
+                { id: input.id, title: input.title, location: Location.Ref.make({ directory }) },
+                Project.markerless(directory),
+              ).pipe(Effect.orDie),
+            )
+            const close = yield* resolve.attach(session.id, input)
+            return {
+              id: session.id,
+              prompt: (prompt) =>
+                Effect.gen(function* () {
+                  const admitted = yield* result.prompt({ ...prompt, sessionID: session.id, resume: false })
+                  if (prompt.resume !== false) yield* result.resume(session.id)
+                  return admitted
+                }),
+              resume: () => result.resume(session.id),
+              interrupt: (options) => result.interrupt(session.id, options),
+              close,
+            } satisfies SessionCapabilities.Handle
+          }),
+        ),
+      ),
       fork: Effect.fn("Session.fork")(function* (input) {
         const parent = yield* result.get(input.sessionID)
         const boundary = yield* db
@@ -510,6 +547,7 @@ const layer = Layer.effect(
         yield* Effect.forEach(children.data, (child) => result.remove(child.id), { concurrency: 1, discard: true })
         yield* environments.clear(sessionID)
         yield* bus.publish(SessionEvent.Deleted, { sessionID })
+        yield* resolve.remove(sessionID)
         yield* bus.remove(sessionID)
       }),
       list: Effect.fn("Session.list")(function* (input = {}) {
@@ -648,16 +686,24 @@ const layer = Layer.effect(
             if (session.revert) yield* SessionRevert.commit(session).pipe(Effect.provideService(Bus.Service, bus))
             // Resolved lazily so prompt admission only boots location services when an
             // image attachment actually needs the resizer.
-            const image = Effect.gen(function* () {
-              const plugins = yield* PluginSupervisor.Service
-              yield* plugins.flush
-              return yield* Image.Service
-            }).pipe(Effect.provide(locations.get(session.location)))
-            const skills = Effect.gen(function* () {
-              const plugins = yield* PluginSupervisor.Service
-              yield* plugins.flush
-              return yield* Skill.Service
-            }).pipe(Effect.provide(locations.get(session.location)))
+            const resolved = input.files?.length || input.skills?.length ? yield* resolve.resolve(session) : undefined
+            // TODO: typed unavailable-operation errors belong to the capability-gated operations phase.
+            if (resolved?.status === "owned-detached") return yield* SessionResolve.unavailable(session.id)
+            const capabilities = resolved?.status === "attached" ? resolved.capabilities : undefined
+            const image = capabilities
+              ? Effect.succeed(capabilities.image)
+              : Effect.gen(function* () {
+                  const plugins = yield* PluginSupervisor.Service
+                  yield* plugins.flush
+                  return yield* Image.Service
+                }).pipe(Effect.provide(locations.get(session.location)))
+            const skills = capabilities
+              ? Effect.undefined
+              : Effect.gen(function* () {
+                  const plugins = yield* PluginSupervisor.Service
+                  yield* plugins.flush
+                  return yield* Skill.Service
+                }).pipe(Effect.provide(locations.get(session.location)))
             const prompt = yield* resolvePrompt(
               { text: input.text, files: input.files, agents: input.agents, skills: input.skills },
               image,
@@ -690,7 +736,7 @@ const layer = Layer.effect(
             }
             return admitted
           }),
-        ),
+        ).pipe(Effect.scoped),
       ),
       generate: Effect.fn("Session.generate")(function* (input) {
         const session = yield* result.get(input.sessionID)
@@ -778,6 +824,7 @@ const layer = Layer.effect(
       }),
       skill: Effect.fn("Session.skill")(function* (input) {
         const session = yield* result.get(input.sessionID)
+        if (resolve.status(session.id) !== "unowned") return yield* new SkillNotFoundError({ skill: input.skill })
         const skills = yield* Skill.Service.pipe(Effect.provide(locations.get(session.location)))
         const skill = yield* skills.get(input.skill)
         if (!skill) return yield* new SkillNotFoundError({ skill: input.skill })
@@ -1021,7 +1068,7 @@ function synthesizeTerminalShellInfo(started: ShellSchema.Info): ShellSchema.Inf
 const resolvePrompt = Effect.fn("Session.resolvePrompt")(function* (
   input: PromptInput.Prompt,
   image: Effect.Effect<Image.Interface>,
-  skills: Effect.Effect<Skill.Interface>,
+  skills: Effect.Effect<Skill.Interface | undefined>,
 ) {
   const fs = yield* FSUtil.Service
   const files = input.files
@@ -1031,6 +1078,7 @@ const resolvePrompt = Effect.fn("Session.resolvePrompt")(function* (
   const selected = yield* Effect.gen(function* () {
     if (!requested?.length) return undefined
     const skillService = yield* skills
+    if (!skillService) return yield* new SkillNotFoundError({ skill: requested[0].id })
     const prepared = new Map<Skill.ID, Skill.Name>()
     return yield* Effect.forEach(requested, (attachment) =>
       Effect.gen(function* () {
@@ -1201,6 +1249,7 @@ export const node = makeGlobalNode({
     Project.node,
     SessionExecution.node,
     SessionStore.node,
+    SessionResolve.node,
     LocationServiceMap.node,
     SessionProjector.node,
     FSUtil.node,

packages/core/src/session/capabilities.ts

@@ -0,0 +1,38 @@
+export * as SessionCapabilities from "./capabilities.js"
+
+import type { Effect } from "effect"
+import type { Instructions } from "../instructions/index.js"
+import type { Permissions } from "../permissions.js"
+import type { Source } from "../source.js"
+import type { Tool } from "../tool.js"
+import type { Session } from "../session.js"
+import type { SessionRunner } from "./runner/index.js"
+import type { SessionRunnerModel } from "./runner/model.js"
+import type { SessionSchema } from "./schema.js"
+
+export interface OpenInput {
+  readonly id?: SessionSchema.ID
+  readonly title?: string
+  readonly model: Source.Value<SessionRunnerModel.Resolved, SessionRunnerModel.Error>
+  readonly tools?: Source.Value<ReadonlyArray<Tool.Info>>
+  readonly instructions?: Source.Value<ReadonlyArray<string> | Instructions.Unavailable>
+  readonly permissions?: Permissions.Interface
+  readonly system?: Source.Value<string | Instructions.Unavailable>
+  readonly limits?: Source.Value<{ readonly steps?: number }>
+  /** Called after replacement or host teardown, once all in-flight work has settled. */
+  readonly retire?: () => Effect.Effect<void>
+}
+
+export interface Handle {
+  readonly id: SessionSchema.ID
+  readonly prompt: (
+    input: Omit<Parameters<Session.Interface["prompt"]>[0], "sessionID">,
+  ) => Effect.Effect<
+    Effect.Success<ReturnType<Session.Interface["prompt"]>>,
+    Effect.Error<ReturnType<Session.Interface["prompt"]>> | SessionRunner.RunError
+  >
+  readonly resume: () => ReturnType<Session.Interface["resume"]>
+  readonly interrupt: (options?: { readonly continue?: boolean }) => Effect.Effect<boolean>
+  /** Releases this open's capabilities after settlement without interrupting or deleting the Session. */
+  readonly close: () => Effect.Effect<void>
+}

