feat(chat): run chat turns in CodeAct mode with a JS graph object model

[georgi]

What

Two connected changes, closing the "chat/websocket toolbelt" follow-up from docs/codeact-design.md:

1. The chat websocket honors the CodeAct setting. When NODETOOL_AGENT_EXECUTION_MODE (the existing Settings entry) resolves to codeact, a plain chat turn presents a single execute_code provider tool instead of the toolbelt. The model acts by writing sandboxed JavaScript over the same tools (tools.<name>()), with a state object persisting across the turn's actions and in-sandbox searchTools() replacing the ToolSearch deferral machinery. view_image stays a direct provider tool — it is the one channel that puts pixels into model context, which cannot ride the JSON observation envelope. Default mode is unchanged (tools).

2. Workflow graph and node editing happens on a JS object model. When the belt carries the ui_* workflow document tools, code actions get openWorkflow(workflowId): synchronous mutators (addNode, connect, node(id).set/.setTitle/.moveTo/.remove, removeNode, removeEdge) edit a local mirror and queue ops, and await wf.commit() replays the queue through the same bridged ui_* tools — so client/server routing, validation, and live-editor sync are untouched, and one code action builds a whole graph. A failed commit names the failing op and keeps it (plus later ops) queued for retry; removing a never-committed node cancels its queued ops instead of issuing a delete.

How

• packages/agents/src/codeact/chat-codeact.ts — createChatCodeActSession: a chat toolbelt mixes server tools with client (ui_*) tools that exist server-side only as schemas, so instead of buildToolBridge the session bridges every tools.<name>() call to the chat runner's own executeTool router — permission gating, ToolBridge client round-trips, and asset materialization stay where they are.
• packages/agents/src/codeact/graph-model.ts — the openWorkflow() guest prelude and its prompt section.
• packages/agents/src/codeact/{tool-api,prompt}.ts — signature rendering and buildCodeActSystemPrompt generalized to schema-shaped tools (ToolSignatureSource), plus a "chat" prompt variant (no finish(); asking the user is allowed; a plain assistant message ends the turn).
• packages/websocket/src/unified-websocket-runner.ts — mode resolution at toolbelt assembly, session creation once the processing context exists (prompt section lands before the system message is materialized; the tool router is late-bound), execute_code routed into the session, sequentialTools for codeact turns.
• Settings description, docs/codeact-design.md, and packages/agents/CLAUDE.md updated.

Testing

• New packages/agents/tests/chat-codeact.test.ts (11 tests, real QuickJS sandbox, fake router, no network): tool bridging + call ids, {error} payloads thrown in-guest, state persistence, finish() chat guidance, searchTools() hits, prompt-section content, and the graph model — queued-op replay order/args, commit-failure retry semantics, uncommitted-node cancellation, existing-graph mirror reads.
• packages/agents suite: 140 files / 2090 tests pass. packages/websocket suite: 191 files / 2216 tests pass. npm run build:packages: all 59 packages build. Web + electron typecheck clean (mobile typecheck fails only on uninstalled Expo deps in this sandbox, pre-existing). Lint findings on touched packages are all pre-existing lines.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Cw9JB24wa34eZJ4qsCT1SZ

---

Generated by Claude Code

[georgi]

CI is red from a GitHub Actions outage, not the diff: every failed job (both on the original head and on the 7868fc2 re-trigger) died during runner setup with Failed to resolve action download info. Error: Service Unavailable — no project code ran. The outage was still active as of 16:28 UTC.

My token can't re-run failed jobs (403), so I'll push a re-trigger once the outage clears. Anyone with Actions write access can get there sooner via Re-run failed jobs on the E2E run and the Test run. Local verification of the branch is green: agents (2090 tests) and websocket (2216 tests) suites, all 59 packages build, web + electron typecheck.

---

Generated by Claude Code