The components leg ran the session UI and app component suites back to
back on one runner, which made it the slowest Windows e2e leg (session UI
3.2 min plus app 3.5 min plus setup, about 8.8 min, against 7-8 min for
each app shard). Each suite now gets its own leg.
The app suite no longer inherits the session UI suite's Vite optimizer
cache, so it always starts cold. The Storybook dependency pre-bundling
below this change is what keeps a cold start from reloading pages.
main.ts removes storybook:optimize-deps-plugin and there are no HTML
files under the Vite root, so Vite's startup scan finds nothing and
discovers every dependency on first import. Each discovery re-runs the
optimizer and broadcasts a full reload to every page in both workers,
which aborts tests mid-import ("Execution context was destroyed",
"Failed to fetch dynamically imported module") and is why the app
component suite is flaky on every CI run.
optimizeDeps.entries now covers the stories, the *.worker.ts files (the
scan does not follow new Worker(new URL(...))) and the modules component
tests load by URL, read from the ./ and ../ literals in the specs so a
new spec needs no config change. Based on the approach in #55632.
On top of that, preview.tsx is an entry and the Storybook preview
runtime is included. Without them the preview's addons and the runtime
are still found late, and one reload remains whenever the first browser
request arrives after the startup optimize has finished, which is the
usual case because the webServer is ready once the manager page loads.
The webServer's stdout is now piped, so any "optimized dependencies
changed. reloading" line shows up in CI logs instead of being dropped.
retain-on-failure records and encodes a video for every component test
only to drop it when the test passes, which shows up in CI as 50-70 ms of
context close per passing test. In CI, record it on the first retry
instead, next to the trace; first-attempt failures still keep their
screenshot and error context. Local runs have no retries, so they keep
retain-on-failure.
addon-a11y also loads axe-core and scans the page after every story
render. preview.tsx only marks results as todo and no component test
reads them, so mount now always sets the a11y.manual global, which makes
the addon skip that run. Callers' globals are merged over it, and
browsing Storybook directly still runs axe.
mobile-timeline-scroll sets its own launchOptions, so its 37 tests get a
separate worker hash and, without fullyParallel, form one serial group.
In CI that group takes 3 to 5 minutes. Unsharded it is always the last
file and mostly runs alone; on shard 2 it starts first and sets the
length of the whole leg.
A second project now runs only that file, with fullyParallel on.
Playwright queues projects in declaration order, so these tests start
after every chromium group and spread over all workers. Both projects
use one regex, so the file always runs in exactly one project; a rename
that does not update the regex quietly returns it to the serial schedule,
without losing or duplicating tests. The spec and its launchOptions are unchanged, the shard
split stays at 225/224 tests, and the performance configs list the same
tests. `--project chromium`, UI mode and the VS Code extension now skip the file
by default.
The cost is more contention for these touch tests, which now run beside
each other on every worker, so CI retries for this file need watching.
vite preview checks the case of every requested file on case-insensitive
file systems (Windows, macOS) with a synchronous readdir of its parent
directory. The fixture build has about 3100 files in dist/_assets, so each
of the roughly 150 asset requests per page load blocks the one server
that every worker shares. That matches the CI traces, where Windows asset
waits and first navigations grow with the number of tests running.
e2e/utils/serve.ts serves dist with Bun instead and keeps vite preview's
routing rules: files resolve only inside dist, a malformed escape is a
404, and index.html is the fallback only for GET or HEAD requests whose
Accept is missing, empty, text/html or */*, never for /favicon.ico. A
missing script fetched as JavaScript or an API request asking for JSON
still gets a 404. Bun fails on a busy port like --strictPort did, and
reuseExistingServer and the dev-mode command are unchanged.
The performance configs keep vite preview so their baselines stay
comparable.
Most app e2e tests let the production build register its PWA service
worker, which then precaches the whole 35 MB build from the shared preview
server while the test runs. No app e2e spec depends on the worker, and
e2e/service-worker keeps that coverage under its own config, so block it
in the shared config. The 11 per-file blocks are now redundant; they are
left alone to keep this change to one file.
retain-on-failure records and encodes a video for every test only to drop
it when the test passes. In CI, record it on the first retry instead,
next to the trace; first-attempt failures still keep their screenshot and
error context. Local runs have no retries, so they keep retain-on-failure.
The performance, uncapped, timeline-stability and projection configs
spread this config and now block service workers too. CI runs none of
them.
@corvu/drawer 0.2.4 starts its opening transition two frames after it opens. A dismissal before then slid the drawer open while it was closed. A dismissal in that frame started no transition, or cancelled it, so the transitionend that finishes a close never arrived, leaving focus trapped and body pointer-events disabled. Patch it to skip the late opening step once closed, and to finish a close when no transition is running.
Also stop non-fullscreen story pages from scrolling by 32px inside Storybook's padding. Wait for web fonts before the composer overflow story measures the toolbar. Add composer tests that dismiss the variant drawer before and as it starts opening; both fail without the patch.
The overflow story also waits for focus to return to the add button after Escape before measuring, so the menu focus restore cannot land between select-all and Delete.
TanStack resets isScrolling on a 150ms timer that reports the offset
captured at the last scroll event. If the virtualizer wrote scrollTop
after that event (prepend anchoring or size compensation) and the
browser had not yet dispatched the write's own scroll event, the timer
put the old offset back. The next size compensation was then computed
from it and undid the write, so the transcript jumped to the top of a
newly loaded page or skipped rows while scrolling up.
When a scroll settles, the reconnect-aware offset observer now reports
the element's live scrollTop. Detached elements keep the event offset.
Chromium's media controls answer the second click of a double click with a
native fullscreen request for the <video>. The same double click opens the
in-window theatre. Where fullscreen is granted (Chrome, Playwright), the
fullscreened video covers the theatre backdrop, so a backdrop click no longer
closes it. That is the cause of the flaky review.spec.ts video theatre test.
controlslist="nofullscreen" would also remove Chromium's click-to-play,
which it only offers on videos that can go fullscreen. Instead, the video is
held as an open popover while the controls handle a multi-click, and released
before the next frame renders. The fullscreen element ready check refuses an
open popover, so only the fullscreen request fails. Chromium's play/pause
bookkeeping is unchanged. Desktop already denies the fullscreen permission,
so nothing changes there.
The e2e test now also checks that a single click still starts playback, and
that nothing went native fullscreen by the end of the flow.
Chromium recreates every CSS-connected FontFace as "unloaded" whenever a
lazily loaded stylesheet with @layer rules becomes active ahead of a sheet
that is already applied. The route's settings and text-input CSS chunks do
this when they finish out of order, which is common on Windows runners. The
downloaded font file is kept, so the first code that lays out still uses
IBM Plex Mono, but the face's status stays "unloaded" and polling
document.fonts.check() timed out.
Wait for document.fonts.ready, then mount one <code> and check once that
its first layout already has the font loaded. A preload that only downloads
the file still fails, because that first layout reports "loading".
turbo.json defines a test task per package and no generic one, so the
test scripts of client, gui-extensions, http-recorder, merman, plugin,
plugin-browser, script, sdk, simulation, theme and util never ran in CI
(119 test files). Register them like the other suites.
These suites had no turbo test task, so they rotted unnoticed.
- client: update tests for intended API changes (session.view idle is a
DateTime, interrupt's continue became resume, new API groups, the
Effect service no longer imports protocol), and assert the service
shutdown per platform since Windows cannot catch SIGTERM.
- http-recorder: run the record-mode WebSocket test with CI unset;
CI=true forces replay mode and the test waited forever for "open".
- util: reset the retry test's socket after the request arrives instead
of racing the client's write; set TEMP for the child on Windows and
compare with the native realpath; exclude colocated tests from the
build, which compiled them into dist where bun test ran them twice.
- sdk: update two tests for intended behaviour changes, keeping the
unknown-session error on the HTTP transport.
- merman: hold the established layout families to their original
ceilings and give decision-loops its own, drop the wall-clock
assertion, and run the cheap label distance check before the clearance
check (same output, half the audit CPU).
websearch.query resolved the provider without waiting for the Location's
plugins to activate, so the first query on a cold Location failed with
websearch_provider_not_found for a provider a plugin was about to
register. It now waits like the model and generate handlers do, bounded
by the same 5s activation wait. websearch.providers stays non-blocking.
command-selection typed into the variant dialog before its filter had
focus (DialogSelect focuses on a 1ms timer). Markdown text is drawn only
after the tree-sitter worker replies, which can take longer than the 20
frames waitForFrame allows on a busy runner. Wait for focus, and add
waitForHighlightedFrame to wait for pending highlights inside the
predicate.
The dialog stack keeps a closing dialog for 100ms before disposing it,
and useDialog().active kept returning it, so the global shortcut handler
dropped every shortcut in that window. Reopening the command palette
right after closing it did nothing. active now returns the top dialog
that has not started closing, the same rule the focus trap already uses.
- whats-new: What's New opens as a modal 500ms after the changelog
response and hides #root, so on Windows start() often looked for the
Settings button after it was hidden. The dialog tests hold the
changelog route until Home has been checked.
- subagent-child-navigation: the parent title was measured while the
loading header could still be swapped for the timeline header, so it
sometimes measured a detached heading (left 0). Wait for the timeline
row first.
- session-summary: the tooltip was hovered during the timeline's first
mount, whose final scroll closes it. Wait for the rows to be revealed.
- tabs: page.clock.pauseAt(new Date()) raced the still-running page
clock ("Cannot fast-forward to the past"). Install at an explicit time
and pause at a fixed offset from it.
Windows CI installed with --linker hoisted to work around
oven-sh/bun#28147, where an isolated install re-applied a patched
dependency per peer variant and failed with EPERM while hardlinking. That
issue was fixed in July, and a full test run on Windows with the isolated
linker installs and passes. The install step drops from about 95s to
about 36s on each Windows job.
A test task's hash only covered its direct dependencies' build tasks. tui
and script have no build task, so core, codemode, plugin-browser and ui
never reached cli#test's hash, and Linux replayed cached cli test passes
from the runner's remote cache for commits that changed core. Every test
task now also depends on a transit task, which runs nothing but folds the
whole workspace closure into its hash. The root package.json is a global
dependency, so a Bun upgrade through packageManager invalidates results.
