Agent /goal Loop

What /goal is

/goal is a slash command that turns an agent prompt into a persistent agent looping plan → act → test → review → iterate until a stop condition is met, the user pauses, or the token budget runs out. Internally called the "Ralph loop."

Agents with the /goal feature right now: Codex, Claude Code, and Hermes Agent.

Key difference from a normal prompt: when a turn ends but the goal isn't met, the agent auto-continues instead of waiting for input.

Lifecycle states: pursuing, paused, achieved, unmet, budget-limited.

When monitoring a running /goal, every check should include a one-line update to the user: what the agent is doing and whether it is on track. Keep it extremely concise.

Not: a budget command, a safety boundary, "run forever", or a replacement for /plan. It's a contract enforcer with a verification loop.

Requirements

• An agent with the /goal feature — right now: Codex, Claude Code, or Hermes Agent
• The goals feature enabled in the agent's config
• Subscription auth — API-key auth does not work. A pro-tier plan is the realistic minimum for long runs.

When to Use it

Use only when all three are true:

1. Task is >30 min of mechanical work.
2. There's a verifiable stop condition (tests pass, coverage hit, eval ≥ X, build green).
3. Repo is agent-ready (working build, decent tests, AGENTS.md present).

Fits: migrations, coverage lifts, TDD feature builds, refactors with contract tests, prompt/eval optimization, deploy retry loops, bug-repro-then-fix.

Bad fits: exploratory work, vague "improve this", anything without a "done" definition, prod credentials, destructive shared-infra ops.

The 5-part contract (every goal needs this)

1. Objective — one sentence, one concrete outcome.
2. Constraints — what must NOT change (public API, files, libs, conventions).
3. Validation command — the exact shell command that proves progress (pytest -q, pnpm test, etc.).
4. Stop condition — verifiable: "Stop when X passes" OR "when further changes need human/product input."
5. Documentation — one sentence instructing the agent to write concise, targeted docs for every change, either creating new .md files or updating existing ones.

Plus: tell the agent what to read first, ask it to work in checkpoints with a short progress log.

Writing a goal (the core deliverable)

When the user wants a quick /goal instruction, produce a structured markdown block with one line per contract item (proper newlines, not flowing prose). Do not prefix the output with /goal — the user adds the slash command themselves in the composer. Emit only the contract body. Template:

**Objective:** <one-sentence objective>
**Read first:** <files/PLAN.md/issue>
**Constraints:** <what not to change, libs, conventions>
**Validate:** `<exact command>` after each change
**Document:** Write concise, targeted documentation for all changes — create new `.md` files or update existing docs as needed.
**Checkpoints:** work in checkpoints and log progress briefly
**Stop when:** <verifiable condition>, OR when further changes require human/product input

Example (migration)

**Objective:** Migrate this project from Pydantic v1 to v2.
**Read first:** pyproject.toml, src/, tests/
**Constraints:** no public API changes; keep imports backwards-compatible via shims if needed; no new dependencies
**Validate:** `pytest -q` after each change
**Checkpoints:** work in checkpoints; log progress briefly
**Stop when:** full suite passes with zero deprecation warnings, OR when a change requires architecture decisions

Example (coverage lift)

**Objective:** Raise coverage in src/auth/ from ~38% to ≥75%.
**Read first:** src/auth/, tests/auth/, AGENTS.md
**Constraints:** no new deps; mirror existing test style; do not modify production code unless strictly required for testability
**Validate:** `pytest --cov=src/auth --cov-report=term-missing`
**Checkpoints:** work in checkpoints; log coverage delta each one
**Stop when:** coverage ≥75% AND all tests pass, OR when uncovered code needs design changes

Writing rules

• One objective, one stop condition. Not a backlog.
• Documentation is mandatory. Every /goal prompt must include a single sentence committing the agent to concise, targeted docs — new .md files or focused updates to existing docs.
• Never instruct the agent to create new ADRs — ADRs require the user's explicit approval, so goal prompts must not pre-approve or encourage them.
• Forbid reward-hacking explicitly: "Do not delete, skip, weaken, or narrow tests to make the goal pass." Otherwise the agent may game the stop condition.
• 4,000-char limit on the objective. If longer, put detail in a file (PLAN.md/GOAL_BRIEF.md) and make the goal point to it — keep the goal itself compact.
• Use literal strings for paths, commands, issue numbers — exact.
• Forbid scope creep explicitly: "Do not refactor unrelated code. Do not add dependencies."
• Tell the agent when to pause: "If , pause and ask before proceeding."
• Short, vague goals burn tokens for no extra value vs. a normal prompt.

Meta-prompting trick (highest-leverage)

Hand-written goals under-specify. Ask a second AI session (Claude with the codebase loaded, ChatGPT with project connected, or a separate agent thread in the same dir) to: (1) inspect the codebase, (2) surface hidden assumptions/constraints/edge cases, (3) emit a structured /goal markdown block using the 4-part contract. Paste that into the agent. Order-of-magnitude better runs.

Claude Code cmux note: after Claude finishes, it may prefill a predicted next user message; that draft is Claude, not the user speaking.

Self-goal setting

The agent can now write and set its own goal natively (the create_goal tool). Instead of crafting the contract yourself, give it your high-level intent and tell it to set the goal: "Inspect this repo, then write yourself a /goal with a verifiable stop condition and pursue it." It's the meta-prompting trick done inline — the agent turns your intent into the contract. Still give it the same raw materials (files to read, constraints, the validation command) so the goal it writes is grounded. Add: "ask clarifying questions before committing if the intent is underspecified" — catches ambiguity up front and prevents the self-set goal from drifting.

Launching

1. cd <repo> (goals run scoped to the working directory).
2. Launch the agent bare (opens the TUI). Not exec/headless mode — /goal is a TUI slash command only.
3. Sign in with subscription auth (not an API key).
4. Type /goal <your contract> in the composer, Enter.
5. Walk away.

Controlling a running goal

Resuming across sessions: goal state is persisted server-side. cd back into the repo, launch the agent, /goal for status, /goal resume.

Budget-limited state: the agent doesn't stop abruptly — it summarizes, notes what's left, saves state. /goal resume works after budget refresh or upgrade.

When a goal drifts

• Minor drift: just type a correction in the composer (auto-pauses, folds it in, resumes).
• Loose objective: /goal pause, read status, then /goal <tighter version> — replaces the contract. Don't pile instructions on a vague goal.
• Bad mess: /goal clear, git status or git stash, rewrite with the meta-prompting trick, restart.

Don't let a drifting goal keep running "to see where it goes." Tokens burn, diffs compound.

Operational tips

• Inspect status periodically with bare /goal.
• Always review the diff before merging — long autonomy means more code to validate, not less. Human oversight becomes more critical, not optional.
• Keep approvals/sandboxing tight; default permissions are correct.
• First run: pick a 30-min scoped task so you learn how /goal actually stops before trusting it overnight.
• Bake recurring policy into AGENTS.md so every goal inherits it without restating: adversarial self-review before declaring done, an extra QA pass even when tests pass, and the standard validation command. Saves repeating it in each goal paragraph.

Troubleshooting

Mental model

/goal is a contract enforcer with a verification loop, not a "run forever" button. The shift: stop writing prompts, start writing specifications with stop conditions. Spend the time upfront defining "done"; the run takes care of itself.

Limitations

• Adapted from davidondrej/skills; verify local paths, tools, credentials, and agent features before acting.
• For commands, remote access, scheduling, browser automation, or file-changing workflows, get explicit user approval and confirm the target environment first.