continuity-brief

continuity-brief

Estelle's session-start brief — continuity of self that a model, reset to amnesia every session, is denied by its own design. Estelle keeps the memory the model can't: save_context stashed the last session's state under a label, RunHistory recorded every action, the TemporalLog holds the team's live decisions, PresenceLog knows who worked while you were away, and the ExperienceStore holds what you learned. This is the exact mirror of baton-pass and context-checkpoint: they SAVE the handoff at session END; continuity-brief LOADS and synthesizes it at session START. The model wakes up with a memory instead of a blank slate.

When to run

At the very start of a session on a repo you (or a sibling agent) have worked before — before the first real task. Run it to answer "who am I on this repo, what did I do last, what's pending, what changed while I was away, and what have I already learned not to repeat?" — so the first action is informed, not a rediscovery.

Procedure

1. Load the last handoff bundle. list_contexts for the saved labels, then load_context(label) (POST /context/load) to pull the exact context bundle baton-pass / context-checkpoint wrote at the last session's end — its done / in-progress / next / key-decisions, verbatim. If no label fits, search_contexts by the current task to find the relevant handoff by meaning.
2. Read what you actually did last. RunHistory.recent(limit) returns the newest-first agent runs — which branches, what was taken versus refused by the ceiling, the artifacts (diffs / PRs) produced; runs_to_markdown renders them team-readably. This is a record of your own actions, not a guess about them.
3. Establish who you are on THIS repo. Query the team's live decisions: facts() with neither key nor at returns TemporalLog.current() — every still-open decision you must not contradict. Recall the swept surface and conventions so you write in-house from the first line, not after a rejected round.
4. Measure how much time passed. Read the clock and presence.report(now): overnight() tells you someone (or an agent) worked the night window while you were away, files_in_use flags the files other members are touching right now so a fresh start doesn't collide, and handoffs surfaces the questions left for you.
5. Reload what you learned. ExperienceStore.recall(task) pulls the Reflexion lessons from past failures on this repo — the mistakes not to repeat — plus the graduated InstinctEngine reflexes the team has already proven, so the session opens with hard-won context, not just facts.
6. Synthesize ONE brief, not five dumps. Fuse the bundle + run history + current facts + presence + lessons into a single short waking brief: who you are on this repo, what you did last, what's pending (the pinned resume point), what changed while you were away, and the decisions/lessons that constrain what you may do.
7. Resolve staleness before acting. A loaded fact can be superseded — if a decision looks old, check facts(key=...) for its history() or facts(at=...) for what was believed then, and treat a grounding lesson whose warned symbols now exist (expire_known) as retired, so you never act on a fact that stopped being true while you were gone.
8. Land on the resume point. End the brief on the single next action from the handoff's "next" list — the precise file, branch, or command to start from. Make "where do I begin?" a non-question, the exact mirror of what baton-pass pinned when it saved.