Windows unit jobs had no Turbo cache at all. Restore one keyed by week on
Windows, save it from pushes to dev, and skip it on manual runs.
On the Windows runners, restoring the ~720 MB bun cache spends about 37s
extracting it with tar onto NTFS, and the warm install still takes about
64s. A cold install without the cache is faster than restore plus warm
install in every paired run we sampled, so Windows skips the restore and
the save. Linux keeps the cache.
Split each OS's e2e job into three legs: two Playwright shards of the app
e2e suite and one leg for the session UI and app component tests. The
service worker suite reads the production build, so it runs on shard 2
after the e2e step that builds it.
Keep "e2e (linux)" and "e2e (windows)" as the names of small aggregator
jobs, since the branch rules require them. They pass only when every leg
passed, or when e2e was not needed, and fail if the affected job failed.
Move the compiled-service and Node build smoke steps out of the unit job
into a parallel build job. They took about 2.5 minutes after the tests on
each OS and do not depend on anything the tests produce.
Remove the Turbo actions/cache steps. They cached node_modules/.cache/turbo,
which Turbo 2 never writes (its local cache is .turbo/cache), so they never
restored or saved anything. On Windows, hashing the cache key still cost
about 30 seconds before and after the tests.
Check out two commits on pull requests instead of the full history. The
affected-package diff only needs the merge commit and its first parent.
The Storybook Vite config aliased @tanstack/solid-query to
packages/app/node_modules/@tanstack/solid-query. That path only exists with
bun's isolated linker. Windows CI installs with --linker hoisted, so Vite
returned 404 for every module importing solid-query and the app component
tests timed out on Windows.
Dedupe the package instead; Storybook declares it, so it resolves to one copy
under both linkers and both runtimes.
Also run the e2e job for Storybook-only changes. No app package depends on
Storybook, so those PRs skipped every component test and reported green.
Takes the v2 tree, with packages/console, packages/web, packages/stats, packages/function, infra and sst.config.ts merged up to latest dev, and retargets V2 publishing and deploys from the v2 branch to dev.
Session rows have no id, so reconcile matched them by position. Each page of older history shifted every row into a different store object and remounted the visible transcript. Newly mounted markdown is blank until highlighting finishes, so the transcript blinked once per page. This was most visible with Low verbosity, where opening a session loads several pages to complete the oldest group.
Cross-compiled darwin binaries ship with an invalid Bun signature, which breaks codesign verification and publisher-based allowlists. Sign them on macOS with the Developer ID certificate and hardened runtime, matching V2, then repack the release zips and publish the signed binaries to npm and Homebrew.
Refs #46313
Users who were migrated to the new interface will now see a dismissible notice in Settings explaining the old interface was phased out, so they are not surprised by the change. The new-interface toggle is also visually elevated with a card style so it is easier to find. Settings labels and descriptions across General have been rewritten to be shorter and more direct, so it is quicker to understand what each option controls. The "New" badge color is also adjusted for better readability.
Compute all stat dimensions and both grains in one GROUPING SETS query
so each sync pass scans the source table once instead of eight times.
Hourly passes now only recompute the current ISO week; a daily full pass
refreshes the whole display window. Restarted daemons resume the hourly
cadence from the last completed sync instead of immediately re-running a
pass, and the workgroup kills any query scanning more than 2 TB.
Cherry-picks the packages/core changes from the v2 branch onto dev:
- combined ordered stdout/stderr in AppProcess + bash structured output
- edit/apply-patch return FileDiff info with status and line stats
- ignore no-op model switches; record reasoning timestamps
- return unexpected local tool defects to the model and continue
- keep OAuth account metadata out of request bodies
- nest OpenAI reasoning effort/summary options
- load OpenCode provider config asynchronously; batch plugin boot
- export latest public event manifest
Includes the supporting schema reasoning time field and regenerated
client/SDK types.
const commentAge = now - new Date(complianceComment.created_at).getTime();
if (commentAge < twoHours) {
core.info(`${kind} #${item.number} still within 2-hour window (${Math.round(commentAge / 60000)}m elapsed)`);
if (commentAge < seventyTwoHours) {
core.info(`${kind} #${item.number} still within 72-hour window (${Math.round(commentAge / 60000)}m elapsed)`);
continue;
}
const closeMessage = isPR
? 'This pull request has been automatically closed because it was not updated to meet our [contributing guidelines](../blob/dev/CONTRIBUTING.md) within the 2-hour window.\n\nFeel free to open a new pull request that follows our guidelines.'
: 'This issue has been automatically closed because it was not updated to meet our [contributing guidelines](../blob/dev/CONTRIBUTING.md) within the 2-hour window.\n\nFeel free to open a new issue that follows our issue templates.';
? 'This pull request has been automatically closed because it was not updated to meet our [contributing guidelines](../blob/dev/CONTRIBUTING.md) within the 72-hour window.\n\nFeel free to open a new pull request that follows our guidelines.'
: 'This issue has been automatically closed because it was not updated to meet our [contributing guidelines](../blob/dev/CONTRIBUTING.md) within the 72-hour window.\n\nFeel free to open a new issue that follows our issue templates.';
await github.rest.issues.createComment({
owner: context.repo.owner,
@@ -129,5 +129,5 @@ jobs:
});
}
core.info(`Closed non-compliant ${kind} #${item.number} after 2-hour window`);
core.info(`Closed non-compliant ${kind} #${item.number} after 72-hour window`);
If the issue is NOT compliant and the author association is not OWNER or MEMBER, start the comment with:
<!-- issue-compliance -->
Then explain what needs to be fixed and that they have 2 hours to edit the issue before it is automatically closed. Also add the label needs:compliance to the issue using: gh issue edit ${{ github.event.issue.number }} --add-label needs:compliance
Then explain what needs to be fixed and that they have 72 hours to edit the issue before it is automatically closed. Also add the label needs:compliance to the issue using: gh issue edit ${{ github.event.issue.number }} --add-label needs:compliance
If duplicates were found, include a section about potential duplicates with links.
@@ -124,7 +124,7 @@ jobs:
**What needs to be fixed:**
- [specific reasons]
Please edit this issue to address the above within **2 hours**, or it will be automatically closed.
Please edit this issue to address the above within **72 hours**, or it will be automatically closed.
"message":"Extension code runs side effects through the SDK: createKeyed(source, fn, { otherwise }) per provider generation or value, createMemo or a plain function for derived values, createLatest for async data, createVisitState for per-visit state, and the handler for logic a user action causes. An escape hatch needs an oxlint-disable comment with a reason."
"message":"Extension code runs side effects through the SDK: createKeyed(source, fn, { otherwise }) per provider generation or value, createMemo or a plain function for derived values, createLatest for async data, createVisitState for per-visit state, and the handler for logic a user action causes. An escape hatch needs an oxlint-disable comment with a reason."
}
],
"patterns":[
{
"regex":"^@opencode/(app|desktop)(/|$)",
"message":"GUI extensions never import the app or desktop packages. Use the SDK."
},
{
"regex":"^@/",
"message":"GUI extensions never import app internals. Use the SDK."
},
{
"group":["../*/*","!../*/contract","!../sdk/*"],
"message":"Import another extension only through its contract.ts."
},
{
"regex":"\\.css$",
"message":"Import CSS with ?inline and contribute it with ctx.add(Style, css)."
}
]
}
]
}
},
{
"files":["packages/gui-extensions/src/sdk/**"],
"rules":{
"no-restricted-imports":[
"error",
{
"patterns":[
{
"regex":"^@opencode/(app|desktop)(/|$)",
"message":"GUI extensions never import the app or desktop packages. Use the SDK."
},
{
"regex":"^@/",
"message":"GUI extensions never import app internals. Use the SDK."
},
{
"group":["../*/*","!../*/contract","!../sdk/*"],
"message":"Import another extension only through its contract.ts."
},
{
"regex":"\\.css$",
"message":"Import CSS with ?inline and contribute it with ctx.add(Style, css)."
- After changing the public Protocol or Server `HttpApi`, run `bun run generate` from `packages/client`. Do not edit generated client files directly.
- Keep runtime dependencies directed from Schema to Core and Protocol, then from Core and Protocol to Server. Client runtime code may depend on Schema and Protocol but never Core or Server; `sdk` composes Client, Core, and Server.
- Current implementation changes belong in `packages/core`, `packages/cli`, `packages/server`, `packages/protocol`, `packages/schema`, and related generated client surfaces when required.
- This repository does not use Changesets. Do not add `.changeset` files; follow the existing release workflow instead.
- The default branch in this repo is `v2`.
- Default new branches and worktrees to `v2`, or `origin/v2` when the local `v2` ref is unavailable, and default pull requests to target `v2`. Use another base or target branch when the requester explicitly instructs it.
- Local `main` ref may not exist; use `v2` or `origin/v2` for diffs.
- The default branch is `dev`. Base branches and worktrees on `origin/dev`, target PRs at `dev`, and diff against `origin/dev` unless told otherwise.
## Live V2 TUI Testing
## Live TUI Testing
- Run `bun run dev:live` from a development worktree to test its TUI against the currently elected `opencode` background server and live sessions.
- Pass a directory after the script when needed, for example `bun run dev:live /path/to/project`.
- The script discovers the server with `opencode service status`, injects its private local credential from `opencode service get password`, and uses the `dev` TUI storage channel so tabs and other client-local state match the installed client.
- Prefer `dev:live` over plain `bun run dev` for this workflow. An implicit managed-service connection may replace the live server when the worktree client version differs; explicit `--server` warns and continues without replacing it.
## V2 TUI Stories
## TUI Stories
- When a user asks for a TUI story, add a fixture-driven story under `packages/tui/src/feature-plugins/system/storybook` and register it in `index.tsx`.
- Render the real production component rather than a visual copy. Keep submissions and other side effects local to the story so it is safe to explore repeatedly.
