goal-file-autonomous-run-until-evidencelisted
Install: claude install-skill sblattj/agent-chandlery
# goal-file-autonomous-run-until-evidence
One file (`.chandlery/goal.json`) holds the goal. The `chandlery goal` CLI is the only writer, so the gates hold: an action can't close without evidence, the goal can't finish with open actions, and the budget stops the loop.
## Do this
```bash
chandlery goal set "Make the parser handle quoted flags" \
--success "npm test passes" --success "no change to the public API" \
--constraint "do not touch src/legacy" --max-minutes 240
chandlery goal plan add a1 "write a failing test for quoted flags"
chandlery goal plan add a2 "fix the tokenizer"
chandlery goal status
```
A long handoff goes in a file: `chandlery goal set --brief-file handoff.md` (kept verbatim; the label comes from its first line).
While working:
```bash
chandlery goal act a1 in_progress
chandlery goal act a1 done --claim "test fails on main" --evidence "node --test: 1 fail, expected 'a b'" --verdict pass
chandlery goal act a2 blocked --reason "needs the vendor's spec, not in the repo"
chandlery goal complete --summary "quoted flags parse" --evidence "npm test: 48 pass, 0 fail"
chandlery goal block "the deploy needs a production token I don't have"
```
## Keep the agent going
- **Claude Code:** `chandlery goal hook install` adds a Stop hook to `.claude/settings.json`. Each time the agent tries to stop, the hook counts a turn and, while the goal is active, sends it back with the continuation prompt. It goes quiet once the goal is achieved, blocked or paused.
- **A