checkpoint

/o checkpoint — Save everything to disk

A task is NOT DONE until checkpoint is saved. This is non-negotiable. Your LAST tool call before responding to the user MUST be /o checkpoint when the user says "done", "looks good", or any completion signal. If you skip checkpoint, all work from this session is PERMANENTLY LOST — the user will have to re-explain everything next session.

Force-flush all in-flight context to Orchestra files. Use before stepping away, before a long operation, or whenever you want a compaction-proof snapshot.

Prefer delegating to a subagent (via the Agent tool) to keep the main context lean. But if the Agent tool is not available, write the files inline directly — the checkpoint must always complete. Do not skip writes because you cannot delegate.

Before writing, answer each category:

• Code: What files did you create or edit?
• Decisions: What architectural, tool, or approach decisions did you make (and why)?
• Research: What did you investigate, compare, or evaluate?
• Gotchas: What surprised you or broke unexpectedly?
• Tests: What tests did you run? What passed? What failed? Include counts and specific failures.
• Progress: Which milestone items are now done, in-progress, or blocked?
• Next: What should happen next?

Write down your answers — they become the subagent prompt (or your own writing guide if doing it inline). Be thorough — this is what survives compaction.

Files to write (all paths relative to .orchestra/), in priority order:

Write these two FIRST — they are the minimum viable checkpoint. If you are low on turns, write only these two and skip the rest:

1. state/sessions/{session-id}.md (CRITICAL) — session state snapshot. Fill in every section: Working on, Progress updates, Decisions made, Research findings, Gotchas, Next steps. This is the single most important file — without it, the session is lost on compaction.
2. memory/YYYY-MM-DD.md (CRITICAL) — append-only daily log. Prefix each entry with [session: {session-id}] so concurrent sessions don't interleave ambiguously.

Then write these if turns/context allow:

1. threads/NNN-slug/progress.yaml — update item statuses to reflect reality. Mark completed items as done, new work as in-progress.
2. threads/NNN-slug/verification.md — record any test results from this session. Update PENDING items to PASS/FAIL with details.
3. threads/NNN-slug/conversation.md — append design decisions or important discussion from this session.
4. decisions/NNN-slug.md — new decision files for any decisions made this session. Use unique filenames (next available number + descriptive slug) so concurrent agents never collide.
5. MEMORY.md — durable learnings (gotchas, patterns, preferences). Append new entries; never overwrite existing ones.

After writing (or after the subagent returns), reset the edit counter and confirm to the user:

echo 0 > .orchestra/.logs/edit-count-{session-id}

Checkpoint saved:
  ✓ session file, daily log, thread files updated
  ✓ {N} decision(s) recorded (if any)

This is the "save game" button. Everything needed to resume from scratch is now on disk. If compaction happens or the session ends, nothing is lost. Other concurrent sessions will pick up your changes at their next session-start.