- Rely on type inference when possible; avoid explicit type annotations or interfaces unless necessary for exports or clarity
- Prefer functional array methods (flatMap, filter, map) over for loops; use type guards on filter to maintain type inference downstream
- In `src/config`, follow the existing self-export pattern at the top of the file (for example `export * as ConfigAgent from "./agent"`) when adding a new config module.
- In `src/config`, follow the existing self-export pattern at the top of the file (for example `export * as ConfigAgent from "./agent.js"`) when adding a new config module.
- In Effect generators, bind services to named variables before calling methods. Do not use nested service yields such as `yield* (yield* Foo.Service).bar()`.
Reduce total variable count by inlining when a value is only used once.
- Run `bun run check` from the repository root as the canonical full lint and type-check verification.
- During focused iteration, run `bun typecheck` from the affected package directory (for example, `packages/core`). Never run `tsc` directly.
## V2 Session Core
## Session Core
- Keep durable events minimal: record irreducible new facts and do not repeat state derivable by folding the ordered aggregate history. Enrich projections and read models with previous or derived state when consumers need self-contained views.
- Keep durable prompt admission separate from model execution. `Session.prompt(...)` publishes `session.inbox.enqueued`, whose projection inserts one durable `session_inbox` row, before scheduling advisory `SessionExecution.wake(sessionID)` unless `resume: false` requests admit-only behavior. Delivery publishes `session.inbox.delivered`; its projection consumes the inbox row and inserts the visible message in the same transaction. `session_inbox` stores only unconsumed work.
- Reusing a Session ID adopts the existing Session. Reusing a user or synthetic inbox item ID is idempotent when Session and type match: the first admission wins and the retried payload, metadata, and delivery mode are ignored, whether the item is still pending or already delivered (reconciled from the projected message without retained enqueue history). Cross-Session or cross-type reuse fails. Control items keep their operation-specific conflict behavior.
- Keep `SessionExecution` process-global and Session-ID based. Its local implementation owns the process-local Session coordinator and discovers placement through `SessionStore` plus `LocationServiceMap.get(session.location)` only when a drain starts; no layer should take a Session ID. V2 interruption targets the active process-local ownership chain for that Session; interruption of a known but idle or locally unowned Session is a no-op, while the public API rejects an unknown Session.
- Keep `SessionExecution` process-global and Session-ID based. Its local implementation owns the process-local Session coordinator and discovers placement through `SessionStore` plus `LocationServiceMap.get(session.location)` only when a drain starts; no layer should take a Session ID. Interruption targets the active process-local ownership chain for that Session; interruption of a known but idle or locally unowned Session is a no-op, while the public API rejects an unknown Session.
- Keep `SessionRunner`, model resolution, tool registry, permissions, and filesystem Location-scoped. Omitted `Location.workspaceID` means implicit-local placement; explicit workspace identity remains reserved for future placement semantics.
-Preserve one explicit `llm.stream(request)` call per Physical Attempt and reload projected history before durable continuation. A logical Step may use generic pre-output retries, one full-context retry after continuation rejection, incomplete-stream continuation, or one overflow-compaction rebuild. Generic retries retain the logical step number and do not consume another agent-step allowance. Do not delegate orchestration to an in-memory tool loop.
-Each step is one explicit `llm.stream` call that continues from reloaded history. Do not add an in-memory tool loop.
- Keep local Session drains process-local until clustering is implemented. `SessionRunCoordinator` joins explicit same-Session resumes, coalesces prompt wakeups, and allows different Sessions to run concurrently. A write-ahead execution claim marks a process-local busy period for restart recovery: terminal completion, failure, or user interruption releases it, while shutdown interruption and process death preserve it. Startup recovery resumes claimed top-level Sessions with durable per-execution attempt accounting. The claim is a recovery marker, not clustered ownership, fencing, or an exactly-once guarantee.
- Keep delivery vocabulary explicit. Prompts steer by default. At safe step boundaries, steered compaction takes priority up to the first steered move control; other steers retain enqueue order. At an idle boundary, steers take priority; otherwise exactly one queued item delivers before the runner reevaluates continuation. Inbox items may be cancelled or changed between queue and steer before delivery. Promoting new user input resets the selected agent's step allowance; a batch of steers resets it once.
- One step is one logical LLM call; its durable record covers only the model-visible span. Do not write "provider turn", and do not use bare "turn" for a single call: "turn" is reserved for the future assistant-turn unit containing all steps from prompt promotion until the session would go idle.
- Keep event replay ownership separate from clustered Session execution ownership.
- Keep the Instructions algebra and built-ins in `src/instructions`; keep instruction producers with their observed domains, and keep Session History selection plus `InstructionState` and `InstructionEntry` persistence Session-owned. `InstructionDiscovery` observes ambient global and upward-project instructions. The runner composes built-ins, discovery, guidance, and entries explicitly in `loadInstructions`; there is no instruction registry.
-`session.instructions.updated` stores changed source keys and content hashes and may freeze rendered chronological update text. Blob values live once in `instruction_blob`; the projected `instruction_state` row is the normal boundary-processing source of current and initial values. Request assembly renders the epoch baseline from stored values, while later frozen updates enter history as durable System messages. Completed compaction moves the instruction epoch; Session movement retains it so destination instruction changes are chronological, while committed revert clears it. Forks adopt the parent's newest instruction values even when copied message history ends at an earlier boundary. Unavailable sources retain the last value and block only the initial complete delta.
- **Remote Device Pairing**: In **Settings → Pairing**, select **Pair device** to enable OpenTunnel remote access and get a QR code or copyable link. ([#55306](https://github.com/anomalyco/opencode/pull/55306))
- **Provider Account Management**: Use the connect dialog or **Settings → Providers** to switch, rename, remove, or add provider accounts. ([#54101](https://github.com/anomalyco/opencode/pull/54101))
- **Automatic Response Continuation**: Responses that reach their output-token limit now continue automatically, preserving partial output instead of stopping abruptly. ([#53876](https://github.com/anomalyco/opencode/pull/53876))
- **Trackpad Image Navigation**: Pinch to zoom and use two-finger scrolling to pan image and SVG previews. Select **Fit** to reset the view. ([#54590](https://github.com/anomalyco/opencode/pull/54590))
### Core
#### Improvements
- Automatically continue responses after an output-token limit, with bounded retries and clearer recovery for truncated tool calls. ([#53876](https://github.com/anomalyco/opencode/pull/53876))
- Stabilize access policies under the top-level `policies` configuration key while retaining compatibility with `experimental.policies`. ([#55297](https://github.com/anomalyco/opencode/pull/55297))
#### Bug fixes
- Give centrally managed configuration precedence over local and user configuration. ([#55267](https://github.com/anomalyco/opencode/pull/55267))
- Report working sessions and sessions awaiting permission or input through terminal program-status notifications. ([#55172](https://github.com/anomalyco/opencode/pull/55172))
### Desktop
#### Improvements
- Pair devices from **Settings → Pairing** through OpenTunnel remote access, with a QR code and copyable link. ([#55306](https://github.com/anomalyco/opencode/pull/55306))
- Manage provider accounts from the connect dialog and **Settings → Providers**, including switching, renaming, removal, and adding another account. ([#54101](https://github.com/anomalyco/opencode/pull/54101))
- Show session context tokens and accumulated family cost in the session view. ([#53785](https://github.com/anomalyco/opencode/pull/53785))
- Pinch to zoom and use two-finger scrolling to pan image and SVG previews; select **Fit** to reset. ([#54590](https://github.com/anomalyco/opencode/pull/54590))
- Cycle side-panel tabs with the configured previous-tab and next-tab shortcuts while the side panel is focused. ([#54899](https://github.com/anomalyco/opencode/pull/54899))
- Refresh the question composer and clarify timeline notices that apply only to a model variant. ([#54822](https://github.com/anomalyco/opencode/pull/54822), [#54880](https://github.com/anomalyco/opencode/pull/54880))
- Refine diff colors and move unread indicators into session tabs. ([#52551](https://github.com/anomalyco/opencode/pull/52551), [#53570](https://github.com/anomalyco/opencode/pull/53570))
#### Bug fixes
- Show steered messages as pending immediately and render inline attachments before their server echo arrives. ([#55251](https://github.com/anomalyco/opencode/pull/55251), [#54650](https://github.com/anomalyco/opencode/pull/54650))
- Preserve raw media data up to 256 MiB so artifact previews remain available. ([#55018](https://github.com/anomalyco/opencode/pull/55018))
- Retain session model selections when the model catalog refreshes. ([#54481](https://github.com/anomalyco/opencode/pull/54481))
- Close browser tabs from the address field and suppress incorrect missing-file notices for restored or already-open tabs. ([#54501](https://github.com/anomalyco/opencode/pull/54501), [#55022](https://github.com/anomalyco/opencode/pull/55022))
- Prevent a project-store eviction loop during launch and remove hover gaps between vertical tabs. ([#55078](https://github.com/anomalyco/opencode/pull/55078), [#54866](https://github.com/anomalyco/opencode/pull/54866))
### SDK
#### Improvements
- Export `lastAssistantWithUsage` from the Solid client to consistently locate the assistant step that determines current context usage. ([#53785](https://github.com/anomalyco/opencode/pull/53785))
#### Bug fixes
- Include all paginated child sessions when loading a session family for usage totals. ([#53785](https://github.com/anomalyco/opencode/pull/53785))
- Admit inline data-URL files into optimistic pending messages so attachments render immediately. ([#54650](https://github.com/anomalyco/opencode/pull/54650))
