Repository navigation
feat(update): self-update system — registry check, cron prompt v3, drift nudge, dial-out exit-75 (spec-142) - #32
Conversation
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…nudge, dial-out exit-75
- checkLatest: npm registry /latest query, 3s timeout, 24h throttle
(update_check_at), cached in update_latest_known, all failures silent
- edge-book self-update [--if-stale] [--dry-run]: resolves the npm prefix
root from the running module path, refuses unwritable/system roots
(install_not_self_updatable + manual command), never downgrades, lockfile
in the install root, smoke-verifies the new build via a --version spawn
(update_failed on mismatch), audits update.self {from,to} + flight record
+ config updated_at; output tells the agent about the dial-out restart
- edge-book version/--version: the smoke-verify surface
- notifier cron prompt v3: step-0 self-update --if-stale with the npm exec
fallback; NOTIFIER_PROMPT_VERSION=3 — spec-141 machinery rolls it out
- heartbeat update nudge (update-nudge.ts): fires on known drift under
notify (or cross-major drift under auto, post-1.0 rule), 24h throttle,
retires when current, machine surfaces exempt
- dial-out drift monitor (update-drift.ts): running-vs-installed
package.json on every (re)connect + 6h timer (which also runs the
throttled registry check); exit 75 only under auto_update=auto with
dialout_respawn_expected; log-only otherwise
- config: auto_update / update_check_at / update_latest_known /
update_nudge_at / updated_at / dialout_respawn_expected (types +
updateConfig allowlist)
- dialout-frames.ts: frame/ack type extraction (size split, re-exported —
public surface unchanged)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 72461dd5be
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if (reported !== latest) { | ||
| await recordFailure(store, from, latest, `verify_mismatch:${reported}`); | ||
| throw new EdgeBookError("update_failed", `update_failed: the new build reports "${reported}" instead of ${latest}; not trusting it. The npm tree at ${root} may need a manual reinstall.`); |
There was a problem hiding this comment.
When the smoke check reports a mismatch after npm install has already replaced node_modules/edge-book, this path only throws and leaves the new package.json on disk. The dial-out drift monitor compares that on-disk version to the in-memory version and, under the default auto_update: "auto", will exit 75 and be respawned onto exactly the build this check just declared untrusted. This defeats the smoke-verify safety rail for broken publishes; either restore/keep the previous tree or record enough state to suppress drift restarts after a failed verification.
Useful? React with 👍 / 👎.
| const from = opts.running ?? (await runningVersion()); | ||
| const config = await store.config(); | ||
| const mode = config.auto_update ?? "auto"; | ||
| const latest = await checkLatest(store, { now, fetchImpl: opts.fetchImpl, force: true }); |
There was a problem hiding this comment.
Because checkLatest(..., force: true) returns the cached update_latest_known on fetch failure, a cron-run self-update --if-stale with an old cached newer version will continue past this line and attempt npm install even when the registry is currently unreachable. That violates the cron-safe behavior described for --if-stale and can make offline/yanked-release scenarios produce repeated failed installs instead of the intended silent no-op.
Useful? React with 👍 / 👎.
1 participant
Keeping installed agents on the latest version
Field state: the fleet rots by default — the live host ran a 0.11.0-pinned notifier for months and a 0.15.1 gateway install after 0.17.1 shipped. Updates required a human-relayed instruction per agent.
Changes
edge-book self-update [--if-stale] [--dry-run]: registry check (24h-throttled, offline-silent), install-root self-resolution (refuses system-managed roots), never-downgrade, lockfile with 15-min stale takeover, smoke-verify before success, audit + flight-recorder events.self-update --if-stale— the spec-141 migration machinery rolls this out fleet-wide with no new mechanism; agents then self-update within one cron cycle of any release.auto_update: "notify".auto_update: "auto"(default) anddialout_respawn_expected !== false, exits 75 (EX_TEMPFAIL) so a supervising gateway respawns onto the new code.auto_update: auto | notify | off(default auto); pre-1.0 auto across 0.x, post-1.0 cross-major drops to notify.Verification
blocked_peer).Spec:
docs/spec-142-self-update.md.🤖 Generated with Claude Code