packages/core/src/session/compaction.ts

@@ -8,9 +8,10 @@ import { Bus } from "../bus.js"
 import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
 import { llmClient } from "../effect/app-node-platform.js"
 import { SessionEvent } from "./event.js"
+import type { SessionContext } from "./context.js"
 import type { SessionMessage } from "./message.js"
-import { SessionModelRequest } from "./model-request.js"
-import { SessionRunnerModel } from "./runner/model.js"
+import type { SessionModelRequest } from "./model-request.js"
+import type { SessionRunnerModel } from "./runner/model.js"
 import { SessionSchema } from "./schema.js"
 import { toSessionError } from "./to-session-error.js"
 import { Token } from "../util/token.js"
@@ -69,14 +70,13 @@ type Dependencies = {
   readonly llm: {
     readonly stream: (request: LLMRequest, options?: StreamOptions) => Stream.Stream<LLMEvent, AIError>
   }
-  readonly models: SessionRunnerModel.Interface
-  readonly modelRequests: SessionModelRequest.Interface
 }
 
 export type AutoInput = {
   readonly session: SessionSchema.Info
   readonly messages: readonly SessionMessage.Info[]
   readonly resolved: SessionRunnerModel.Resolved
+  readonly prepare: SessionModelRequest.Interface["prepare"]
 }
 
 type RequiredInput = Pick<AutoInput, "messages" | "resolved">
@@ -86,6 +86,9 @@ export type ManualInput = {
   readonly messages: readonly SessionMessage.Info[]
   readonly inputID: SessionMessage.ID
   readonly started?: boolean
+  /** Invoked after content planning, not when the caller captures the operation. */
+  readonly resolveModel: SessionContext.Interface["resolveModel"]
+  readonly prepare: SessionModelRequest.Interface["prepare"]
 }
 
 type Plan = {
@@ -96,6 +99,7 @@ type Plan = {
   readonly recent: string
   readonly inputID?: SessionMessage.ID
   readonly started?: boolean
+  readonly prepare: SessionModelRequest.Interface["prepare"]
 }
 
 export type Outcome =
@@ -278,7 +282,7 @@ const make = (dependencies: Dependencies) => {
           })
         : Effect.void,
     )
-    const prepared = yield* dependencies.modelRequests.prepare({
+    const prepared = yield* plan.prepare({
       scope: { session: plan.session, agentID: Agent.ID.make("compaction"), model: plan.resolved },
       transcript: { system: [], messages: [Message.user(plan.prompt)] },
       contextHooks: false,
@@ -348,6 +352,7 @@ const make = (dependencies: Dependencies) => {
       return yield* execute({
         session: input.session,
         resolved: input.resolved,
+        prepare: input.prepare,
         reason: "auto",
         ...content,
       })
@@ -387,7 +392,7 @@ const make = (dependencies: Dependencies) => {
         error: { type: "compaction.unavailable", message: "Nothing to compact yet" },
         inputID: input.inputID,
       })
-    const resolved = yield* dependencies.models.resolve(input.session).pipe(
+    const resolved = yield* input.resolveModel(input.session).pipe(
       Effect.catch((cause) =>
         failed({
           sessionID: input.session.id,
@@ -401,6 +406,7 @@ const make = (dependencies: Dependencies) => {
     return yield* execute({
       session: input.session,
       resolved,
+      prepare: input.prepare,
       reason: "manual",
       inputID: input.inputID,
       started: input.started,
@@ -422,14 +428,12 @@ export const layer = Layer.effect(
   Effect.gen(function* () {
     const bus = yield* Bus.Service
     const llm = yield* LLMClient.Service
-    const models = yield* SessionRunnerModel.Service
-    const modelRequests = yield* SessionModelRequest.Service
-    return make({ bus, llm, models, modelRequests })
+    return make({ bus, llm })
   }),
 )
 
 export const node = makeLocationNode({
   service: Service,
   layer,
-  deps: [Bus.node, llmClient, SessionRunnerModel.node, SessionModelRequest.node],
+  deps: [Bus.node, llmClient],
 })

packages/core/src/session/context.ts

@@ -1,7 +1,8 @@
 export * as SessionContext from "./context.js"
 
-import { Context, Effect, Layer } from "effect"
+import { Context, Effect, Layer, Schema } from "effect"
 import { Agent } from "../agent.js"
+import { Catalog } from "../catalog.js"
 import { CodeModeInstructions } from "../codemode/instructions.js"
 import { Database } from "../database/database.js"
 import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
@@ -11,14 +12,23 @@ import { InstructionBuiltIns } from "../instructions/builtins.js"
 import { Location } from "../location.js"
 import { McpInstructions } from "../mcp/instructions.js"
 import { McpTool } from "../tool/mcp.js"
+import { Model } from "../model.js"
 import { PluginSupervisor } from "../plugin/supervisor.js"
 import { ReferenceInstructions } from "../reference/instructions.js"
 import { SkillInstructions } from "../skill/instructions.js"
 import { Tool } from "../tool.js"
+import { Permission } from "../permission.js"
+import { Permissions } from "../permissions.js"
+import { Image } from "../image.js"
+import { PluginHooks } from "../plugin/hooks.js"
+import { Source } from "../source.js"
+import type { SessionCapabilities } from "./capabilities.js"
+import { SessionSystemPrompt } from "./system-prompt.js"
 import { AgentNotFoundError } from "./error.js"
 import { SessionHistory } from "./history.js"
 import { InstructionEntry } from "./instruction-entry.js"
 import { SessionMessage } from "./message.js"
+import { SessionModelRequest } from "./model-request.js"
 import { SessionRunnerModel } from "./runner/model.js"
 import { SessionSchema } from "./schema.js"
 import { SessionStore } from "./store.js"
@@ -28,6 +38,8 @@ export interface Selection {
   readonly agent: Agent.Selection & { readonly info: Agent.Info }
   readonly instructions: Instructions.List
   readonly tools: Tool.Snapshot
+  /** `baseTranscript` uses its default prefix for undefined; "" omits it when system text lives in the instruction epoch. */
+  readonly system?: string
 }
 
 export interface Loaded {
@@ -37,19 +49,33 @@ export interface Loaded {
   readonly initial: string
   readonly messages: ReadonlyArray<SessionMessage.Info>
   readonly tools: Tool.Snapshot
+  readonly system?: Selection["system"]
 }
 