### Extensions
#### Improvements
- Add `ctx.permission.assert` for Effect and Promise plugins, allowing plugins to use configured permission rules and the native approval prompt. ([#54930](https://github.com/anomalyco/opencode/pull/54930))
#### Bug fixes
- Use Figma's public endpoint and default OAuth configuration, and support Git installations and development releases. ([#55252](https://github.com/anomalyco/opencode/pull/55252), [#55304](https://github.com/anomalyco/opencode/pull/55304))
- Resolve the Figma plugin peer dependency from its workspace. ([#55269](https://github.com/anomalyco/opencode/pull/55269))
- Associate each rejected Promise hook with its own tool call. Thanks [@1052326311](https://github.com/1052326311). ([#55263](https://github.com/anomalyco/opencode/pull/55263))
## v2.0.28 — 2026-10-13
### Highlights
- **Per-Project Auto-Accept**: Set **Settings → Projects → Auto-accept permissions** to override the default for one project. You can also press `mod+shift+a` in a session. ([#54839](https://github.com/anomalyco/opencode/pull/54839))
- **Release Highlights in the TUI**: After updating, the home screen shows the release highlights once. Run `/highlights` to reopen them, or disable automatic display with `highlights: false` in `cli.json`. ([#54969](https://github.com/anomalyco/opencode/pull/54969))
- **Native Windows ARM64 Support**: Windows ARM64 downloads, CLI installs, and desktop updates now use native ARM64 builds instead of x64 emulation. ([#54689](https://github.com/anomalyco/opencode/pull/54689), [#54971](https://github.com/anomalyco/opencode/pull/54971))
- **Continuous Gemini Responses**: Long Gemini and Vertex Gemini responses split by the provider now continue automatically as one response, including recovery from a dropped stream after a continuation checkpoint. ([#55056](https://github.com/anomalyco/opencode/pull/55056))
- **Official WinGet Packages**: Install V2 with `winget install --id Anomaly.OpenCode -e` or `winget install --id Anomaly.OpenCode.Desktop -e`. OpenCode upgrades WinGet installations through WinGet. ([#54808](https://github.com/anomalyco/opencode/pull/54808))
### Core
#### Improvements
- Continued split Gemini and Vertex Gemini responses automatically using provider continuation tokens. ([#55056](https://github.com/anomalyco/opencode/pull/55056))
- Added `settings.encoding` with `zstd`, `gzip`, and `none` values for provider, model, and variant HTTP request bodies. ChatGPT authentication defaults to Zstandard compression when no custom `baseURL` is set. Thanks [@esafak](https://github.com/esafak). ([#54932](https://github.com/anomalyco/opencode/pull/54932))
- Added provider-level compatibility defaults and `compatibility.cacheFormat` for Anthropic- and Alibaba-style prompt-cache markers. ([#54911](https://github.com/anomalyco/opencode/pull/54911), [#54945](https://github.com/anomalyco/opencode/pull/54945))
- Added official WinGet packages and WinGet-aware upgrade and uninstall support. ([#54808](https://github.com/anomalyco/opencode/pull/54808))
- Raised the default streamed-chunk timeout from 5 to 15 minutes for models that pause while reasoning. ([#55032](https://github.com/anomalyco/opencode/pull/55032))
#### Bug fixes
- Disabled Bun’s fetch idle timeout for provider requests, preventing long-running streams from ending early. ([#55007](https://github.com/anomalyco/opencode/pull/55007))
- Preserved mixed CRLF and LF line endings when the edit tool changes files. ([#53166](https://github.com/anomalyco/opencode/pull/53166))
- Preserved unknown V1 OpenAI-compatible model options in migrated request bodies. ([#54694](https://github.com/anomalyco/opencode/pull/54694))
- Stopped querying Azure resources during startup; Azure connections now ask for the resource name directly. ([#54870](https://github.com/anomalyco/opencode/pull/54870))
- Loaded the system configuration directory at the lowest precedence. Thanks [@rntdrts](https://github.com/rntdrts). ([#54921](https://github.com/anomalyco/opencode/pull/54921))
- Installed the native CLI and extracted ripgrep without PowerShell on Windows ARM64. ([#54955](https://github.com/anomalyco/opencode/pull/54955), [#54981](https://github.com/anomalyco/opencode/pull/54981))
### TUI
#### Improvements
- Added the post-update highlights dialog, `/highlights`, and the `highlights` alert setting. ([#54969](https://github.com/anomalyco/opencode/pull/54969))
#### Bug fixes
- Restricted slash-command autocomplete to the start of the prompt. Thanks [@OpeOginni](https://github.com/OpeOginni). ([#54857](https://github.com/anomalyco/opencode/pull/54857))
- Reported x64 emulation correctly when the TUI runs on Windows ARM64. ([#54984](https://github.com/anomalyco/opencode/pull/54984))
- Removed the stale `todowrite` label from tool output. ([#53966](https://github.com/anomalyco/opencode/pull/53966))
### Desktop
#### Improvements
- Added per-project auto-accept permission overrides under **Settings → Projects**, with `mod+shift+a` controlling the current project. ([#54839](https://github.com/anomalyco/opencode/pull/54839))
- Added a native **Windows (ARM64)** option to the download page. ([#54971](https://github.com/anomalyco/opencode/pull/54971))
#### Bug fixes
- Bundled native ARM64 terminal and CLI binaries in Windows ARM64 desktop builds, and moved emulated x64 installations to native ARM64 updates. Thanks [@mweinbach](https://github.com/mweinbach). ([#54689](https://github.com/anomalyco/opencode/pull/54689), [#54983](https://github.com/anomalyco/opencode/pull/54983))
- Included subagent sessions in displayed session costs. ([#55052](https://github.com/anomalyco/opencode/pull/55052))
- Loaded image and audio attachments in the review panel and opened video previews in an in-window theatre view. ([#55015](https://github.com/anomalyco/opencode/pull/55015), [#54696](https://github.com/anomalyco/opencode/pull/54696))
- Showed submitted prompts immediately and kept the **Working** marker above undelivered steers. ([#54047](https://github.com/anomalyco/opencode/pull/54047), [#55023](https://github.com/anomalyco/opencode/pull/55023))
- Fixed tab closing when collapsed side-panel tabs exist and restored focus correctly after stacked dialogs close. ([#54695](https://github.com/anomalyco/opencode/pull/54695), [#55038](https://github.com/anomalyco/opencode/pull/55038))
### SDK
#### Improvements
- Added request-body encoding and model/provider cache-format fields to generated client types. ([#54932](https://github.com/anomalyco/opencode/pull/54932), [#54945](https://github.com/anomalyco/opencode/pull/54945))
- Exported `CredentialApi` from the Promise client API. ([#52797](https://github.com/anomalyco/opencode/pull/52797))
### Extensions
#### Improvements
- Added an optional official Figma plugin with its MCP connection and six bundled Figma skills. ([#55049](https://github.com/anomalyco/opencode/pull/55049), [#55061](https://github.com/anomalyco/opencode/pull/55061))
- Exposed `credential.activate` to Effect and Promise server plugins. Thanks [@remorses](https://github.com/remorses). ([#52797](https://github.com/anomalyco/opencode/pull/52797))
#### Bug fixes
- Converted rejected Promise-based plugin tools into normal tool failures instead of crashing their execution path. ([#54913](https://github.com/anomalyco/opencode/pull/54913))
- Fixed the VS Code extension launching the V2 CLI with the removed `--port` argument. Thanks [@uyencss](https://github.com/uyencss). ([#49084](https://github.com/anomalyco/opencode/pull/49084))
## v2.0.27 — 2026-10-12
### Highlights
- **Built-In Worktree Tools**: Ask OpenCode to use `worktree_create`, `worktree_list`, and `worktree_remove` in Code Mode. You can create isolated worktrees, move sessions into them, and remove them with normal edit approval. ([#53981](https://github.com/anomalyco/opencode/pull/53981))
- **Safer Remote Pairing**: The first `opencode pair --remote` now warns that it will expose your server and asks for confirmation. For non-interactive use, pass `opencode pair --remote --yes`. ([#54561](https://github.com/anomalyco/opencode/pull/54561))
- **Installable Web App**: Use Chrome or Edge’s **Install** button to install the OpenCode web app, including when it runs behind an authenticating proxy. ([#53905](https://github.com/anomalyco/opencode/pull/53905))
- **Clear Session Failures**: Desktop and TUI timelines now show execution errors that happen before or between model steps instead of ending without an explanation. ([#53826](https://github.com/anomalyco/opencode/pull/53826))
- **Automatic Project Cleanup**: Projects whose directories no longer exist are automatically hidden from the project list. Opening the directory again restores the project. ([#54354](https://github.com/anomalyco/opencode/pull/54354))
### Core
#### Improvements
- Added the built-in Code Mode tools `opencode.worktree_create`, `opencode.worktree_list`, and `opencode.worktree_remove`. Removing a worktree requires edit approval, and `force: true` can discard modified or untracked files. ([#53981](https://github.com/anomalyco/opencode/pull/53981))
- Added confirmation before `opencode pair --remote` first enables internet access, plus `--yes`/`-y` for scripts and other non-interactive use. ([#54561](https://github.com/anomalyco/opencode/pull/54561))
- Automatically archived projects when all known directories disappear. Reopening a project restores it, while surviving clones and linked worktrees keep it active. ([#54354](https://github.com/anomalyco/opencode/pull/54354))
- Raised the ChatGPT OAuth input limit for `gpt-6.1-sol` and `gpt-6.1-sol-fast` to 500,000 tokens. ([#54668](https://github.com/anomalyco/opencode/pull/54668))
- Added Vertex AI routes for Interactions and Mistral models. ([#54175](https://github.com/anomalyco/opencode/pull/54175), [#54058](https://github.com/anomalyco/opencode/pull/54058))
#### Bug fixes
- Fixed updater detection for symlinked home directories and package-manager installations, increased probe timeouts, and now report unavailable updates clearly when the installation method cannot be detected. ([#54609](https://github.com/anomalyco/opencode/pull/54609), [#54622](https://github.com/anomalyco/opencode/pull/54622), [#54672](https://github.com/anomalyco/opencode/pull/54672))
- Fixed GitHub Copilot routing for Gemini, Anthropic, and Responses models when live model metadata is unavailable, and accepted interleaved Claude reasoning values. ([#54208](https://github.com/anomalyco/opencode/pull/54208), [#54210](https://github.com/anomalyco/opencode/pull/54210), [#54201](https://github.com/anomalyco/opencode/pull/54201))
- Fixed Google Vertex Application Default Credentials when `CLOUDSDK_CONFIG` points to a custom gcloud configuration directory. Thanks [@H-Max](https://github.com/H-Max). ([#54164](https://github.com/anomalyco/opencode/pull/54164))
- Blocked `file://` access through Code Mode’s `fetch`, which now returns a clear filesystem-access error. Thanks [@ferdinandyb](https://github.com/ferdinandyb). ([#54510](https://github.com/anomalyco/opencode/pull/54510))
- Fixed remote workspace file operations so driver-provided filesystem overrides are used correctly.
