From task to checked result

Replume runs beside the agent. It keeps a local record, checks key actions while the work happens, and checks the result before calling it done.

1. 01

   Start in the repository

   replume init writes project-local hook settings. Give your agent a task as usual; in auto mode the first prompt establishes the task and its working scope.

   See: .claude/settings.json (or your agent's hook config) and .replume/contract.yaml.

2. 02

   Follow each step

   The hooks record tool calls in a local ledger. Replume can warn or steer when work moves beyond the task.

   See: .replume/events.jsonl or run replume watch.

3. 03

   Protect the checks

   Protected files such as tests and the test command cannot be quietly weakened by an agent action the hook can intercept. Read the reason when a write is blocked.

   See: the BLOCK line from replume watch and its ledger entry.

4. 04

   Check before the agent stops

   The Stop gate runs the test command. It compares failures with a baseline from before the change, so a pre-existing failure stays visible. When the agent introduces a failing test, the gate can send it back to fix the change; retries are bounded.

   See: .replume/baseline.json, TEST and RETRY lines in replume watch.

5. 05

   Read the outcome

   PASS means required checks passed in this run. FAIL means a required check failed. INCOMPLETE means Replume could not verify enough; it never counts as PASS. Read what was checked and what remains unknown.

   See: .replume/reports/ or run replume report.

6. 06

   Share when you choose

   replume share previews redacted report sections before creating an unlisted link. replume github comment --repo OWNER/REPO --pr N previews a PR comment before posting. Neither command runs as part of normal verification.

   See: the preview printed by either command; use --dry-run to inspect without sending.

For setup details, read the Start guide. For the broader direction, see the roadmap.