[kinqsradio]

See CONTRIBUTING before starting engineering work.

Refs #258. Stacked on #259: the base is feat/project-chat-dock. Retarget to feat/web-rebuild once #259 merges.

Change

• Project scope reaches the model on every run. Since the runner refactor (34dd881b), scoped facts were checked at admission and then dropped, so "which project is selected?" had nothing to go on. contextFor now rereads the scope (agent/chat-context.ts) with the job owner's current access on every attempt and resume. A scope that was kept out or lost access stops the run before any provider call. The lead and specialist prompts name the selected project or task. Titles appear as escaped, one-line, bounded data.
• Workspace view with each turn. chatRequest.view is optional and strict: {pane, path, query}, where the path is an /app/... pathname. The server derives any open task, page or project from the path. It drops that item unless the assistant may read it (visibility, kept out, team assistant setting), and takes titles from the database only. Under restrictive read rules only the screen is named. The view is a location hint and never changes the chat's scope.
• Web dock: sends the working frame's pane, path and query at send time. The api-client now forwards view.
• Navigation: dock links already go through the working route scope into WorkspaceController. No change.
• Refresh: project reads now also refresh on review news, which is what approving a proposal announces. The approving tab already refreshes the planner store.
• Approvals: unchanged. Applying a proposal stays people-only.
• Finalization: a reviewed answer that only finalizes skips the scope reread, so an applied plan isn't reported as failed after its task is deleted. If the checker blocks that plan and it goes back to the model, the scope is checked again first.

Checks and limits

• Run locally: npm run format:check, npm run typecheck (all workspaces), npm run build -w web, scripts/check-code-organization.py, scripts/check-project-docs.py.
• Pure tests pass, 36/36: backend/tests/assistant-chat-context.unit.test.ts (new), web/tests/workspace-chat-guards.test.ts and backend/tests/web-workspace.unit.test.ts.
• New DB tests, not run locally (no Postgres on this machine); CI runs them. They are in backend/tests/assistant-runs.test.ts and cover:
  • the provider sees the selected project and the checked view;
  • a foreign or kept-out project in the view never reaches the provider;
  • a waiting project chat makes no provider call after its project is kept out.
• No browser or live provider check.
• Threat model (design mode, quick) covered three risks: view → context leaks, resume after access loss, and prompt text → tools. The two leak and resume risks are checked by the new DB tests above.
• An independent review found a restricted-rules title leak, a finalization regression and a blocked-plan recheck gap. All three are fixed here and confirmed by the reviewer.
• Not in scope:
  • other tabs and devices still refresh the planner store by polling;
  • docs/architecture.md still describes the retired fixed graph;
  • citations from capability reads are still not shown.

AI usage

• Tools: Claude Code (desktop app), Git, GitHub CLI, npm, Prettier, tsc, Node test runner.
• Models: Claude Opus 5.5 at medium effort led: design, threat model, backend implementation and fixes. Opus 5.5 subagents at medium effort did the code trace and the independent review. A Claude Sonnet 5.5 subagent at medium effort implemented the web client slice.
• Human role: asked for Wire workspace dock chat to project context, views, and workspace actions #258 to be resolved under the workspace rules. Review pending.
• External sources: none copied.

Workspace

• Used the outer Orbyn Engineering workspace with this target nested under repo/, and read the relevant project guides.

This contributor confirmation is not an automated claim that a host loaded policies.

🤖 Generated with Claude Code