- Preserved subagent notifications when location services are evicted, and propagated well-known configuration updates across locations with bounded HTTP timeouts. ([#54072](https://github.com/anomalyco/opencode/pull/54072), [#53821](https://github.com/anomalyco/opencode/pull/53821))
### TUI
#### Bug fixes
- Displayed session execution failures that occur before or between model steps directly in the timeline. ([#53826](https://github.com/anomalyco/opencode/pull/53826))
- Preserved a saved session’s unavailable model selection instead of silently replacing it with a fallback. ([#53822](https://github.com/anomalyco/opencode/pull/53822))
- Added syntax highlighting for C++ module interface files such as `.cppm`, `.ixx`, and `.cxxm`. ([#53853](https://github.com/anomalyco/opencode/pull/53853))
- Summarized integrations with multiple credentials as a connection count, preventing long labels from crowding the integration name. ([#54054](https://github.com/anomalyco/opencode/pull/54054))
### Desktop
#### Improvements
- Made the web app installable from Chrome and Edge, including deployments behind cookie-based authenticating proxies. ([#53905](https://github.com/anomalyco/opencode/pull/53905))
- Updated the session column layout and text wrapping for clearer long content. ([#54020](https://github.com/anomalyco/opencode/pull/54020))
#### Bug fixes
- Displayed session execution failures that occur before or between model steps directly in the timeline. ([#53826](https://github.com/anomalyco/opencode/pull/53826))
- Fixed command-palette selection so newly arriving higher-ranked results become active until you explicitly choose a result. ([#54275](https://github.com/anomalyco/opencode/pull/54275))
- Preserved bold, code, emphasis, and strike-through formatting inside Markdown link labels. ([#54451](https://github.com/anomalyco/opencode/pull/54451))
- Fixed timeline file links and complete patch rendering, including unusual filenames and patches whose original alignment can be reused. ([#53641](https://github.com/anomalyco/opencode/pull/53641), [#54017](https://github.com/anomalyco/opencode/pull/54017))
- Showed complete tool error text when an error is expanded. ([#53816](https://github.com/anomalyco/opencode/pull/53816))
- Made the third-party notices dialog scrollable. ([#53911](https://github.com/anomalyco/opencode/pull/53911))
### SDK
#### Improvements
- Added startup plugin seeding to Effect embedded hosts and workerd through `OpenCode.create({ plugins })` and `OpenCodeWorkerd.create({ plugins })`.
- Re-exported `WorkspaceDriver`, `EnvironmentDriver`, and `EnvironmentFiles` from `@opencode/sdk/effect`.
#### Bug fixes
- Ensured embedded hosts activate supplied plugins before model discovery, generation, and recovery begin. ([#54219](https://github.com/anomalyco/opencode/pull/54219))
### Extensions
#### Improvements
- Extended `BridgeLayout` with viewport dimensions and optional border information so native extension surfaces can track window resizing and preserve rounded card borders.
#### Bug fixes
- Kept desktop browser-panel corner masks and borders aligned while resizing or using custom themes. ([#52461](https://github.com/anomalyco/opencode/pull/52461))
- Kept the browser site button clickable after the embedded page takes focus. ([#53907](https://github.com/anomalyco/opencode/pull/53907))
## v2.0.26 — 2026-10-08
### Highlights
- **Google Cloud Credentials for Vertex**: Run `/connect`, choose **Vertex**, then select **Google Cloud credentials (gcloud auth or environment)** to use Application Default Credentials. You can choose or enter the Google Cloud project and location. ([#53798](https://github.com/anomalyco/opencode/pull/53798))
- **Faster, Safer Session Snapshots**: Session snapshots now add less overhead, recover from interrupted or concurrent captures, restore large selections faster, and pack stored objects to reduce disk usage. ([#51996](https://github.com/anomalyco/opencode/pull/51996))
- **Working Transcript Links**: You can click links anywhere in a transcript, including any row of a wrapped URL. Hyperlinks also work in more terminals, and `FORCE_HYPERLINK=1` can enable them when automatic detection does not. ([#53903](https://github.com/anomalyco/opencode/pull/53903), [#53904](https://github.com/anomalyco/opencode/pull/53904))
- **Step 5 Preview Free**: You can now select **Step 5 Preview Free** from OpenCode’s free model lineup for a limited time.
### Core
#### Improvements
- Added **Google Cloud credentials (gcloud auth or environment)** to the Vertex connection flow, with project and location choices discovered from environment variables, gcloud configurations, and Application Default Credentials. ([#53798](https://github.com/anomalyco/opencode/pull/53798))
- Reduced session snapshot overhead, accelerated multi-file restores, made capture resilient to crashes and concurrent processes, and added lossless background packing of snapshot objects. ([#51996](https://github.com/anomalyco/opencode/pull/51996))
- Added **Step 5 Preview Free** to OpenCode’s free models for a limited time and removed the retired **Fledge Alpha Free** listing.
#### Bug fixes
- Prevented working-tree diffs from timing out or stalling the server in repositories with many untracked files by batching Git operations. ([#53449](https://github.com/anomalyco/opencode/pull/53449))
- Fixed large revert selections failing when their paths exceeded the operating system’s command-line limit. ([#53956](https://github.com/anomalyco/opencode/pull/53956))
- Fixed snapshot and undo handling for literal filenames containing Git pathspec characters, corrupted or locked snapshot indexes, concurrent captures, and project directories whose names begin with `..`. ([#51996](https://github.com/anomalyco/opencode/pull/51996))
- Changed `opencode service` remote-access URLs to use a persistent random subdomain rather than a predictable route. Disabling and re-enabling remote access generates a new address. ([#53962](https://github.com/anomalyco/opencode/pull/53962))
### TUI
#### Improvements
- Enabled OSC 8 hyperlinks in more terminals, including VS Code and other xterm.js terminals, Zed, mintty, and newer VTE terminals; `FORCE_HYPERLINK=1` now overrides terminal detection. ([#53904](https://github.com/anomalyco/opencode/pull/53904))
#### Bug fixes
- Fixed transcript links not opening when clicked, including URLs wrapped across multiple rows. ([#53903](https://github.com/anomalyco/opencode/pull/53903))
- Kept activity summaries in place when expanding or collapsing details with `session.verbosity: "low"`. ([#52357](https://github.com/anomalyco/opencode/pull/52357))
- Prevented the home-screen version footer from overlapping keyboard hints in short terminals. Thanks [@fihaaade](https://github.com/fihaaade). ([#53891](https://github.com/anomalyco/opencode/pull/53891))
@@ -81,7 +81,7 @@ Bug fixes, chores, and tests must reference an existing issue. Documentation, re
Before implementing new functionality, open a feature request describing the problem, why it belongs in OpenCode, and your proposed approach if you have one. Wait for design approval before opening the implementation PR.
Base branches on `v2`, not `dev`, and complete the provided pull request template.
Base branches on `dev` and complete the provided pull request template.
-Prefer `HttpClient.HttpClient` / `HttpClientResponse.HttpClientResponse` over web `fetch` / `Response` at package boundaries.
-Use `Stream.Stream` for streaming data flow. Avoid ad hoc async generators or manual web reader loops unless an Effect `Stream` API cannot model the behavior.
-Use Effect Schema codecs for JSON encode/decode (`Schema.fromJsonString(...)`) instead of direct `JSON.parse` / `JSON.stringify` in implementation code.
-In `Effect.gen`, yield yieldable errors directly (`return yield* new MyError(...)`) instead of `Effect.fail(new MyError(...))`.
- Use `Effect.void` instead of `Effect.succeed(undefined)` when the successful value is intentionally void.
-**Forward compatibility.** Providers ship new models, options, and values constantly. Accept unknown pass-through values (`Known | (string & {})` for string options); use closed literals only when OpenCode branches on the value.
-**One canonical model.** The Schema classes in `src/schema/` are the data model. Do not adda parallel shape or a second construction path.
-**Independent of Session.** Auth, permissions, plugins, and orchestration belong in Core. `LLM.stream`/`LLM.generate` run exactly one model call.
-**Reuse protocols.** Providers compose existing protocols with `Route.make(...)` instead of cloning them.
## Conventions
## Code style
Per-type constructors live on the type, not as top-level re-exports. Use `Message.system(...)`, `Message.user(...)`, `Message.assistant(...)`, `Message.tool(...)`, `LanguageModel.make(...)`, `ToolDefinition.make(...)`, `ToolCallPart.make(...)`, `ToolResultPart.make(...)`, `ToolChoice.make(...)`, `ToolChoice.named(...)`, `SystemPart.make(...)`, and `GenerationOptions.make(...)` directly. The top-level `LLM` namespace is reserved for request-shaped call APIs: `LLM.request`, `LLM.generate`, `LLM.stream`, and `LLM.generateObject`. Use `LLMRequest.update(...)` when deriving canonical request data; do not add a duplicate `LLM.updateRequest(...)` path. Two ways to construct the same thing is one too many.
-Keep provider-defined string enums forward-compatible. Expose known values for autocomplete while accepting future values with `Known | (string & {})`; use `Schema.String` at runtime unless rejecting unknown values is required for correctness.
- Order reasoning-effort values from lowest to highest: `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max`. Provider-specific subsets follow the same relative order in types, schemas, option lists, and tests.
- Use `HttpClient`, `Stream`, and Schema codecs instead of `fetch`, async generators, and `JSON.parse`/`JSON.stringify`.
-`return yield* new MyError(...)` instead of `Effect.fail(...)`; `Effect.void` instead of `Effect.succeed(undefined)`.
-Only `src/promise.ts` may use Promises.
## Tests
- Use `testEffect(...)` from `test/lib/effect.ts` for tests requiring Effect layers.
