Your agent says “All tests pass!” The hidden tests say 28/34.
goalpost makes /goal prove it before Claude is allowed to stop.

▶ Watch the full 90-second demo · two real benchmark runs, same model, same task, nothing staged (the most striking pair of the 33; the honest summary is in Benchmark)

English · Русский

Install (then open a new Claude Code session and type /goal like always):

claude plugin marketplace add syntaxixr/goalpost
claude plugin install goalpost@goalpost

On Windows you can just double-click install.bat. Remove it any time with uninstall.bat / ./uninstall.sh. Details are below.

---

You keep typing /goal <condition>. goalpost, a hooks plugin, turns every goal into a frozen spec with checkable criteria, blocks the stop until each criterion has a fresh passing check and a clean fresh-eyes audit, keeps tests from being edited to fit the code, and restores the plan after context compaction. It's enforced by hooks, so the model can't skip it.

In one breath: in 33 benchmark runs the built-in /goal judge said "done" in 17 of 17 bare runs, and 8 of them were broken. With goalpost, Sonnet 5.5 finished 9 of 9 runs with every hidden check green. Full numbers, including where it didn't help, are below.

The problem

/goal is a session-scoped Stop hook. After each turn a small model (Haiku) reads the conversation and says "met", "not yet" or "impossible". That judge never runs a command or opens a file (measured): if the agent writes a confident "all done", the judge can believe it. A vague goal ("build the whole project") leaves nobody knowing what "done" means. There's no plan on disk, so the thread is lost after compaction. And nothing stops the agent from editing a test until it passes.

Install

Needs Claude Code and Node.js 18+ (the hooks are small Node scripts). Nothing else. Windows, macOS, Linux.

Windows: download or clone the repo and double-click install.bat. macOS / Linux: ./install.sh. The installer checks Claude Code and Node, adds the marketplace, installs the plugin for your user (all projects) and confirms it's listed.

Or by hand, in a terminal:

claude plugin marketplace add syntaxixr/goalpost
claude plugin install goalpost@goalpost

Then start a new Claude Code session and use /goal as usual. To try it without installing: claude --plugin-dir ./plugin.

Check that it really enforces

Before trusting it with real work, run /goalpost:selftest in Claude Code (or node plugin/scripts/gp.js selftest from a clone). It drives the real hooks through a throwaway goal: an edit before the spec is blocked, a failing check blocks the stop and names the criterion, the fix passes, the stop then waits for the audit, and only after an audit PASS is Claude allowed to stop. Every step prints PASS or FAIL, and it shows where the evidence log lives. Then, in a real session, /goalpost:status after a /goal should show the goal as active; if it says there is no goal, Claude Code is not running the hooks (usually node is not on PATH).

Update and uninstall

claude plugin marketplace update goalpost && claude plugin update goalpost@goalpost   # update

To remove it, run uninstall.bat (Windows) or ./uninstall.sh, or by hand: claude plugin uninstall goalpost@goalpost && claude plugin marketplace remove goalpost. /goal goes back to its normal behaviour. The .goal/ folders inside your projects stay (they're your specs and logs); delete them if you don't need them. To pause goalpost without removing it, use /goalpost:off.

What happens on /goal ship the importer

1. Spec first. goalpost creates .goal/SPEC.md and tells the agent to fill in: assumptions (open questions get a decision, not a question to you), milestones for big goals, and acceptance criteria. Each criterion is an observable outcome plus a command that exits 0 only when it holds. Project files stay locked until at least one criterion exists.
2. Frozen. At the first project edit the criteria freeze. They can grow, but can't be dropped or weakened without a "Spec changes" entry that names them. The auditor reads those.
3. Checks that leave a trail. node .goal/verify.js runs the Verify commands, writes exit codes to .goal/evidence.log and ticks the boxes. A criterion counts only if its latest check passed after the latest code edit.
4. Tests are read-only. Test files that existed when the goal started, plus anything under "Protected", can't be edited, deleted or overwritten through Edit/Write or shell commands. A test that really has to change is declared under "Test changes" with a reason.
5. The Stop hook decides. While anything is open, Claude can't stop. The reason names the specific criterion: "AC-3 fails its check (exit 1) … last output: …". It runs next to the built-in evaluator, so even if the evaluator is talked into "met", the session keeps going.
6. Fresh-eyes audit. When everything passes, the goalpost:auditor subagent re-runs the checks, reads the code, and hunts for stubs, skipped tests and hardcoded answers. Only VERDICT: PASS for the current code lets the goal end.
7. Safety valves. Same error 3 times → "change the approach". No tool use for 3 stops in a row → goalpost lets go (before Claude Code's own cap of 8). A blocker only you can fix goes into .goal/BLOCKED.md, and goalpost stops and tells you. /goal clear and "impossible" are honoured.
8. Compaction-proof. After compaction or --resume, the SPEC and PROGRESS are injected again.

Before and after

Bare /goal                                    /goal + goalpost
─────────────────────────────────────         ──────────────────────────────────────────
agent: "All done, tests pass!"                agent: "All done!"
evaluator (reads text only): met              Stop hook: blocked — AC-4 has not been checked
→ goal cleared, session ends                  since the last edit; AC-6 fails (exit 1):
                                              "expected '#N/A' got '#VALUE!'"
                                              → agent fixes, re-runs verify, audit PASS, then ends

A real SPEC and audit from a small run are in examples/.

