session-resume

session-resume — ≤500-byte state summary for new sessions

Role: First skill called by main session in codenook-shell.md after /clear or new-session bootstrap. Reads workspace state and returns a compact summary so main session can ask "Continue T-007 (implement)?".

CLI

resume.sh [--workspace <dir>] [--json]

Output (--json, ≤500 bytes UTF-8)

{
  "active_tasks": [
    {"task_id","plugin","phase","status","last_event_ts","one_liner"}
  ],
  "current_focus": "T-NNN" | null,
  "last_session_summary": "<≤300 char tail of history/sessions/latest.md>",
  "suggested_next": "Continue T-NNN (<phase>)?" | "N active tasks — pick one?" | "No active task, awaiting user input"
}

The output also exposes M1-compat keys (active_task, phase, iteration, summary, hitl_pending, next_suggested_action, last_action_ts, total_iterations) chosen from the most-recent task. These are part of the same 500-byte budget; under multi-task CJK pressure the chattiest of these are dropped first (see _resume.py:truncate_to_bytes).

Algorithm (M4)

1. Read .codenook/state.json → active_tasks[], current_focus.
2. For each tid in active_tasks, read .codenook/tasks/<tid>/state.json and project to the active-task shape above (last_event_ts = last history[] ts or created_at).
3. last_session_summary = tail(history/sessions/latest.md, 300 chars).
4. Pick suggested_next based on whether current_focus matches an in_progress task, else by count.
5. Truncate the JSON until UTF-8 size ≤500 bytes (drops in priority order: tail → one_liner → per-entry secondary fields → legacy chatty keys).

Backward compat (M1)

When .codenook/state.json does not exist, falls back to the legacy M1 behaviour: scan every tasks/*/state.json, exclude phase==done or status ∈ {done,cancelled}, sort by updated_at descending. The M1 bats suite (10 tests) continues to pass.

→ Design basis: implementation.md §3.4, architecture.md §3.1.4