-Keep provider tests fixture-first. Live provider calls must stay behind`RECORD=true` and required API-key checks.
## Errors
-`AIError` wraps a union of tagged reason errors. It stores only `reason`, derives `message` from the reason, and exposes the reason as its `cause`.
- Each reason owns its readable `message`, category-specific fields, and optional `body`, `http`, and underlying exception in `cause`.
-`reason.body` is the sole original-response or triggering-event payload field. Preserve original text before schema decoding removes fields; do not replace the complete event with only its nested error.
-`reason.http` describes an observed HTTP response with required `url`, `status`, and response `headers`. Do not invent status codes or derive a separate request ID from headers.
- Reclassification and transport recovery must preserve the reason's body, HTTP context, and underlying cause. Error `message` and `cause` are non-enumerable: copy them explicitly when constructing an enriched reason with its constructor or `AIErrorReason.make`.
## Architecture
This package is an Effect Schema-first LLM core. The Schema classes in `src/schema/` are the canonical runtime data model. Convenience functions in `src/llm.ts` are thin constructors that return those same Schema class instances; they should improve callsites without creating a second model.
Session integration lives in `packages/core/src/session`: `runner/llm.ts` owns orchestration, `model-request.ts` lowers Session state into `LLMRequest`, and `model-transport.ts` selects transport behavior.
Keep this package independent of Session concerns. Session auth, permissions, plugins, telemetry headers, and runtime selection belong in Core.
`LLM.request(...)` builds an `LLMRequest`. `LLMClient.generate(...)` reads the executable route carried by `request.model.route`, builds the provider-native body, asks the route's transport for a real `HttpClientRequest.HttpClientRequest`, sends it through `RequestExecutor.Service`, parses the provider stream into common `LLMEvent`s, and finally returns an `LLMResponse`.
Use `LLMClient.stream(request)` when callers want incremental `LLMEvent`s. Use `LLMClient.generate(request)` when callers want those same events collected into an `LLMResponse`.
Filter or narrow `LLMEvent` streams with `LLMEvent.is.*` (camelCase guards, e.g. `events.filter(LLMEvent.is.toolCall)`). The kebab-case `LLMEvent.guards["tool-call"]` form also works but prefer `is.*` in new code.
### Routes
A route is the runnable composition of four orthogonal pieces:
- **`Protocol`** (`src/route/protocol.ts`) — semantic API contract. Owns request body construction (`body.from`), the body schema (`body.schema`), the streaming-event schema (`stream.event`), and the event-to-`LLMEvent` state machine (`stream.step`). `Route.make(...)` validates and JSON-encodes the body from `body.schema` and decodes frames with `stream.event`. Examples: `OpenAIChat.protocol`, `OpenResponses.protocol`, `OpenAIResponses.protocol`, `AnthropicMessages.protocol`, `Gemini.protocol`, `BedrockConverse.protocol`.
- **`Endpoint`** (`src/route/endpoint.ts`) — URL construction. The host, path, and route query live on the endpoint. `Endpoint.path("/chat/completions", { baseURL })` is the common case; pass a function for paths that embed the model id or a body field (e.g. `Endpoint.path(({ body }) => `/model/${body.modelId}/converse-stream`)`).
- **`Auth`** (`src/route/auth.ts`) — per-request transport authentication. Provider facades configure credentials onto the route before model selection, usually via `Auth.bearer(apiKey)` or `Auth.header(name, apiKey)`. Routes that need per-request signing (Bedrock SigV4, future Vertex IAM, Azure AAD) implement `Auth` as a function that signs the body and merges signed headers into the result.
- **`Framing`** (`src/route/framing.ts`) — bytes → frames. SSE (`Framing.sse`) is shared; Bedrock keeps its AWS event-stream framing as a typed `Framing<object>` value alongside its protocol.
Compose them via `Route.make(...)`:
```ts
exportconstroute=Route.make({
id:"openai-chat",
provider:"openai",
protocol: OpenAIChat.protocol,
endpoint: Endpoint.path("/chat/completions",{
baseURL:"https://api.openai.com/v1",
}),
auth: Auth.bearer(Auth.config("OPENAI_API_KEY")),
framing: Framing.sse,
})
```
Route defaults are request-shaping defaults such as `headers`, `limits`, `generation`, `providerOptions`, and `http`. Endpoint host/query belongs on the route endpoint. Selected `LanguageModel` values carry only model id, provider id, and the configured route value. Model capability/catalog metadata lives outside this package; protocol support is enforced by request lowering and typed `AIError`s.
The four-axis decomposition is the reason DeepSeek, TogetherAI, Cerebras, Baseten, Fireworks, and DeepInfra all reuse `OpenAIChat.protocol` verbatim — each provider owns a small `Route.make(...)` composition instead of a protocol clone. Bug fixes in one protocol propagate to every consumer of that protocol in a single commit.
When a provider supports multiple physical transports, selection remains execution policy below its semantic route. `OpenResponsesChannel.transport(...)` owns the provider-neutral Responses WebSocket concept: it prepares one final request, executes HTTP by default, strips WebSocket-disallowed fields, and passes a generic channel exchange to a per-call `WebSocketChannelExecutor` when supplied. Provider-specific Responses routes opt in with handshake and connection-age policy. `Route.streamPrepared` owns decoding and acknowledges channel completion only after successful full consumption.
### URL Construction
`Endpoint` owns `{ baseURL, path, query }`. Each protocol route includes a canonical endpoint when the provider has one (e.g. `https://api.openai.com/v1`); provider helpers override endpoint fields by configuring the route before selecting a model. Generic OpenAI-compatible routes have no canonical URL and require configuration before execution.
For providers where the URL is derived from typed inputs (Azure resource name, Bedrock region), the provider helper configures the route endpoint before calling `.model(...)`. Use `AtLeastOne<T>` from `route/auth-options.ts` for inputs that accept either of two derivation paths (Azure: `resourceName` or `baseURL`).
### Provider Facades
Provider-facing APIs are configured facades over route values. Endpoint/auth/resource/API-version setup happens before model selection, and model selectors accept only a model or deployment id:
- Use branded `ProviderID.make(...)` and `ModelID.make(...)` where ids are constructed directly.
- Use `model` for the default API path and named methods for provider-native alternatives such as OpenAI `responses` and `chat`.
- Put provider-specific setup on `.configure(...)`; do not add `model(id, overrides)` as a duplicate construction path.
- Export lower-level `routes` arrays separately only when advanced internal wiring needs them.
- Prefer `apiKey` as provider-specific sugar and `auth` as the explicit override; keep them mutually exclusive in provider option types with `ProviderAuthOption`.
- Resolve `apiKey` → `Auth` with `AuthOptions.bearer(options, "<PROVIDER>_API_KEY")` (it honors an explicit `auth` override and falls back to `Auth.config(envVar)` so missing keys surface a typed `Authentication` error rather than a runtime crash).
- Use separate top-level facades for products with different required setup, such as `CloudflareAIGateway` and `CloudflareWorkersAI`.
- Give every named provider its own file and top-level export. Keep its endpoint, auth defaults, and route setup in that file. Compose shared protocols directly; do not nest named provider presets under generic compatible facades or keep their endpoints in a shared provider profile registry.
`Provider.make(...)` remains available for simple static provider definitions, but new built-in providers should prefer plain configured facades unless a helper removes real duplication without adding runtime behavior.
### Provider Package Entrypoints
Catalog-selected native providers use package-like export paths from `@opencode/ai`. They are internal entrypoints in one npm package, not separately published provider packages. Every entrypoint implements `ProviderPackage.Definition` and exposes `model(modelID, settings)`, where settings are one flat serializable object: the connection keys the entrypoint declares (`apiKey`, `baseURL`, `region`, …), the common `headers` and `body` overlays, and the protocol's request options (`reasoningEffort`, `thinking`, …) side by side. Each entrypoint destructures its own connection keys and passes the rest to the route as `providerOptions`; there is no nested `providerOptions` at the entrypoint.
Keep semantic APIs as separate entrypoints, such as OpenAI `chat` and `responses`. Transport is execution policy: OpenAI Responses uses HTTP by default and may receive a per-call WebSocket channel executor through `StreamOptions` without changing model or route identity.
Do not expose `Route` in provider package settings. Route composition stays an implementation detail behind `model(...)`.
The dependency arrow points down: `providers/*.ts` files import protocol routes and auth-option utilities; protocol modules import `endpoint`, `auth`, `framing`, and transport pieces. Protocols do not import provider facades. Lower-level modules know nothing about provider catalog metadata. `OpenAIResponses` composes the provider-neutral `OpenResponses` protocol; the baseline never imports the OpenAI extension.
### Shared protocol helpers
`ProviderShared` exports a small toolkit used inside protocol implementations to keep them focused on provider-native shapes:
-`joinText(parts)` — joins an array of `TextPart` (or anything with a `.text`) with newlines. Use this anywhere a protocol flattens text content into a single string for a provider field.
-`parseToolInput(route, name, raw)` — Schema-decodes a tool-call argument string with the canonical "Invalid JSON input for `<route>` tool call `<name>`" error message. Treats empty input as `{}`.
-`parseJson(route, raw, message)` — generic JSON-via-Schema decode for non-tool bodies.
-`validateWith(decoder)` — maps Schema decode errors to `InvalidRequest`. `Route.make(...)` uses this for body validation; lower-level routes can reuse it.
-`matchToolChoice(provider, choice, branches)` — branches over `LLMRequest["toolChoice"]` for provider-specific lowering.
If you find yourself copying a 3-to-5-line snippet between two protocols, lift it into `ProviderShared` next to these helpers rather than duplicating.
### Chronological System Updates
`LLMRequest.system` is the initial privileged prompt that applies ahead of the conversation. `Message.system(...)` is a separate, provider-neutral chronological operator update inside `LLMRequest.messages`; it applies only from its position in history onward and accepts text content only.
Native chronological system messages are route/model-specific. Open Responses lowers them to standard `developer` messages, while Anthropic Messages lowers them to native system messages for Claude Opus 4.8 (`claude-opus-4-8`). Other routes and models intentionally lower the update in place into ordinary user-compatible text using this stable escaped representation:
```text
<system-update>
...
