checkpoint

name	checkpoint
description	Write a committed handoff checkpoint so a fresh session can resume with zero loss. Use when crossing the 40% context line, on every stage transition, and at session end. Writes .trace/checkpoints/<plan>-<seq>.md while the agent is still sharp.

A checkpoint is institutional memory. Write it while still under 40% — a handoff authored by a degraded agent is worthless.

Triggers (write a checkpoint on ALL of these)

1. Crossing the 40% context line (bin/ctx.sh exited 2).
2. Every lifecycle stage transition.
3. Session end.

Procedure

1. Pick the next sequence number for this plan (look at existing .trace/checkpoints/<plan>-*.md).
2. Write .trace/checkpoints/<plan>-<seq>.md:

---
plan: <plan>
seq: <NNN>
stage: <current-stage>
ctx_pct_at_checkpoint: <n>
prev: <plan>-<seq-1 or none>
---
## Done
- <what is verifiably complete — tests passing, slices green>
## Now
- <what is mid-flight, exact state>
## Next (resume here)
- <the very first action the next session should take>
## Decisions
- <choices made + where the rationale lives>
## Open questions / blockers
- <or "none">

3. Update the active exec-plan's Now/Next to match.
4. Drop any odor in the backlog. Noticed a smell outside this slice's scope? Don't fix it inline (that breaks slice discipline) and don't lose it — append it to .trace/garden-backlog.md (- [ ] <date> | <file:line> | <smell> | <low|med|high> | <note>). See the garden skill.
5. bin/trace.sh <agent> <stage> checkpoint seq=<NNN> ctx_pct=<n>.
6. These files are committed (unlike .trace/runtime/). Commit them.

Each committed checkpoint is a gardening "visit": writing this one may tip harness.sh status to garden: DUE (≥5 since the last sweep). The main agent acts on that at the next plan boundary — see garden.

Quality bar

A good checkpoint lets a cold agent resume in one read. If "Next" is vague, the checkpoint failed.