feat(cli): ensure the daemon explicitly before attaching in the CLI and MCP

[thymikee]

Summary

Layer 3 of 3 for #3754 (umbrella #3747); based on #3765. The CLI and MCP stop managing the daemon implicitly.

• withEnsuredDaemon wraps the CLI's default transport and each MCP tool call's client. The first request that needs a local daemon calls ensureDaemon() once; every request then attaches. A command that never reaches the daemon starts none.
• Intentional CLI behavior change: a daemon from another version or build is replaced only if idle. A busy one fails the command with daemon-busy instead of losing its sessions; one without the quiesce capability fails with daemon-incompatible. Recorded in the changelog and CLI help, and in the client docs.
• agent-device proxy is unchanged and stays on the managed path.
• A request that finds its ensured daemon gone ensures again and retries once; nothing was sent.
• Pass-through, unchanged: remote --daemon-base-url connections and a one-shot replay/test on its private ephemeral daemon. A timeout no longer retires an unresponsive daemon from the CLI or MCP (changelog).

11 files incl. docs. Closes #3754.

Validation

pnpm check:affected --run --head e1f0bca89a: passed. Tests cover ensure-once, idle replacement, busy refusal before send, remote and private-replay pass-through, and re-ensure after the daemon vanishes. An independent review of layers 1-3 found 7 issues. Five are fixed. Uploads and lease-less holds being outside the census, and quiesce over HTTP with an auth hook, are documented in ADR 0006. A hung but connectable daemon is no longer retired by a CLI timeout; the changelog says so.

🤖 Generated with Claude Code

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-17 08:59 UTC

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.34 MB	5.34 MB	+3.1 kB
Package (unpacked)	5.33 MB	5.34 MB	+3.1 kB
Package (download)	1.61 MB	1.61 MB	+958 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.4 ms	26.4 ms	+0.1 ms
CLI --help	75.7 ms	75.4 ms	-0.3 ms

[cubic-dev-ai]

All reported issues were addressed across 8 files

_{You’re at about 92% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 5 files (changes from recent commits).

_{You’re at about 92% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 4 files (changes from recent commits).

_{You’re at about 93% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 5 files (changes from recent commits).

_{You’re at about 93% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

<sub>Requires human review: Auto-approval blocked because this review re-detected 1 unresolved issue already reported by Cubic.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[thymikee]

This PR is ready at 125b4cc. I found no problems that should hold it up. All 9 checks that are not passing were cancelled as superseded runs, so there is no failure to attribute, and they need a completed run at 125b4cc before merge. The diff touches their routes, because the CLI default transport is now the ensured one and daemon-replace-exit-flush.test.ts changed. No conflicts.

Not blocking, so take or leave these: daemon-client-ensuring.ts:95 maps a daemon that was force-retired as unresponsive to reason 'incompatible', so cli.ts:88 prints "Replaced idle daemon ... it did not match this client", and the 'unreachable' label is inferred from 'missing' plus a previous daemon, so the notice should name how the previous daemon left and ensureDaemon's onStart could report that cause as a declared value; daemon-lifecycle-errors.ts:77, line 61 and daemon-client-timeout.ts:195 now reach CLI and MCP users with hints that name ensureDaemon(), so the hints should hold on every surface (for example, stop it with agent-device daemon stop --state-dir X); command-tools.ts:251 wiring through the default createClient has no test, so one MCP tool-call test could assert that daemonLifecycle 'attach' reaches the transport and that a second call ensures again; and in daemon-client-ensuring.ts:50 the "Settled promises are not kept" comment is wrong (only rejected ones are dropped), ?? process.env.AGENT_DEVICE_STATE_DIR repeats what resolveExplicitStateDir reads, the MCP cast at command-tools.ts:258 duplicates the one in cli.ts and goes away if createEnsuredDaemonTransport returns AgentDeviceDaemonTransport, the test at ensuring.test.ts:252 should assert details.reason instead of matching /busy/, and the re-ensure test at ensuring.test.ts:183-188 should build daemon-not-running through daemonNotRunningError.

I did not run the tests or check:affected, and I did not measure the latency of the extra per-tool-call ensure in MCP. I took the busy/idle census behavior from the base layer without reading its implementation.

Nine inline threads from cubic-dev-ai (all P2) no longer apply at this head, so the author can resolve them: r4299317645 (entry deleted only when it matches the relied one), r4299317658 (rejected ensure is dropped), r4299317674 (cli-help excludes device status and release --stale), r4299317719 (commands.md names the exclusions), r4299317737 (CHANGELOG names the census exclusions), r4299430976 (test binds both listeners before closing), r4299430986 (CHANGELOG no longer says idle for an unresponsive daemon), r4299793590 (import promise is memoized), and r4299837292 (a pre-aborted request skips ensuring; a mid-ensure abort is documented and the caller's guard settles it).

[thymikee]

Thanks. Taken in e9cab3f:

• Force-retired daemon: a daemon retired for timing out on quiesce and failing liveness probes is now reported through onStart as missing with a previous (the same as an unreachable one), so the CLI prints "Replaced unreachable daemon" and never "Replaced idle daemon". The onStart contract says so.
• Hints hold on every surface (see feat(client): daemonLifecycle attach mode and ensureDaemon() #3765).
• MCP wiring: new test drives the default client factory twice and asserts one ensureDaemon per tool call and a single daemon start.
• Nits: the stale "Settled promises" comment is fixed, the redundant AGENT_DEVICE_STATE_DIR fallback is dropped, the busy-refusal test asserts details.reason, and the re-ensure test builds its error with daemonNotRunningError.
• Not taken: the MCP cast. createEnsuredDaemonTransport returning the client transport type would loosen CliDeps.sendToDaemon, which the CLI tests inject fakes into; the cast stays at the one boundary where the request types differ.
• agent-device proxy stays on the managed path (see feat(client): daemonLifecycle attach mode and ensureDaemon() #3765), so nothing about it changes here.

pnpm check:affected --run passes at e9cab3f, plus the full unit and Node integration suites. CI is re-running on this head.

[thymikee]

The PR is ready at e1f0bca. The earlier pass at 125b4cc was clean, and the nine Cubic threads are now fixed at this head: the lost-daemon cache cleanup, the dropped rejected promise, the help and docs exceptions, the changelog wording, the test port reuse, the shared import promise, and the already-aborted signal case. I read the code but did not run the tests, so the MCP spy behavior through the dynamic import comes from reading, not from a run.

Not blocking: daemon-client-ensuring.ts:69 compares details.reason to the bare string 'daemon-not-running', while daemon-lifecycle-errors.ts keeps DAEMON_LIFECYCLE_REASONS.notRunning module-private. You could export a predicate such as isDaemonNotRunningError(error) from the owner and use it there, or leave it as is.

The threads from the other reviewer all apply at head, so you can resolve them: #3766 (comment), #3766 (comment), #3766 (comment), #3766 (comment), #3766 (comment), #3766 (comment), #3766 (comment), #3766 (comment), #3766 (comment).

No conflicts are known. CI on e1f0bca has not finished: the 22 checks that have not passed are cancelled or queued runs left over from the rebase, and none has failed. Please rerun them and wait for green checks before merge.