/goal Deterministic Gate

/goal keeps Claude working until a small Haiku checker reads the transcript and decides the condition holds. Its load-bearing limit (verified, 2.1.170): the checker runs no tools and reads no files — it can only confirm that text claiming success appears in the conversation. A maker that hallucinates "✓ all tests passed", or runs the wrong command, sails straight through.

The fix Anthropic's docs themselves point to: push the real, deterministic check into a Stop-hook script, and let /goal adjudicate only the semantic residue. The script actually runs the command and exit-2-blocks the stop (feeding the failure back as guidance) until it genuinely passes.

The pattern (two layers)

1. /goal <semantic condition> — drives continuation and handles the fuzzy part (e.g. "the auth refactor is complete and the new endpoints are documented").
2. A Stop hook running goal-stop-gate.sh — the objective backstop: runs your verify command (pytest -q, npm test, tsc --noEmit, cargo build) and:
   • exit 0 → allow the stop (genuinely done);
   • exit 2 → block the stop; stderr (the failing output) becomes Claude's next-turn guidance.

This is the §"force unforgeable evidence" rule from loop-designer/references/verifier-discipline.md made mechanical — the gate reads the real exit code, not the agent's narration of it.

Install (template: templates/goal-stop-gate.sh)

Copy the template, set your verify command, make it executable, and register it as a Stop hook in .claude/settings.json (project) — keep it project-scoped so it only gates the loop you mean:

{ "hooks": { "Stop": [ { "hooks": [
  { "type": "command", "command": "CLAUDE_GOAL_VERIFY_CMD='pytest -q' .claude/hooks/goal-stop-gate.sh", "timeout": 120000 }
] } ] } }

The script respects stop_hook_active (Claude Code's loop-guard) so it blocks once per stop attempt, never infinitely — and fails open (allows the stop) if the verify command can't run, so a broken gate never traps the session.

When to use which

• /goal alone — the condition is fully demonstrable in the transcript (the agent pastes the raw pytest output) and you trust it to. Cheapest.
• /goal + this gate — the condition must be machine-true, not transcript-claimed. Use for anything that ships, merges, or runs unattended.
• agent-handler Stop hook (experimental) — when the check needs a subagent to read files/run several commands and reason, not a single deterministic command.

Companions: loop-designer (the /goal gotcha + non-rotting gate), evaluator-optimizer (semantic maker≠checker), long-running-harness (features.json verify-before-flip), quality-gates.