You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[bug] the LifeOSHooks session-start stamp ignores the bridge's reply and is never… #2798
the LifeOSHooks session-start stamp ignores the bridge's reply and is never retried when it fails; the per-prompt retry from #2753 only covers a changed session id
Summary
Follow-up to #2753 (closed). That fix re-reads the session id on each prompt and retries the stamp bridge verb until it is acknowledged, but only when the id changed: stampPending is set only inside rebindSession.
The session-start hook sends the same stamp verb and discards the reply. If that first stamp returns a not-ok reply or times out, nothing sets stampPending, so the session's first id is never stamped and never retried.
Suggested fix, in sessionStart of skills/LifeOSHooks/hooks/LifeOSHooks.ts: keep the reply and, when it is not ok: true, set s.stampPending = s.sessionId. The existing per-prompt retry in rebindSession then covers the start path too, and a failed stamp still never blocks the session.
Reproduce: a unit test with a fake bridge that answers stamp with a not-ok reply (or rejects) at session start, then accepts; before the change only one stamp call is ever made, after it the stamp is retried on each prompt until acknowledged and not sent again.
Details
Kind: bug
LifeOS version: 8.761.0
Filed with the LifeOS /feedback command. The reporter approved this exact text before it was sent; paths, secrets and personal details were stripped locally.
A related gap found by a static audit of the plugin on LifeOS 8.761.0 (plugin files unchanged on main): the stamp is not a liveness signal, so a stamp that was written once can outlive the plugin.
The bridge commits the stamp and its capability set once. mergeStampCaps returns early when nothing changed, so the mtime is not refreshed after the first prompt, and hookDrawn trusts the stamp for 7 days (hooks/README.md says 24 h). The native hooks that defer to it (memory, nudges, drift, time, naming, satisfaction, last-response cache, voice, subagent logging, ISA sync echo, format gate) then stay silent for the rest of the session in these cases:
the module is hot-reloaded and does not re-arm (register resets state to armed: false; whether the engine re-fires session.start after a reload is not shown);
the plugin is disabled and /reload-plugins runs mid-session;
prompt-context times out on the plugin side after an earlier success.
Suggested fix: a per-prompt proof. prompt-context writes the sha of the prompt it composed, and each native hook skips only when the sha of its own stdin prompt matches and was written within a few seconds; the same for turn-complete and voice.
Details
Kind: feedback
Filed with the LifeOS /feedback command. The reporter approved this exact text before it was sent; paths, secrets and personal details were stripped locally.
Activity
tbouquet commentedon Nov 18, 2026
[feedback] stamp liveness follow-up
What happened
Comment on #2798: stamp liveness follow-up
Summary
A related gap found by a static audit of the plugin on LifeOS 8.761.0 (plugin files unchanged on main): the stamp is not a liveness signal, so a stamp that was written once can outlive the plugin.
The bridge commits the stamp and its capability set once. mergeStampCaps returns early when nothing changed, so the mtime is not refreshed after the first prompt, and hookDrawn trusts the stamp for 7 days (hooks/README.md says 24 h). The native hooks that defer to it (memory, nudges, drift, time, naming, satisfaction, last-response cache, voice, subagent logging, ISA sync echo, format gate) then stay silent for the rest of the session in these cases:
armed: false; whether the engine re-fires session.start after a reload is not shown);Suggested fix: a per-prompt proof. prompt-context writes the sha of the prompt it composed, and each native hook skips only when the sha of its own stdin prompt matches and was written within a few seconds; the same for turn-complete and voice.
Details
Filed with the LifeOS
/feedbackcommand. The reporter approved this exact text before it was sent; paths, secrets and personal details were stripped locally.