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
LifeOSHooks: promptSubmit puts its composed context on the result after next(), which the engine does not attach, so memory, nudges and the ORIGIN warning never reach the model #2822
What.skills/LifeOSHooks/hooks/LifeOSHooks.tspromptSubmit builds its per-turn entries before next() (the time line, then the prompt-context bridge reply: the memory hot layer, AlgorithmNudge, VersionDrift, DriftReminder, ModelDriftGuard, PresenceContext). After next(input) resolves, it appends the FAST-TOGGLE and ORIGIN lines and returns { ...r, context }.
The client's own plugin declarations (PromptSubmitResult.context, as written by 2.1.322) say:
A hook attaches context on the way down; one put here after next resolved is not attached (the prompt had entered), and is logged.
What happens. None of the plugin's prompt context reaches the model. In one live session, no time line, memory block or ORIGIN: line appeared in the model's context on any prompt, including prompts the plugin logged with origin task-notification. Context from native command hooks (a PostToolUse additionalContext) did arrive in the same session. Because native MemoryTurnStart exits early in an armed session, an armed session gets no hot memory at all.
The ORIGIN line is the warning that a prompt did not come from the principal and carries no authority. It is lost the same way.
Suggested fix. Build every entry before next() and pass it down, then return the result unchanged. Keep the side effects that need an accepted prompt (scheduleFast, the memory ack) after next():
Related. The memory ack in the same hook depends on this. It marks chunks consumed after next(), but those chunks were never attached, so the ack confirmed deliveries that hadn't happened. #2608 (closed, fixed for the next release) moves the ack after an accepted next() and carries the emission id back from prompt-context; its note does not say whether the context itself moved before next(). If that fix already passes the entries down through next(), this issue is covered and can be closed.
A second issue that shows up once context is delivered. When one composer inside prompt-context throws, the bridge withdraws the prompt-context cap and the native command hooks run for that prompt as well, so that turn carries some entries twice. Dropping the plugin's entries in that case doesn't fix it cleanly. Native TimeContext and VersionDrift are registered async: true, so their output arrives a turn late (and TimeContext falls back to UTC when no timezone is configured). VersionDrift has also already started its cooldown during the bridge's run, so its warning would be lost. PromptProcessing and SatisfactionCapture run in the bridge before coverage is decided, so their model request and rating pulse repeat natively either way. A fallback that is equivalent to the bridge would let the bridge report which composers failed, so that only those fall back to their native hooks.
Activity