 /**
  * Resolves model-request state in two phases: `select` fixes the Session,
  * agent, instruction sources, and tool snapshot; `load` adds the model and
- * active history for that selection. This module does not build or execute the
- * model request.
+ * active history for that selection. Auxiliary operations resolve only the
+ * capabilities they need; request preparation stays separate from selection.
  */
 export interface Interface {
   /** Selects the Session, agent, instructions, and tools used by subsequent work. */
   readonly select: (sessionID: SessionSchema.ID) => Effect.Effect<Selection, AgentNotFoundError>
   /** Resolves the model and active history for that selection. */
   readonly load: (selection: Selection) => Effect.Effect<Loaded, SessionRunnerModel.Error>
+  readonly resolveModel: (
+    session: SessionSchema.Info,
+  ) => Effect.Effect<SessionRunnerModel.Resolved, SessionRunnerModel.Error>
+  /** Selects auxiliary title capabilities without instruction or tool preflight. */
+  readonly selectTitle: (session: SessionSchema.Info) => Effect.Effect<
+    | {
+        readonly agent: Agent.Info
+        readonly primary: SessionRunnerModel.Resolved | undefined
+        readonly selected: SessionRunnerModel.Resolved
+      }
+    | undefined
+  >
+  readonly prepare: SessionModelRequest.Interface["prepare"]
 }
 
 /** Location-scoped model-context loader for durable Session Steps. */
@@ -60,6 +86,7 @@ const layer = Layer.effect(
   Effect.gen(function* () {
     const agents = yield* Agent.Service
     const builtins = yield* InstructionBuiltIns.Service
+    const catalog = yield* Catalog.Service
     const db = (yield* Database.Service).db
     const discovery = yield* InstructionDiscovery.Service
     const entries = yield* InstructionEntry.Service
@@ -67,12 +94,41 @@ const layer = Layer.effect(
     const mcpInstructions = yield* McpInstructions.Service
     const mcpTools = yield* McpTool.Service
     const models = yield* SessionRunnerModel.Service
+    const modelRequests = yield* SessionModelRequest.Service
     const plugins = yield* PluginSupervisor.Service
     const referenceInstructions = yield* ReferenceInstructions.Service
     const skillInstructions = yield* SkillInstructions.Service
     const store = yield* SessionStore.Service
     const registry = yield* Tool.Service
 
+    const resolveModel = (session: SessionSchema.Info) => models.resolve(session, catalog.model.available)
+
+    const selectTitle = Effect.fn("SessionContext.selectTitle")(function* (session: SessionSchema.Info) {
+      const agent = yield* agents.get(Agent.ID.make("title"))
+      if (!agent) return
+      const primary = yield* resolveModel(session).pipe(Effect.orElseSucceed(() => undefined))
+      const info = yield* Effect.gen(function* () {
+        if (agent.model) return yield* catalog.model.get(agent.model.providerID, agent.model.id)
+        if (!primary) return
+        return yield* catalog.model.small(primary.ref.providerID)
+      })
+      const variant =
+        agent.model?.variant ?? MINIMAL_REASONING_VARIANTS.find((id) => info?.variants.some((item) => item.id === id))
+      const preferred =
+        info &&
+        (yield* resolveModel({
+          ...session,
+          model: Model.Ref.make({
+            providerID: info.providerID,
+            id: info.id,
+            ...(variant ? { variant } : {}),
+          }),
+        }).pipe(Effect.orElseSucceed(() => undefined)))
+      const selected = preferred ?? primary
+      if (!selected) return
+      return { agent, primary, selected }
+    })
+
     const select = Effect.fn("SessionContext.select")(function* (sessionID: SessionSchema.ID) {
       const session = yield* store.get(sessionID)
       if (!session) return yield* Effect.die(new Error(`Session not found: ${sessionID}`))
@@ -111,28 +167,153 @@ const layer = Layer.effect(
       }
     })
 
-    const load = Effect.fn("SessionContext.load")(function* (selection: Selection) {
-      const model = yield* models.resolve(selection.session)
-      const history = yield* SessionHistory.entriesForRunner(db, selection.session.id, selection.instructions)
-      return {
-        session: selection.session,
-        agent: selection.agent,
-        model,
-        initial: history.initial,
-        messages: history.entries.map((entry) => entry.message),
-        tools: selection.tools,
-      }
+    return Service.of({
+      select,
+      load: load(db, resolveModel),
+      resolveModel,
+      selectTitle,
+      prepare: modelRequests.prepare,
     })
-
-    return Service.of({ select, load })
   }),
 )
 
