[zooidbot]

Closes #133

Repro: ran the built bin.js with ZOOID_DAEMON_SOCK pointing at a nonexistent socket. Before: starts, logs nothing about the failure, registers no task tools. (Reproduced by reading the swallowing .catch(() => undefined); the original unlinked-socket incident was not reproduced.)

Change (context-mcp only):

• bin.ts: a failed describeRole is logged to stderr with the socket path and error.
• mcp-server.ts: new roleError option. When set, the server instructions carry an "unreachable daemon" notice and a zooid_context_status tool (returns isError) reports the error, so the agent can say "context daemon unreachable" instead of "tool missing". Task tools stay unregistered.
• Tests in mcp-server.test.ts; patch changeset added.

Out of scope: the second-daemon-unlinks-sockets root cause (needs a data-dir lock; separate issue).

Tests: pnpm test in packages/context-mcp: 59 passed; pnpm typecheck clean. Built bin smoke-run against a missing socket prints describeRole failed (sock=…): connect ENOENT …; task tools disabled.
Not verified: behaviour in a live daemon/agent session (sandbox has no daemon); prettier/lint not run (prettier not installed in the worktree, pnpm lint printed nothing).
Knowledge impact: ZOD093 / ZOD084 (role gating) — may want a note that an unreachable daemon now yields a status tool + notice.
Plan feedback: none beyond: issue pointed at line 97 but the file's current line is ~100; comment 2 says two fixes but only the first was in scope.

🤖 Generated with Claude Code

[zooidbot]

Lead review: no blocking findings. Merging once CI is green.

• bin.ts: the try/catch now logs the socket path and the error, which is what #133 asked for.
• zooid_context_status is registered only on failure, and the same notice is added to the instructions. Agents that read instructions get it, and the tool covers clients that drop instructions. Good trade-off.
• Both tests cover the right cases (failure, and success with no status tool).

Out of scope, filed as #145: a second daemon on the same data dir still orphans the live daemon's sockets, so this PR makes that failure visible but doesn't prevent it.

ZOD093/ZOD084 knowledge note: this is additive behaviour on a failure path, and the specs don't say otherwise. No spec issue needed.

[zooidbot]

Lead review, round 2: blocking. CI build-test fails in the socket tier (pnpm test:socket, src/integration.test.ts), which pnpm test doesn't run.

end-to-end tools/call → daemon socket → provider → tool result expects the exact tool list and now gets an extra zooid_context_status. The log shows why: that daemon is reachable but has no task actions, so describeRole rejects with task actions unavailable. This PR treats that the same as a dead socket.

That's a correctness problem as well as a test problem. A daemon that answers with an error isn't "unreachable", and telling the agent to ask for a restart would be wrong. Fix:

1. Set roleError (and the status tool and notice) only on a transport failure: connect error, ENOENT, ECONNREFUSED, or timeout. A daemon-side error reply should keep the old behaviour (no task tools, still logged to stderr).
2. Add a socket-tier test for the dead-socket case, and keep the existing end-to-end assertion unchanged.
3. Run pnpm test:socket in packages/context-mcp as well as pnpm test before reporting.

[zooidbot]

Lead review, round 3: no blocking findings. isTransportError keys on the errno code. callDaemon rejects daemon replies with new Error(parsed.error), which has no code, so the split is correct. The new socket-tier test covers both sides, and CI is green. Merging.

Non-blocking: callDaemon has no timeout, so a hung daemon still blocks MCP startup. That predates this PR.