Peter SteinbergerandGitHub
06f5f73e47
* refactor(agents): prepare model runtime catalogs Build lifecycle-owned model and auth snapshots, carry prepared stores into hot agent paths, and serialize config/auth publication. Credits @zeroaltitude's #90741 investigation and benchmark approach. * refactor: finish lifecycle-owned model discovery * fix: align prepared model catalog contracts * test: align lifecycle catalog mocks * test: fix prepared catalog type fixtures * refactor: split prepared model runtime ownership * fix: import prepared runtime replacement gate type * test: split media runtime coverage * test: preserve image auth fixture key types * test: isolate lifecycle gate fixtures * chore: keep release changelog owned * refactor: finish prepared model catalog migration * test: keep catalog review fixtures scanner-safe * refactor: preserve lifecycle model runtime ownership * fix: close prepared runtime lifecycle races * fix: preserve compaction workspace fallback * chore: document btw generation rebinding * fix: preserve prepared generation boundaries * fix: keep model-list discovery flag explicit * fix: serialize standalone model runtime activation * refactor: migrate subagent model catalog lookup * refactor: clarify doctor catalog lookup seam * chore: refresh plugin sdk api baseline * test: migrate swarm catalog dependency * refactor(telegram): rename runtime catalog seam * refactor: extract model-aware tool context * test(models): isolate lifecycle catalog fixtures * refactor(agents): avoid btw parameter rebinding * fix(net-policy): align root ipaddr dependency * fix(build): keep net policy dependency bundled * fix(deadcode): document net policy compile dependency
5.5 KiB
title, summary
| title | summary |
|---|---|
| Agent runtime architecture | How OpenClaw structures the built-in agent runtime: code layout, boundaries, resource manifests, and runtime selection. |
OpenClaw owns the built-in agent runtime. Runtime code lives under src/agents/, model/provider transport lives under src/llm/, and plugin-facing contracts are exposed through openclaw/plugin-sdk/* barrels.
Runtime Layout
| Path | Owns |
|---|---|
src/agents/embedded-agent-runner/ |
Built-in attempt loop (run.ts, run/), model selection and provider normalization (model*.ts), per-provider request params (extra-params.*), compaction, transcript and session wiring. |
src/agents/sessions/ |
Session persistence (session-manager.ts), resource discovery (package-manager.ts, resource-loader.ts), in-session extensions loading, prompt templates, skills, themes, and TUI-backed tool renderers (tools/). |
packages/agent-core/ |
Reusable agent core (@openclaw/agent-core): agent loop, harness types, messages, compaction helpers, prompt templates, skills, and session storage contracts. |
src/agents/runtime/ |
OpenClaw facade that wires @openclaw/agent-core to the plugin SDK LLM runtime and re-exports it plus local proxy utilities. |
src/agents/agent-tools*.ts |
OpenClaw-owned tool definitions, parameter schemas, tool policy, before/after tool-call adapters, and host/sandbox edit tools. |
src/agents/agent-hooks/ |
Built-in runtime hooks: compaction safeguard, compaction instructions, context pruning. |
src/agents/harness/ |
Harness registry, selection policy, and lifecycle for the built-in and plugin-registered harnesses. |
src/llm/ |
Model/provider registry, transport helpers, and provider-specific stream implementations (src/llm/providers/). |
Boundaries
Core calls the built-in runtime through OpenClaw modules and SDK barrels; no external agent framework packages remain. Plugins use documented openclaw/plugin-sdk/* entrypoints and do not import src/** internals.
@earendil-works/pi-tui remains a third-party dependency: a terminal component toolkit used by the local TUI and session tool renderers. Internalizing it would be a separate vendoring effort.
Manifests
Resource packages declare OpenClaw resources in package.json metadata. Entries are file paths or globs relative to the package root:
{
"openclaw": {
"extensions": ["extensions/index.ts"],
"skills": ["skills/*.md"],
"prompts": ["prompts/*.md"],
"themes": ["themes/*.json"]
}
}
Resource types not listed in a manifest fall back to discovery of conventional extensions/, skills/, prompts/, and themes/ directories.
Runtime Selection
- The built-in runtime id is
openclaw. The legacy aliaspinormalizes toopenclaw;codex-app-servernormalizes tocodex. - Plugin harnesses register additional runtime ids (for example
codex). - Runtime policy is model/provider-scoped
agentRuntime.idconfig (model entry wins over provider entry). Unset ordefaultresolves toauto. autoselects a registered plugin harness that supports the effective provider route, otherwise the built-in OpenClaw runtime. A provider or model prefix alone never selects a harness.- OpenAI may select
codeximplicitly only for an exact official HTTPS Platform Responses or ChatGPT Responses route with no authored request override. Completions adapters, custom endpoints, and routes with authored request behavior stay onopenclaw; plaintext official HTTP endpoints are rejected. See OpenAI implicit agent runtime.
Model Runtime Generations
Gateway startup and config, plugin, or auth publication build one prepared model runtime generation per configured agent. Each generation owns the discovered auth template, model registry, and projected model catalog as one atomic snapshot. Agent runs fork mutable auth and registry stores from that snapshot; browse, status, cron, doctor, TUI, PDF, and image paths read the published catalog instead of repeating filesystem discovery.
Standalone embedded runtimes publish the same snapshot shape at their activation boundary. A failed or stale generation is never served alongside a newer partial generation; the lifecycle owner must publish a complete replacement first.