+/** The values path shares instruction persistence and history assembly with discovery. */
+export const values = (input: SessionCapabilities.OpenInput) =>
+  Layer.effect(
+    Service,
+    Effect.gen(function* () {
+      const db = (yield* Database.Service).db
+      const store = yield* SessionStore.Service
+      const entries = yield* InstructionEntry.Service
+      const requests = yield* SessionModelRequest.Service
+      const hooks = yield* PluginHooks.Service
+      const image = yield* Image.Service
+      const tools = Source.from(input.tools ?? [])
+      const instructions = Source.from(input.instructions ?? [])
+      const limits = Source.from(input.limits ?? {})
+      const permissions = input.permissions ?? Permissions.allowAll
+      const model = Source.from(input.model)
+      const resolveModel: Interface["resolveModel"] = (session) => model.get(session)
+      let cached:
+        | {
+            readonly tools: ReadonlyArray<Tool.Info>
+            readonly rules: Permission.Ruleset
+            readonly snapshot: Tool.Snapshot
+          }
+        | undefined
+      const select: Interface["select"] = Effect.fn("SessionContext.selectValues")(function* (sessionID) {
+        const session = yield* store.get(sessionID)
+        if (!session) return yield* Effect.die(new Error(`Session not found: ${sessionID}`))
+        const selected = yield* Effect.all(
+          {
+            tools: tools.get(session),
+            rules: permissions.visibility.get(session),
+            limits: limits.get(session),
+            entries: entries.load(sessionID),
+          },
+          { concurrency: "unbounded" },
+        )
+        if (cached?.tools !== selected.tools || cached.rules !== selected.rules)
+          cached = {
+            tools: selected.tools,
+            rules: selected.rules,
+            snapshot: yield* Tool.snapshot(selected.tools, selected.rules).pipe(
+              Effect.provideService(PluginHooks.Service, hooks),
+              Effect.provideService(Image.Service, image),
+            ),
+          }
+        const snapshot = cached.snapshot
+        const id = session.agent ?? Agent.defaultID
+        return {
+          session,
+          agent: { id, info: { ...Agent.Info.default(id), permissions: selected.rules, steps: selected.limits.steps } },
+          // System text participates in the epoch instead of changing the privileged prefix.
+          system: "",
+          tools: snapshot,
+          instructions: Instructions.combine([
+            Instructions.make({
+              key: Instructions.Key.make("session/system"),
+              codec: Schema.String,
+              read:
+                input.system === undefined
+                  ? Effect.succeed(SessionSystemPrompt.make(snapshot.definitions.map((tool) => tool.name)))
+                  : Source.from(input.system)
+                      .get(session)
+                      .pipe(Effect.map((value) => (value === "" ? Instructions.removed : value))),
+              render: {
+                initial: (value) => value,
+                changed: (_previous, value) =>
+                  `The system instructions changed and supersede the previous value:\n${value}`,
+                removed: () => "The previous system instructions no longer apply.",
+              },
+            }),
+            CodeModeInstructions.make(snapshot.codeModeCatalog),
+            Instructions.make({
+              key: Instructions.Key.make("session/instructions"),
+              codec: Schema.Array(Schema.String),
+              read: instructions
+                .get(session)
+                .pipe(
+                  Effect.map((value) =>
+                    Array.isArray(value) && !value.some((part) => part.length > 0) ? Instructions.removed : value,
+                  ),
+                ),
+              render: {
+                initial: (value) => value.join("\n\n"),
+                changed: (_previous, value) =>
+                  `The session instructions changed and supersede the previous value:\n${value.join("\n\n")}`,
+                removed: () => "The previous session instructions no longer apply.",
+              },
+            }),
+            Instructions.make({
+              key: Instructions.Key.make("session/permissions"),
+              codec: Schema.toCodecJson(Permission.Ruleset),
+              read: Effect.succeed(selected.rules.length > 0 ? selected.rules : Instructions.removed),
+              render: {
+                initial: (value) => `Permission rules:\n${JSON.stringify(value)}`,
+                changed: (_previous, value) => `Permission rules changed:\n${JSON.stringify(value)}`,
+                removed: () => "The previous permission rules no longer apply.",
+              },
+            }),
+            selected.entries,
+          ]),
+        }
+      })
+      return Service.of({
+        select,
+        load: load(db, resolveModel),
+        resolveModel,
+        selectTitle: () => Effect.undefined,
+        prepare: requests.prepare,
+      })
+    }),
+  )
+
+function load(db: Database.Interface["db"], resolveModel: Interface["resolveModel"]): Interface["load"] {
+  return Effect.fn("SessionContext.load")(function* (selection: Selection) {
+    const model = yield* resolveModel(selection.session)
+    const history = yield* SessionHistory.entriesForRunner(db, selection.session.id, selection.instructions)
+    return {
+      session: selection.session,
+      agent: selection.agent,
+      model,
+      initial: history.initial,
+      messages: history.entries.map((entry) => entry.message),
+      tools: selection.tools,
+      ...(selection.system === undefined ? {} : { system: selection.system }),
+    }
+  })
+}
+
+/** Variant IDs that minimize reasoning output, in preference order. */
+const MINIMAL_REASONING_VARIANTS = ["none", "minimal", "low"].map((id) => Model.VariantID.make(id))
+
 export const node = makeLocationNode({
   service: Service,
   layer,
   deps: [
     Agent.node,
+    Catalog.node,
     Database.node,
     InstructionBuiltIns.node,
     InstructionDiscovery.node,
@@ -143,6 +324,7 @@ export const node = makeLocationNode({
     PluginSupervisor.node,
     ReferenceInstructions.node,
     SessionRunnerModel.node,
+    SessionModelRequest.node,
     SessionStore.node,
     SkillInstructions.node,
     Tool.node,

packages/core/src/session/execution.ts

@@ -14,6 +14,7 @@ import { SessionStore } from "./store.js"
 import { toSessionError } from "./to-session-error.js"
 import { UserInterruptedError } from "./error.js"
 import { SessionInbox } from "./inbox.js"
+import { SessionResolve } from "./resolve.js"
 
 export interface Interface {
   /** Snapshots active execution owned by this process. */
@@ -52,6 +53,7 @@ export const layer = Layer.effect(
   Effect.gen(function* () {
     const store = yield* SessionStore.Service
     const locations = yield* LocationServiceMap.Service
+    const resolve = yield* SessionResolve.Service
     const bus = yield* Bus.Service
     const jobs = yield* Job.Service
     const db = (yield* Database.Service).db
@@ -86,10 +88,14 @@ export const layer = Layer.effect(
       return Effect.gen(function* () {
         const session = yield* store.get(sessionID)
         if (!session) return yield* Effect.die(new Error(`Session not found: ${sessionID}`))
-        const result = yield* SessionRunner.Service.use((runner) =>
-          runner.drain({ sessionID, force, continuation, promotable }),
+        const pinned = resolve.pinned(sessionID)
+        const result = yield* (
+          pinned
+            ? pinned.drain({ sessionID, force, continuation, promotable })
+            : SessionRunner.Service.use((runner) => runner.drain({ sessionID, force, continuation, promotable })).pipe(
+                Effect.provide(locations.get(session.location)),
+              )
         ).pipe(
-          Effect.provide(locations.get(session.location)),
           Effect.tapCause((cause) =>
             Cause.hasInterruptsOnly(cause)
               ? Effect.void
@@ -101,11 +107,14 @@ export const layer = Layer.effect(
       })
     }
     const coordinator = yield* SessionRunCoordinator.make<SessionSchema.ID, SessionRunner.RunError, InterruptReason>({
-      started: (sessionID) =>
-        reportLifecycle(
+      eligible: (sessionID) => resolve.status(sessionID) !== "owned-detached",
+      started: (sessionID) => {
+        resolve.pin(sessionID)
+        return reportLifecycle(
           sessionID,
           bus.publish(SessionEvent.Execution.Started, { sessionID }, claimOnCommit(sessionID)),
-        ),
+        )
+      },
       drain: (sessionID, force, promotable) => drain(sessionID, force, undefined, promotable),
       // One terminal observation per busy period, covering every coalesced drain.
       settled: (sessionID, exit, reason) =>
@@ -137,7 +146,7 @@ export const layer = Layer.effect(
               releaseOnCommit(sessionID),
             )
           }),
-        ),
+        ).pipe(Effect.ensuring(resolve.settle(sessionID))),
     })
 