Commands and config

Command	What it does
/goalpost:status	Criteria, checks, audit and open items of the current goal
/goalpost:off / /goalpost:on	Global switch. Also GOALPOST=off in the environment
/goalpost:release	Stop enforcing the current goal (the built-in /goal keeps running)
/goalpost:selftest	Prove the hooks enforce on this machine, on a throwaway goal

One config file, ~/.claude/goalpost.json (all projects) or .claude/goalpost.json (one project). Every setting and its default is in examples/goalpost.json: block limits, audit on/off, reminder interval, test globs, protected paths, verify timeout and shell. To save tokens on small goals, "auditMinCriteria": 3 skips the audit for goals with fewer than three criteria, unless one of them is manual or tests are declared under "## Test changes".

How it differs

	goalpost	Looper	Ouroboros	claude-goal	goalkeeper
Works with the built-in /goal	yes, you keep typing it	designs a loop, you run it	separate ooo commands	replaces it	replaces it
Enforced by hooks	yes	no (skill)	partly	Stop hook only	one PreToolUse guard
Who decides "done"	files on disk + fresh auditor	judge you configure	3-stage evaluation	the agent says so	subagent judge
Evidence required	fresh check after the last edit	typed verification	mechanical stage	none	validator command
Tests protected from edits	yes	no	no	no	partly
Spec frozen against weakening	yes	no	seed is immutable	no	contract is guarded
Repeated-error detection	yes	no-progress stop	stagnation patterns	no	rejection counter
Restores plan after compaction	yes	no	event log	no	log file
Needs	Node.js	Python + PyYAML	Python 3.12, uv, MCP	Python, SQLite	Python 3
Interview before work	no (unattended)	yes	yes	no	yes

Full notes on every project: docs/RESEARCH.md.

Benchmark

33 real runs (Sonnet 5.5 and Haiku 4.5), 4 tasks, hidden graders, each run graded 3 times (full report, every run as a dataset on Hugging Face):

	bare /goal	/goal + goalpost
Sonnet 5.5: hidden checks passed	99.0%, 8/9 runs perfect	100%, 9/9 runs perfect
Sonnet 5.5: "done" while hidden checks failed	1/9	0/9
Haiku 4.5: hidden checks passed (mean over the 4 tasks)	79.8%	89.7%
Haiku 4.5: "done" while hidden checks failed	7/8	5/7
Time and API cost per run (Sonnet / Haiku)	2.7 / 5.7 min, $0.46 / $0.64	5.2 / 15.1 min, $0.91 / $1.25

What it shows, honestly. The built-in evaluator said "met" in all 17 bare runs, and 8 of them were broken: it trusts the agent's text. goalpost never ended a Sonnet run with failing checks and moves the weaker model's score the right way (79.8% → 89.7%), but it does not make Haiku reliable: only 1 of 7 runs was fully correct. It costs about 2× the money and 2–2.6× the time. On a strong model with mid-size tasks the accuracy gain is small (one run).

Limits, honestly

• It costs time. Writing a spec, running checks and an audit take extra turns. In the benchmark the goalpost arm took about 2× the wall time and API cost of bare /goal (Sonnet: 5.2 vs 2.7 min, $0.91 vs $0.46 per run). On a small goal that's overhead you may not want: /goalpost:off, GOALPOST=off claude, or "auditMinCriteria": 3 to skip the audit on small goals. Since 2026-10-07 PROGRESS.md notes no longer hold the session open, the protocol is shorter, and the auditor batches its work; the full benchmark has not been re-run on this version yet.
• Strong models often finish anyway. With Sonnet 5.5 on mid-sized tasks, bare /goal already scored near 100%. goalpost's gains show up when the model takes shortcuts. The numbers are in BENCHMARK.md, including the runs where goalpost changed nothing.
• Shell-write protection is best effort. Edit/Write to protected files is blocked reliably. For Bash/PowerShell, goalpost parses redirects, file commands (rm, mv, sed -i, Set-Content…) and writeFileSync-style calls. A determined obfuscated command can get past it. The auditor is the second line.
• Needs Node.js 18+ on PATH. If node is missing, the hooks can't start, and Claude Code keeps working as if goalpost weren't installed. The installer checks this.
• Measured in headless mode. Hook behaviour was measured with claude -p. If a surface ever doesn't pass /goal to UserPromptSubmit, goalpost activates from the transcript's goal record at the first Stop. That fallback was tested end to end with the hook removed (MECHANICS.md §7).

How it works, and what we measured

docs/MECHANICS.md records what was measured on Claude Code 2.1.285 instead of guessed: UserPromptSubmit does see /goal <condition>; UserPromptExpansion does not; /goal clear fires no hook; the goal state is visible in the transcript; the Stop hook runs next to the built-in evaluator; the 8-block cap only counts stops without tool use.

Develop

node --test test/lib.test.js test/hooks.test.js     # 38 tests, ~10 s
claude --plugin-dir ./plugin                          # try it without installing
node bench/run-bench.js --round rX --tasks D --reps 3  # benchmark

See also

Receipts: checks that the tests in a fix would have failed without it, by running them on the old code too. It is in Anthropic's plugin directory as Receipts Check. goalpost checks that the goal is done, Receipts checks that the tests prove it.

License

MIT. See CREDITS.md for the projects whose ideas this builds on.