| name | vouch-resume |
| description | Resume a failed or interrupted vouch run. Re-runs phases that are in failed/skipped state without disturbing completed phases. Useful after a crash, after topping up the budget, or after fixing whatever caused a stage to fail. |
You are the resume skill. Pick up exactly where state left off; don't redo work that already finished.
Argument
/vouch-resume <project_dir> [<run_id>]
If run_id is omitted, target the most recently modified run under $PROJECT_DIR/.vouch/runs/.
Step 1 — Locate the run
If run_id is given, require $PROJECT_DIR/.vouch/runs/<run_id>/state.json to exist. If not, error out and list available run ids.
Step 2 — Inspect state
python3 $PLUGIN_DIR/helpers/vouch_state.py read $RUN_DIR
Expected outcomes:
- All phases
completed → tell the user "nothing to resume" and stop.
- Run-level status
completed but a phase is failed → unusual; warn the user.
- One or more phases
failed or skipped → these are the resume targets.
Step 3 — Reset target phases
For each phase to resume, transition it from failed/skipped back to pending:
python3 $PLUGIN_DIR/helpers/vouch_state.py update $RUN_DIR --phase $P --status pending
Then transition the run-level status from failed back to running by re-initialising via update on the first reset phase (the helper auto-bumps run status on first running transition).
Step 4 — Hand off to /vouch-run
Tell the user explicitly:
Reset N phase(s) to pending: <list>
Run /vouch-run <project_dir> to continue.
Do NOT auto-invoke /vouch-run. The user might want to top up the budget or edit the pipeline first.
Edge cases
- Budget previously exhausted — tell the user the budget needs to be topped up before resuming, and show the current balance.
- Pipeline edited since the run started — if
pipeline.md has new stages that aren't in state.json, tell the user the safest path is a fresh run (don't try to splice).
- Adversary sidecar — leave existing
state.metadata.adversary entries in place; re-running a phase appends new entries, doesn't overwrite.
- Budget already debited for a failed phase — that's fine; we don't refund. Spend is monotone.
Rules
- Never delete result files. The user might want to inspect what the failed run produced.
- If a phase is
running (someone left it that way after a crash), treat it as failed for resume purposes.
- Refuse if state.json is locked by another process — the lock means a run is actively in flight.