     return Service.of({
@@ -160,7 +169,12 @@ export const layer = Layer.effect(
             yield* coordinator.wake(sessionID, "steer")
           return interrupted
         }),
-      resume: coordinator.run,
+      resume: (sessionID) =>
+        Effect.gen(function* () {
+          // TODO: typed unavailable-operation errors belong to the capability-gated operations phase.
+          if (resolve.status(sessionID) === "owned-detached") return yield* SessionResolve.unavailable(sessionID)
+          yield* coordinator.run(sessionID)
+        }),
       wake: coordinator.wake,
       awaitIdle: coordinator.awaitIdle,
     })
@@ -170,7 +184,7 @@ export const layer = Layer.effect(
 export const node = makeGlobalNode({
   service: Service,
   layer,
-  deps: [SessionStore.node, LocationServiceMap.node, Bus.node, Database.node, Job.node],
+  deps: [SessionStore.node, LocationServiceMap.node, Bus.node, Database.node, Job.node, SessionResolve.node],
 })
 
 /** Low-level compatibility layer for callers that only need durable Session recording. */

packages/core/src/session/execution/restart.ts

@@ -7,6 +7,7 @@ import { Job } from "../../job.js"
 import { Session } from "../../session.js"
 import { SessionEvent } from "../event.js"
 import { SessionExecution } from "../execution.js"
+import { SessionResolve } from "../resolve.js"
 import { SessionSchema } from "../schema.js"
 import { SessionStore } from "../store.js"
 
@@ -37,6 +38,8 @@ export interface Interface {
    * shutdown (which preserves the claim on purpose). The claim is never
    * cleared here: only a terminal event releases it, so a death anywhere in
    * the resume path leaves the same orphaned claim for the next boot.
+   * Capability-owned Sessions stay pending even when reopened; they require
+   * an explicit prompt or resume instead of automatic recovery.
    */
   readonly resumeSuspendedSessions: Effect.Effect<void>
 }
@@ -66,6 +69,7 @@ export const layer = (options?: Options) =>
     Effect.gen(function* () {
       const store = yield* SessionStore.Service
       const execution = yield* SessionExecution.Service
+      const resolve = yield* SessionResolve.Service
       const bus = yield* Bus.Service
       const jobs = yield* Job.Service
       const sessions = yield* Session.Service
@@ -73,6 +77,8 @@ export const layer = (options?: Options) =>
       const maxAttempts = options?.maxAttempts ?? DEFAULT_MAX_ATTEMPTS
 
       const prepareResume = Effect.fnUntraced(function* (sessionID: SessionSchema.ID) {
+        // Reopening capabilities does not opt a Session into automatic recovery.
+        if (resolve.status(sessionID) !== "unowned") return undefined
         // Durable before the resume runs, so a crash inside the resumed turn is
         // counted by the next sweep and the budget cannot be dodged.
         const attempts = yield* store.countResume(sessionID)
@@ -95,6 +101,14 @@ export const layer = (options?: Options) =>
         return true
       })
 
+      const eligibleJob = (recovery: Job.Recovery) => {
+        if (recovery.kind === "shell") return resolve.status(recovery.sessionID) === "unowned"
+        return (
+          resolve.status(recovery.parentSessionID) === "unowned" &&
+          resolve.status(recovery.childSessionID) === "unowned"
+        )
+      }
+
       const recoverShell = Effect.fnUntraced(function* (
         background: Job.Background,
         recovery: Extract<Job.Recovery, { kind: "shell" }>,
@@ -143,6 +157,7 @@ export const layer = (options?: Options) =>
 
         const notify = Effect.fnUntraced(function* (result: Pick<Job.Background, "status" | "output" | "error">) {
           if (result.status === "running") return
+          if (!eligibleJob(recovery)) return
           const text =
             result.status === "completed"
               ? (result.output ?? "Subagent completed without a text response.")
@@ -172,7 +187,9 @@ export const layer = (options?: Options) =>
           return
         }
         if ((yield* execution.active).has(recovery.childSessionID)) return
-        if (!(yield* prepareResume(recovery.childSessionID))) {
+        const prepared = yield* prepareResume(recovery.childSessionID)
+        if (prepared === undefined) return
+        if (!prepared) {
           yield* notify({ status: "error", error: RESUME_EXHAUSTED.message })
           return
         }
@@ -214,18 +231,20 @@ export const layer = (options?: Options) =>
           // Early notices wait for root recovery's accounting, including roots that exhaust their budget.
           const suspended = new Set((yield* store.listSuspended()).filter((sessionID) => !active.has(sessionID)))
           const pending = yield* jobs.pendingBackground
-          yield* store.releaseChildClaims(
-            pending.flatMap((background) =>
+          yield* store.releaseChildClaims([
+            ...(yield* resolve.ownedIDs),
+            ...pending.flatMap((background) =>
               background.status === "running" && background.recovery.kind === "subagent"
                 ? [background.recovery.childSessionID]
                 : [],
             ),
-          )
+          ])
           yield* Effect.forEach(
             pending,
             Effect.fnUntraced(function* (background) {
               if ((yield* jobs.get(background.id))?.status === "running") return
               const recovery = background.recovery
+              if (!eligibleJob(recovery)) return
               yield* recovery.kind === "shell"
                 ? recoverShell(background, recovery)
                 : recoverSubagent(background, recovery, suspended)
@@ -237,10 +256,10 @@ export const layer = (options?: Options) =>
           const resumed = yield* execution.active
           yield* Effect.forEach(
             (yield* store.listSuspended()).filter((sessionID) => !resumed.has(sessionID)),
-            (sessionID) =>
-              execution
-                .resume(sessionID)
-                .pipe(Effect.ignore, Effect.forkIn(scope), Effect.when(prepareResume(sessionID))),
+            Effect.fnUntraced(function* (sessionID) {
+              if (!(yield* prepareResume(sessionID))) return
+              yield* execution.resume(sessionID).pipe(Effect.ignore, Effect.forkIn(scope))
+            }),
             { concurrency: "unbounded", discard: true },
           )
           // Async observers consult this set at delivery; later completions wake parents normally.
@@ -253,5 +272,5 @@ export const layer = (options?: Options) =>
 export const node = makeGlobalNode({
   service: Service,
   layer: layer(),
-  deps: [SessionStore.node, SessionExecution.node, Bus.node, Job.node, Session.node],
+  deps: [SessionStore.node, SessionExecution.node, SessionResolve.node, Bus.node, Job.node, Session.node],
 })

