long-horizon-recoverylisted

# Long-horizon recovery Chat/session context is disposable. Operational state should survive when the work is substantial enough to justify continuity, without publishing private working notes into the repository. Continuity is evidence, not context injection. A resumed execution session still starts from Agentit's three-skill core and explicitly re-selects any non-core skills, references, workers and MCPs it still needs. ## Canonical state Default local state: - `.agentit/STATE.md` - `.agentit/checkpoints/*.json` See `docs/PROJECT_CONTINUITY.md`. Do not create or commit `docs/agentit/STATE.md` merely because Agentit is active. If a project already has an intentional tracked team-status document, it may be reused explicitly; otherwise operational state stays local/private. A useful state snapshot includes: - goal and confirmed constraints; - relevant packs; - `complexity: trivial | bounded | substantial | structural` when material; - risk and topology; - previously selected skills, tools, references and workers; - MCPs enabled by this task and cleanup ownership when relevant; - ownership boundaries; - compact `TASK_DECISION` summary and reviewer verdicts; - branch/PR and important artifacts; - latest verification evidence; - blockers and executable next actions. Do not persist private chain-of-thought. Record decisions, evidence and consequences only. ## Commands ```bash agentit continuity status --project . agentit continuity init "goal text" --project . agentit