</system-update>
```
The wrapped-user fallback preserves ordering while visibly lowering authority. Never silently pass a raw chronological `role: "system"` through a route that might reject it. Do not insert raw retrieved documents, tool output, or web content into privileged chronological system updates; keep untrusted content in ordinary user/tool channels.
### Effort Updates
`Message.effort({ effort, previous })` is a chronological "reasoning effort changed here" marker (`undefined` means the model default). Changing a top-level effort invalidates the whole provider prompt cache, so protocols with a native per-message update (`Protocol.supportsEffortUpdates`) keep the top-level effort at the first marker's `previous` and lower each marker in place: Anthropic Messages emits an empty `role: "system"` message with `output_config.effort` plus the `mid-conversation-output-config-2026-07-01` beta, and OpenAI Responses emits `configuration_update` items. `applyEffortUpdates` runs in `prepareRequest` and strips the markers for every other route, so a protocol without support keeps today's plain top-level behaviour. When the last marker disagrees with the effort the request asks for (reverted or forked history), `resolveEffortUpdates` strips the markers and falls back to a plain top-level change.
### Tools
Tool loops are represented in common messages and events:
Routes lower these into provider-native assistant tool-call messages and tool-result messages. Streaming providers should emit `tool-input-delta` events while arguments arrive, then a final `tool-call` event with parsed input.
### Tool dispatch
`LLM.stream(request)` and `LLM.generate(request)` each run exactly one model call. Add tool schemas to `request.tools` with `Tool.toDefinitions(tools)`. When a caller wants the package's typed one-call execution behavior, pass each canonical local `tool-call` event to `ToolRuntime.dispatch(tools, call)`.
// Persist call + dispatched.result, then construct the next request explicitly.
}
```
The dispatcher:
- On `tool-call`: looks up the named tool, decodes input against `parameters` Schema, dispatches to the typed `execute`, encodes the result against `success` Schema, and returns canonical `tool-result` events.
- Does not stream providers, construct Session events, schedule fibers, append history, count steps, or continue model rounds.
- Leaves persistence and continuation to the enclosing product flow.
Handler dependencies (services, permissions, plugin hooks, abort handling) are closed over by the consumer at tool-construction time. Build the tools record inside an `Effect.gen` once and reuse it across many dispatches.
Errors must be expressed as `ToolFailure`. The runtime catches it and emits a `tool-error` event, then a `tool-result` of `type: "error"`, so the model can self-correct on the next step. Anything that is not a `ToolFailure` is treated as a defect and fails the stream. Three recoverable error paths produce `tool-error` events:
- Routes surface the model's call as a `tool-call` event with `providerExecuted: true`, and the provider's result as a matching `tool-result` event with `providerExecuted: true`.
- Callers detect `providerExecuted` on `tool-call` and **skip local dispatch** — no handler is invoked and no `tool-error` is raised for "unknown tool". The provider already executed it.
- Callers that continue should retain both events in explicit history when the protocol requires it. Anthropic encodes them back as `server_tool_use` + `web_search_tool_result` (or `code_execution_tool_result` / `web_fetch_tool_result`) blocks; OpenAI Responses callers typically use `previous_response_id` instead of resending hosted-tool items.
Add provider-defined tools to `request.tools` (no runtime entry needed). The matching route must know how to lower the tool definition into the provider-native shape; right now Anthropic accepts `web_search` / `code_execution` / `web_fetch` and OpenAI Responses accepts the hosted tool names listed above.
## Protocol File Style
Protocol files should look self-similar. Provider quirks belong behind named helpers so a new route can be reviewed by comparing the same sections across files.
### Section order
Use this order for every protocol module:
1. Public model input
2. Request body schema
3. Streaming event schema
4. Parser state
5. Request body construction (`fromRequest`)
6. Stream parsing (`step` and per-event handlers)
7. Protocol and route
8. Protocol route export
### Rules
- Keep protocol files focused on the protocol. Move provider-specific projection, signing, media normalization, or other bulky transformations into `src/protocols/utils/*`.
- Use `Effect.fn("Provider.fromRequest")` for request body construction entrypoints. Use `Effect.fn(...)` for event handlers that yield effects; keep purely synchronous handlers as plain functions returning a `StepResult` that the dispatcher lifts via `Effect.succeed(...)`.
- Parser state owns terminal information. The state machine records finish reason, usage, and pending tool calls; emit one terminal `finish` event (or `provider-error`) for each completed response. If a provider splits reason and usage across events, merge them in parser state before flushing.
- Emit exactly one terminal `finish` event for a completed response, normally after a matching `step-finish`. Use `stream.terminal` to stop reading when the provider has a completion sentinel; use `stream.onHalt` when the final event must be flushed after the framed stream ends.
- Use shared helpers for repeated protocol policy such as text joining, usage totals, JSON parsing, and tool-call accumulation. `ToolStream` (`protocols/utils/tool-stream.ts`) accumulates streamed tool-call arguments uniformly.
- Make intentional provider differences explicit in helper names or comments. If two protocol files differ visually, the reason should be obvious from the names.
- Prefer dispatched per-event handlers (`onMessageStart`, `onContentBlockDelta`, ...) called from a small top-level `step` switch over a long if-chain. The dispatcher keeps the event surface visible at a glance.
- Keep tests in the same conceptual order as the protocol: basic prepare, tools prepare, unsupported lowering, text/usage parsing, tool streaming, finish reasons, provider errors.
### Review checklist
- Can the file be skimmed side-by-side with `openai-chat.ts` without hunting for equivalent sections?
- Are provider quirks named, isolated, and covered by focused tests?
- Does request body construction validate unsupported common content at the protocol boundary?
- Does stream parsing emit stable common events without leaking provider event order to callers?
- Does `toolChoice: "none"` behavior read as intentional?
## Recording Tests
Recorded tests use one cassette file per scenario. A cassette holds an ordered array of `{ request, response }` interactions, so multi-step flows (tool loops, retries, polling) record into a single file. Use `recordedTests({ prefix, requires })` and let the helper derive cassette names from test names:
Replay is the default. `RECORD=true` records fresh cassettes locally and requires the listed env vars; unset `CI` before recording because CI always forces replay. Cassettes are written as pretty-printed JSON so multi-interaction diffs stay reviewable.
Pass `provider`, `protocol`, and optional `tags` to `recordedTests(...)` / `recorded.effect.with(...)` so cassettes carry searchable metadata. Use recorded-test filters to replay or record a narrow subset without rewriting a whole file:
-`RECORDED_PROVIDER=openai` matches tests tagged with `provider:openai`; comma-separated values are allowed.
-`RECORDED_PREFIX=openai-chat` matches cassette groups by `recordedTests({ prefix })`; comma-separated values are allowed.
-`RECORDED_TAGS=tool` requires all listed tags to be present, e.g. `RECORDED_TAGS=provider:togetherai,tool`.
-`RECORDED_TEST="streams text"` matches by test name, kebab-case test id, or cassette path.
Filters apply in replay and record mode. Combine them with `RECORD=true` when refreshing only one provider or scenario.
**Binary response bodies.** Most providers stream text (SSE, JSON). The recorder treats known textual media types (`text/*`, JSON/XML structured types, JavaScript, forms, YAML, and SVG) as text and stores every other response as base64 with `bodyEncoding: "base64"`. This preserves binary formats such as AWS event-stream frames without a lossy UTF-8 round trip.
**Matching strategy.** A runtime request atomically claims the first unused recorded interaction that matches its method, URL, allow-listed headers, and canonical JSON body. Distinct requests may replay in any order or concurrently. Repeated identical requests consume their matching responses in cassette order, preserving deterministic retry and polling behavior. `scriptedResponses` (in `test/lib/http.ts`) is the deterministic counterpart for tests that don't need a live provider; it scripts response bodies in order without reading from disk.
Do not blanket re-record an entire test file when adding one cassette. `RECORD=true` rewrites every recorded case that runs, and provider streams contain volatile IDs, timestamps, fingerprints, and obfuscation fields. Prefer deleting the one cassette you intend to refresh, or run a focused test pattern that only registers the scenario you want to record. Keep stable existing cassettes unchanged unless their request shape or expected behavior changed.
- Use `testEffect(...)` from `test/lib/effect.ts` for tests that need Effect layers.
-Provider tests replay cassettes: recorded HTTP exchanges stored as JSON in `test/fixtures/recordings/`, one file per scenario. Replay is the default and needs no APIkeys.
- To record, unset `CI` and run with `RECORD=true` plus the provider's API key. Re-record only the cassette you are changing; recordings contain volatile IDs that create noisy diffs.
Run `LLMClient.stream(request)` instead of `generate` when you want incremental `LLMEvent`s. The event stream is provider-neutral — same shape across OpenAI Chat, OpenAI Responses, Anthropic Messages, Gemini, Bedrock Converse, and any OpenAI-compatible deployment.
Run `LLM.stream(...)` instead of `generate` when you want incremental `LLMEvent`s. Both accept input or a prebuilt
`LLM.request(...)`. The event stream is provider-neutral — same shape across OpenAI Chat, OpenAI Responses,
Anthropic Messages, Gemini, Bedrock Converse, and any OpenAI-compatible deployment.
### Google Interactions
`Google.configure({ apiKey }).interactions(modelID)` selects the Interactions API; `.model(modelID)` still selects
GenerateContent. The package entrypoint is `@opencode/ai/providers/google/interactions`.
For Vertex AI, use `GoogleVertexInteractions.configure({ project, location: "global" }).model(modelID)` or
`@opencode/ai/providers/google-vertex/interactions`. It uses Application Default Credentials by default, accepts
an explicit `accessToken` or `auth` override, and supports Express mode with `apiKey`. Vertex metadata and raw usage
live under `providerMetadata.vertex`. The default is still full-history replay with `store: false`.
Conversational image generation remains part of the LLM interaction. OpenAI Responses exposes it through its hosted image tool:
@@ -503,17 +764,268 @@ const program = Effect.gen(function* () {
})
```
The hosted result is represented as a provider-executed tool call and tool result. Its image is a `file` content item with a data URI, so retaining `response.message` preserves the generated image for continuation.