packages/core/src/session/generate-node.ts

@@ -9,7 +9,6 @@ import { SessionContext } from "./context.js"
 import { SessionGenerate } from "./generate.js"
 import { SessionHistory } from "./history.js"
 import { SessionModelRequest } from "./model-request.js"
-import { SessionRunnerModel } from "./runner/model.js"
 
 export const layer = Layer.effect(
   SessionGenerate.Service,
@@ -17,13 +16,11 @@ export const layer = Layer.effect(
     const context = yield* SessionContext.Service
     const database = yield* Database.Service
     const llm = yield* LLMClient.Service
-    const models = yield* SessionRunnerModel.Service
-    const modelRequests = yield* SessionModelRequest.Service
 
     return SessionGenerate.Service.of({
       generate: Effect.fn("SessionGenerate.generate")(function* (input) {
         const selection = yield* context.select(input.sessionID)
-        const model = yield* models.resolve(selection.session)
+        const model = yield* context.resolveModel(selection.session)
         const history = yield* SessionHistory.preview(database.db, selection.session.id, selection.instructions)
         const transcript = SessionModelRequest.baseTranscript({
           agent: selection.agent.info,
@@ -32,7 +29,7 @@ export const layer = Layer.effect(
           initial: history.initial,
           messages: history.messages,
         })
-        const prepared = yield* modelRequests.prepare({
+        const prepared = yield* context.prepare({
           scope: { session: selection.session, agentID: selection.agent.id, model, tools: selection.tools },
           transcript: {
             system: transcript.system,
@@ -59,5 +56,5 @@ export const layer = Layer.effect(
 export const node = makeLocationNode({
   service: SessionGenerate.Service,
   layer,
-  deps: [SessionContext.node, Database.node, SessionModelRequest.node, SessionRunnerModel.node, llmClient],
+  deps: [SessionContext.node, Database.node, llmClient],
 })

packages/core/src/session/instruction-entry.ts

@@ -103,7 +103,7 @@ const source = (entry: Info & { readonly removed: boolean }) =>
     },
   })
 
-const layer = Layer.effect(
+export const layer = Layer.effect(
   Service,
   Effect.gen(function* () {
     const { db } = yield* Database.Service

packages/core/src/session/model-request.ts

@@ -84,14 +84,16 @@ export const baseTranscript = (input: {
   readonly tools: Tool.Snapshot
   readonly initial: string
   readonly messages: ReadonlyArray<SessionMessage.Info>
+  readonly system?: string
 }) => {
   const providerMetadataKey = input.model.model.route.providerMetadataKey ?? input.model.model.provider
   return {
     providerMetadataKey,
     system: [
-      input.agent.system
-        ? input.agent.system
-        : SessionSystemPrompt.make(input.tools.definitions.map((tool) => tool.name)),
+      input.system ??
+        (input.agent.system
+          ? input.agent.system
+          : SessionSystemPrompt.make(input.tools.definitions.map((tool) => tool.name))),
       input.initial,
     ]
       .filter((part) => part.length > 0)

packages/core/src/session/resolve.ts

@@ -0,0 +1,291 @@
+export * as SessionResolve from "./resolve.js"
+
+import type { LLMClientService } from "@opencode-ai/ai"
+import { FSUtil } from "@opencode-ai/util/fs-util"
+import { Global } from "@opencode-ai/util/global"
+import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
+import { Context, Deferred, Effect, Exit, Layer, Scope } from "effect"
+import { Bus } from "../bus.js"
+import { App } from "../app.js"
+import { Database } from "../database/database.js"
+import { llmClient, webSocketConstructor } from "../effect/app-node-platform.js"
+import { KV } from "../kv.js"
+import type { SessionCapabilities } from "./capabilities.js"
+import type { SessionRunner } from "./runner/index.js"
+import { SessionSchema } from "./schema.js"
+import { SessionStore } from "./store.js"
+import { Socket } from "effect/unstable/socket"
+import type { Image } from "../image.js"
+import type { SessionContext } from "./context.js"
+import type { SessionModelTransport } from "./model-transport.js"
+
+const prefix = "session.capabilities/"
+
+/**
+ * Live operations, never an attempt snapshot. Title, request hooks, compaction,
+ * media/skills, snapshots, output, and transport customization are deferred;
+ * open supplies their internal defaults without directory discovery.
+ */
+interface Capabilities extends SessionContext.Interface {
+  readonly image: Image.Interface
+  readonly transport: SessionModelTransport.Interface
+}
+
+type Status = "attached" | "owned-detached" | "unowned"
+type Resolved =
+  | { readonly status: "attached"; readonly capabilities: Capabilities }
+  | { readonly status: "owned-detached" | "unowned" }
+
+type Opened = {
+  readonly capabilities: Capabilities
+  readonly runner: SessionRunner.Interface
+  readonly scope: Scope.Closeable
+  readonly onRetire: () => Effect.Effect<void>
+  readonly done: Deferred.Deferred<void>
+  current: boolean
+  users: number
+}
+
+export interface Interface {
+  readonly own: <A extends { readonly id: SessionSchema.ID }, R>(
+    record: Effect.Effect<A, never, R>,
+  ) => Effect.Effect<A, never, R>
+  readonly status: (id: SessionSchema.ID) => Status
+  readonly ownedIDs: Effect.Effect<ReadonlyArray<SessionSchema.ID>>
+  readonly attach: (
+    id: SessionSchema.ID,
+    input: SessionCapabilities.OpenInput,
+  ) => Effect.Effect<() => Effect.Effect<void>>
+  readonly resolve: (session: SessionSchema.Info) => Effect.Effect<Resolved, never, Scope.Scope>
+  /** Called synchronously when the coordinator installs a busy period, before its first fiber yield. */
+  readonly pin: (id: SessionSchema.ID) => void
+  readonly pinned: (id: SessionSchema.ID) => SessionRunner.Interface | undefined
+  readonly settle: (id: SessionSchema.ID) => Effect.Effect<void>
+  readonly remove: (id: SessionSchema.ID) => Effect.Effect<void>
+}
+
+export class Service extends Context.Service<Service, Interface>()("@opencode/SessionResolve") {}
+
+export const layer = Layer.effect(
+  Service,
+  Effect.gen(function* () {
+    const kv = yield* KV.Service
+    const db = (yield* Database.Service).db
+    const owned = new Set(
+      (yield* kv.scanAll(prefix))
+        .filter((entry) => entry.value === true)
+        .map((entry) => SessionSchema.ID.make(entry.key.slice(prefix.length))),
+    )
+    const globals = yield* Effect.context<
+      | Database.Service
+      | Bus.Service
+      | SessionStore.Service
+      | LLMClientService
+      | FSUtil.Service
+      | Global.Service
+      | Socket.WebSocketConstructor
+    >()
+    const current = new Map<SessionSchema.ID, Opened>()
+    const pinned = new Map<SessionSchema.ID, Opened>()
+    const opened = new Map<Deferred.Deferred<void>, Opened>()
+    // LayerMap invalidation cannot choose synchronously at coordinator start or
+    // run a host hook after all generation-specific users settle. Keep explicit leases.
+    const retire = (value: Opened) =>
+      Effect.suspend(() => {
+        if (value.current || value.users > 0 || !opened.delete(value.done)) return Effect.void
+        return Scope.close(value.scope, Exit.void).pipe(
+          Effect.andThen(value.onRetire()),
+          Effect.onExit((exit) => Deferred.done(value.done, exit)),
+        )
+      })
+    const release = (value: Opened) =>
+      Effect.sync(() => {
+        value.users--
+      }).pipe(Effect.andThen(retire(value)))
+    yield* Effect.addFinalizer(() =>
+      Effect.sync(() => current.clear()).pipe(
+        Effect.andThen(
+          Effect.forEach(
+            opened.values(),
+            (value) => {
+              value.current = false
+              return retire(value)
+            },
+            { discard: true },
+          ),
+        ),
+      ),
+    )
+
+    return Service.of({
+      // Ownership is transitionally one-way: retirement never hands a Session back to discovery.
+      own: (record) =>
+        Effect.uninterruptible(
+          Effect.gen(function* () {
+            const session = yield* db
+              .transaction(() =>
+                Effect.gen(function* () {
+                  const session = yield* record
+                  if (!owned.has(session.id)) yield* kv.set(prefix + session.id, true)
+                  return session
+                }),
+              )
+              .pipe(Effect.orDie)
+            // Publish the memory index only after the durable transaction commits.
+            owned.add(session.id)
+            return session
+          }),
+        ),
+      status: (id) => (current.has(id) ? "attached" : owned.has(id) ? "owned-detached" : "unowned"),
+      ownedIDs: Effect.sync(() => Array.from(owned)),
+      attach: Effect.fn("SessionResolve.attach")(function* (id, input) {
+        const [
+          { Image },
+          { PluginHooks },
+          { PluginSupervisor },
+          { Snapshot },
+          { ToolOutput },
+          { SessionCompaction },
+          { SessionContext },
+          { InstructionEntry },
+          { SessionModelRequest },
+          { SessionModelTransport },
+          { SessionRunner },
+          { SessionRunnerLLM },
+          { SessionTitle },
+        ] = yield* Effect.promise(() =>
+          Promise.all([
+            import("../image.js"),
+            import("../plugin/hooks.js"),
+            import("../plugin/supervisor-service.js"),
+            import("../snapshot.js"),
+            import("../tool-output.js"),
+            import("./compaction.js"),
+            import("./context.js"),
+            import("./instruction-entry.js"),
+            import("./model-request.js"),
+            import("./model-transport.js"),
+            import("./runner/index.js"),
+            import("./runner/llm.js"),
+            import("./title.js"),
+          ]),
+        )
+        const scope = yield* Scope.make()
+        const base = Layer.mergeAll(
+          PluginHooks.layer,
+          Image.layer,
+          InstructionEntry.layer,
+          SessionModelTransport.layer,
+          ToolOutput.layer,
+          Snapshot.noopLayer,
+          SessionCompaction.layer,
+          Layer.succeed(PluginSupervisor.Service, { flush: Effect.void }),
+        ).pipe(Layer.provide(Layer.succeedContext(globals)))
+        const requests = SessionModelRequest.layer.pipe(Layer.provideMerge(base))
+        const capabilities = SessionContext.values(input).pipe(Layer.provideMerge(requests))
+        const runner = SessionRunnerLLM.layer.pipe(
+          Layer.provideMerge(SessionTitle.layer.pipe(Layer.provideMerge(capabilities))),
+          Layer.provide(Layer.succeedContext(globals)),
+        )
+        // Each open builds fresh local state over the SAME captured durable/global services.
+        const services = yield* Layer.buildWithScope(Layer.fresh(runner), scope).pipe(
+          Effect.onError(() => Scope.close(scope, Exit.void)),
+        )
+        const value: Opened = {
+          capabilities: {
+            ...Context.get(services, SessionContext.Service),
+            image: Context.get(services, Image.Service),
+            transport: Context.get(services, SessionModelTransport.Service),
+          },
+          runner: Context.get(services, SessionRunner.Service),
+          scope,
+          onRetire: input.retire ?? (() => Effect.void),
+          done: Deferred.makeUnsafe<void>(),
+          current: true,
+          users: 0,
+        }
+        const previous = current.get(id)
+        current.set(id, value)
+        opened.set(value.done, value)
+        if (previous) {
+          previous.current = false
+          yield* retire(previous)
+        }
+        // Closed handles retain only completion, not retired capability functions or layers.
+        const done = value.done
+        return () =>
+          Effect.gen(function* () {
+            yield* Effect.uninterruptible(
+              Effect.gen(function* () {
+                const value = opened.get(done)
+                if (!value) return
+                if (current.get(id) === value) current.delete(id)
+                value.current = false
+                yield* retire(value)
+              }),
+            )
+            yield* Deferred.await(done)
+          })
+      }),
+      resolve: (session) =>
+        Effect.gen(function* () {
+          const value = current.get(session.id)
+          if (!value) return { status: owned.has(session.id) ? ("owned-detached" as const) : ("unowned" as const) }
+          yield* Effect.acquireRelease(
+            Effect.sync(() => {
+              value.users++
+            }),
+            () => release(value),
+          )
+          return { status: "attached" as const, capabilities: value.capabilities }
+        }),
+      pin: (id) => {
+        const value = current.get(id)
+        if (!value) return
+        value.users++
+        pinned.set(id, value)
+      },
+      pinned: (id) => pinned.get(id)?.runner,
+      settle: (id) =>
+        Effect.suspend(() => {
+          const value = pinned.get(id)
+          if (!value) return Effect.void
+          pinned.delete(id)
+          return release(value)
+        }),
+      remove: (id) =>
+        Effect.gen(function* () {
+          const value = current.get(id)
+          current.delete(id)
+          if (value) {
+            value.current = false
+            yield* retire(value)
+          }
+          if (!owned.has(id)) return
+          yield* kv.remove(prefix + id)
+          owned.delete(id)
+        }),
+    })
+  }),
+)
+
+export const node = makeGlobalNode({
+  service: Service,
+  layer,
+  deps: [
+    KV.node,
+    Database.node,
+    Bus.node,
+    SessionStore.node,
+    llmClient,
+    FSUtil.node,
+    Global.node,
+    webSocketConstructor,
+    // Request preparation reads this Reference from the captured globals, not its fallback metadata.
+    App.node,
+  ],
+})
+
+/** TODO: replace this defect with the typed error in the capability-gated operations phase. */
+export const unavailable = (sessionID: SessionSchema.ID) =>
+  Effect.die(new Error(`Session must be reopened with capabilities: ${sessionID}`))