The hosted result is represented as a provider-executed tool call and a tool result whose content carries the generated image as a file. Gemini image-capable models instead emit a first-class `media``LLMEvent` for inline image output (`response.message` then carries a `media` part). Retaining `response.message` preserves the generated image for continuation on both routes.
## Video generation
Video mirrors `Image` with one difference: every provider is asynchronous, so the route is a submit-then-poll
`Generation`. Models come from `.video(...)` selectors on the `Google` (Veo), `XAI`, `Fal`, and `Runway` facades.
- **`LLM.request({...})`** — build a provider-neutral `LLMRequest`. Accepts ergonomic inputs (`system: string`, `prompt: string`) that normalize into the canonical Schema classes.
- **`LLM.generate` / `LLM.stream`** — re-exported from`LLMClient` for one-import use.
- **`LLM.generate` / `LLM.stream`** — run direct input or an `LLMRequest` through`LLMClient` for one-import use.
- **`Message.user(...)` / `Message.assistant(...)` / `Message.tool(...)`** — message constructors from the canonical schema model.
- **`LanguageModel.make(...)` / `ToolCallPart.make(...)` / `ToolResultPart.make(...)` / `ToolDefinition.make(...)`** — model and tool-related constructors from the canonical schema model.
- **`Image.generate({...})`** — generate images through a provider-neutral image request and response model.
- **`Image.request` / `generate` / `stream` / `start` / `resume`** — images over inline, streaming (partial previews), and queued routes through a provider-neutral request and response model.
- **`ImageClient`** — Effect service and layer for image execution, parallel to `LLMClient`.
- **`Media`** — the shared asset type (`Media.Asset`, `Media.Source`) and constructors used by messages, tool results, and media requests.
- **`Generation`** — provider-neutral handle for an in-flight media generation (`await`, `refresh`, `cancel`, `events`) used by queued media routes.
- **`Video.request` / `generate` / `stream` / `start` / `resume`** — queued video generation through a provider-neutral request; `VideoClient` is its Effect service and layer.
- **`Speech.request` / `Speech.generate` / `Speech.stream`** — text-to-speech through a provider-neutral request; `SpeechClient` is its Effect service and layer.
- **`Transcription.request` / `generate` / `stream` / `start` / `resume`** — speech-to-text over inline, streaming, and queued routes; `TranscriptionClient` is its Effect service and layer.
- **`AIClient.layer` / `AIClient.layerWith(executor)`** — every modality client plus the request executor in one layer.
- **`@opencode/ai/promise`** — `AI.make({ layer? })` and a default `ai` client exposing `llm`, `image`, `video`, `speech`, and `transcription` as Promise / `AsyncIterable` APIs, plus `file`, `write`, `bytes`, `base64`, and `materialize` for assets.
## Testing
@@ -576,11 +1088,13 @@ This is different from prompt caching, server-side history storage, or truncatio
Prefer this operation, where supported, when the application owns compaction policy and durable context updates.
```ts
constresult=yield*LLMClient.compact(request)
constnext=LLMRequest.update(request,{
messages: result.replacement,
constcompacted=Effect.gen(function*(){
constresult=yield*LLMClient.compact(request)
constnext=LLMRequest.update(request,{
messages: result.replacement,
})
returnyield*LLMClient.generate(next)
})
constresponse=yield*LLMClient.generate(next)
```
`replacement` replaces the complete input window. Do not append it to the original transcript or extract only the encrypted item: the provider may retain additional messages in its output. Retained user and assistant messages remain ordinary messages with typed text, media, or reasoning parts, in their original order. Provider-specific message IDs, status, and phase use `providerMetadata`, not a raw output array hidden in an assistant message. Unsupported returned item types fail explicitly.
@@ -596,16 +1110,16 @@ The input must still fit the model's context window. Explicit compaction is not
OpenAI Responses also exposes a separate, explicitly selected mechanism:
```ts
constresult=
yield*
LLMClient.compact(request,{
constcheckpoint=Effect.gen(function*(){
constresult=yield*LLMClient.compact(request,{
mechanism:"trigger",
webSocket,// Optional: without it, the request uses HTTP/SSE.
This appends a native `compaction_trigger` control item to the full input and sends a normal Responses request, with tools and instructions retained, `stream: true`, `store: false`, and parallel tool calls enabled. It removes normal-answer text/output-format controls, forced tool choices, output-token/tool-call limits, and automatic `context_management`. Body overlays cannot replace `input` or supply `previous_response_id`/`conversation`; the complete canonical history is required for safe stateless replay. Request metadata, auth, headers, query parameters, service tier, and supported prompt-cache settings are preserved.
@@ -619,9 +1133,11 @@ The supplied WebSocket executor can reuse a compatible append baseline for the c
Trigger support is separate from endpoint support. Only the OpenAI Responses route advertises it; Azure, xAI, Chat, and compatible Responses routes do not inherit it. Untyped calls still fail before sending: missing route capabilities return `UnsupportedOperation`, while unknown mechanism names and invalid inputs return `InvalidRequest`. Dynamic callers must narrow for the selected mechanism:
This capability describes protocol implementation, **not universal availability on OpenAI API deployments**. The host application owns subscription/deployment eligibility, OAuth, endpoint selection, and deployment-specific headers. Local protocol/socket tests do not establish live provider support.
`providerOptions.contextManagement` lets the provider decide when to compact during an ordinary `generate` or `stream` call. This is an advanced option for callers that own persistence and recovery: persist the complete assistant message, including its checkpoint, before continuing. Enabling the option does not provide durable checkpoint storage, interruption recovery, or model-switch policy. Keep the prior context until a successful checkpoint has been persisted.
Inside an `Effect.gen`, enable OpenAI compaction with typed provider options:
Enable OpenAI compaction with typed provider options:
Included LLM providers: OpenAI, Anthropic, Google (Gemini), Google Vertex, Amazon Bedrock, Azure OpenAI, Baseten, Cerebras, Cloudflare AI Gateway, Cloudflare Workers AI, DeepInfra, DeepSeek, Fireworks, Groq, Mistral, OpenRouter, TogetherAI, and xAI. Z.ai currently exposes image generation. Generic Chat Completions, Responses, and Anthropic Messages-compatible entrypoints support custom endpoints.
Included LLM providers: OpenAI, Anthropic, Google (Gemini), Google Vertex, Amazon Bedrock, Azure OpenAI, Baseten, Cerebras, Cohere, Cloudflare AI Gateway, Cloudflare Workers AI, DeepInfra, DeepSeek, Fireworks, Groq, Mistral, OpenRouter, TogetherAI, Vercel AI Gateway, and xAI. Z.ai currently exposes image generation. Generic Chat Completions, Responses, and Anthropic Messages-compatible entrypoints support custom endpoints.
Each named provider owns its module, endpoint, authentication, and route setup. Providers with the same wire format compose the shared protocol directly:
@@ -802,13 +1321,14 @@ APIs have separate entrypoints:
OpenAI Responses has one semantic route and uses HTTP by default. Advanced callers may supply a per-call WebSocket channel executor through `StreamOptions`; transport policy does not change provider settings, model identity, or route identity. The provider-neutral Open Responses implementation owns the reusable WebSocket request and event contract, while each provider opts in with its own handshake and connection policy. Azure follows the same Chat/Responses split at `providers/azure/chat` and `providers/azure/responses`. Generic OpenAI-compatible Chat remains at `providers/openai-compatible`; the Responses adapter at `providers/openai-compatible/responses` uses the provider-neutral Open Responses protocol. OpenAI Responses extends that baseline with OpenAI tools, event variants, metadata, and defaults. Generic Anthropic Messages-compatible providers use `providers/anthropic-compatible`, which the named Anthropic provider composes. Google Gemini and Amazon Bedrock expose their single native API through their existing provider paths.
Vertex Gemini, Vertex Chat, Vertex Responses, and Vertex Messages are separate API entrypoints. All accept `project`, `location`, and an optional `accessToken`; when no explicit token or auth override is supplied they lazily use Google Application Default Credentials. Vertex Gemini instead selects express mode when `apiKey` or `GOOGLE_VERTEX_API_KEY` is present. Vertex Chat targets MaaS models through the OpenAI-compatible Chat Completions endpoint, while Vertex Responses targets Grok models and defaults `store` to `false` as required by Vertex. `providers/google-vertex` remains the default alias for `providers/google-vertex/gemini`.
Vertex Gemini, Vertex Interactions, Vertex Chat, Vertex Responses, and Vertex Messages are separate API entrypoints. All accept `project`, `location`, and an optional `accessToken`; when no explicit token or auth override is supplied they lazily use Google Application Default Credentials. Vertex Gemini and Interactions instead select express mode when `apiKey` or `GOOGLE_VERTEX_API_KEY` is present (explicit `project` or `location` selects ADC over an ambient key). Vertex Chat targets MaaS models through the OpenAI-compatible Chat Completions endpoint, while Vertex Responses targets Grok models and defaults `store` to `false` as required by Vertex. `providers/google-vertex` remains the default alias for `providers/google-vertex/gemini`.
Tuned Vertex Gemini deployments use model ids shaped like `endpoints/1234567890` and require OAuth or ADC; Vertex express-mode API keys support publisher models only.
@@ -873,7 +1393,7 @@ Compose a route with `Route.make({ protocol, endpoint, auth, framing, ... })`. T
## Effect
This package is built on Effect. Public methods return `Effect` or `Stream`; provide `LLMClient.layer`for LLM dispatch and `ImageClient.layer` for image dispatch, then import the provider/protocol modules for the routes you use. The example at `example/tutorial.ts` is a runnable walkthrough.
This package is built on Effect. Public methods return `Effect` or `Stream`; provide `AIClient.layer`(or `AIClient.layerWith(executor)`) for every modality, then import the provider/protocol modules for the routes you use. The example at `example/tutorial.ts` is a runnable walkthrough.
Loaded 100 of 5053 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Filip Hejmowski