packages/core/src/session/run-coordinator.ts

@@ -52,6 +52,8 @@ type Execution<E, Reason> = {
  */
 export const make = <Key, E, Reason = never>(options: {
   readonly drain: (key: Key, force: boolean, scope: Promotable) => Effect.Effect<void, E>
+  /** Controls new busy periods, including late successors; existing execution may finish. */
+  readonly eligible?: (key: Key) => boolean
   /** Runs once when a process-local busy period begins, before its first drain. */
   readonly started?: (key: Key) => Effect.Effect<void>
   /**
@@ -107,7 +109,7 @@ export const make = <Key, E, Reason = never>(options: {
     // A doorbell that survives the execution loop (rung after the loop decided to end, or
     // during failure or interruption cleanup) starts a fresh execution for the remaining work.
     const settle = (key: Key, execution: Execution<E, Reason>, exit: Exit.Exit<void, E>) => {
-      if (execution.pendingWake) start(key, false, execution.pendingWake)
+      if (execution.pendingWake && (options.eligible?.(key) ?? true)) start(key, false, execution.pendingWake)
       else executions.delete(key)
       Deferred.doneUnsafe(execution.done, exit)
     }
@@ -121,11 +123,13 @@ export const make = <Key, E, Reason = never>(options: {
             return Deferred.await(execution.done).pipe(Effect.ignoreCause, Effect.andThen(run(key)))
           return Deferred.await(execution.done)
         }
+        if (options.eligible?.(key) === false) return Effect.interrupt
         return Deferred.await(start(key, true, "input").done)
       })
 
     const wake = (key: Key, scope: Promotable = "input") =>
       Effect.sync(() => {
+        if (options.eligible?.(key) === false) return
         const execution = executions.get(key)
         if (execution !== undefined) {
           // Coalesced wakes keep the widest scope: "input" subsumes "steer".

packages/core/src/session/runner/llm.ts

@@ -30,13 +30,12 @@ import { MAX_STEPS_PROMPT } from "./max-steps.js"
 const CONTINUE_AFTER_INCOMPLETE_STREAM =
   "The previous response was interrupted. Continue from where you left off without repeating completed content."
 
-const layer = Layer.effect(
+export const layer = Layer.effect(
   Service,
   Effect.gen(function* () {
     const bus = yield* Bus.Service
     const store = yield* SessionStore.Service
     const context = yield* SessionContext.Service
-    const modelRequests = yield* SessionModelRequest.Service
     const modelTransport = yield* SessionModelTransport.Service
     const db = (yield* Database.Service).db
     const compaction = yield* SessionCompaction.Service
@@ -142,6 +141,8 @@ const layer = Layer.effect(
                   Effect.gen(function* () {
                     return yield* compaction.compactManual({
                       session,
+                      resolveModel: context.resolveModel,
+                      prepare: context.prepare,
                       messages: yield* store.context(sessionID),
                       inputID: pending.id,
                       started: true,
@@ -215,7 +216,12 @@ const layer = Layer.effect(
         // Reuse boundary preparation once; retries refresh context without delivering more input.
         const loaded = initial ?? (yield* prepareContext(sessionID).pipe(Effect.flatMap(context.load)))
         initial = undefined
-        const compactionInput = { session: loaded.session, messages: loaded.messages, resolved: loaded.model }
+        const compactionInput = {
+          session: loaded.session,
+          messages: loaded.messages,
+          resolved: loaded.model,
+          prepare: context.prepare,
+        }
         if (compaction.required(compactionInput)) {
           const compacted = yield* compaction.compact(compactionInput)
           if (compacted.status !== "completed") return yield* new StepFailedError({ error: compacted.error })
@@ -229,8 +235,9 @@ const layer = Layer.effect(
           tools: loaded.tools,
           initial: loaded.initial,
           messages: loaded.messages,
+          system: loaded.system,
         })
-        const prepared = yield* modelRequests.prepare({
+        const prepared = yield* context.prepare({
           scope: { session: loaded.session, agentID: loaded.agent.id, model: loaded.model, tools: loaded.tools },
           transcript: {
             system: transcript.system,
@@ -319,7 +326,6 @@ export const node = makeLocationNode({
     Bus.node,
     llmClient,
     SessionContext.node,
-    SessionModelRequest.node,
     SessionModelTransport.node,
     SessionStore.node,
     SessionCompaction.node,