pi-goal-moonland

What is a goal?

A normal prompt tells the agent what to do next. A goal says what must become true:

/goal Fix every failing test in test/auth, rerun npm test until it exits 0,
      and stop and report if a failure needs changes outside test/ or src/auth

Once a goal is active, the agent keeps working. Each time a run ends with the goal still active, pi-goal queues a hidden continuation and the agent picks the work back up — slice by slice, across as many runs as it takes. The loop ends only when the agent marks the goal complete (with evidence) or blocked (with a concrete reason), or when you stop it.

Install

Requires pi ≥ 0.87 — the continuation loop runs on pi's actionable agent_before_settle boundary; older hosts get an error notification instead.

pi install git:github.com/huahuadeliaoliao/pi-goal       # from git
pi install npm:pi-goal-moonland                          # from npm

On npm the package is pi-goal-moonland — the bare pi-goal name was already taken.

Or ad-hoc, without installing: pi -e /path/to/pi-goal (the directory or index.ts both work).

Manual copy still works too:

git clone https://github.com/huahuadeliaoliao/pi-goal.git
cp -r pi-goal ~/.pi/agent/extensions/goal

Usage

The agent can also start goals itself via its goal tool. Say "help me write a goal" and the bundled write-goal skill walks the agent through drafting a verifiable completion contract — you approve the wording in conversation, and the goal starts without a second prompt.

Flags:

• --goal "<objective>" — start a TUI/RPC session with a goal already running
• --goal-confirm — ask for your approval before an agent-created goal starts
• --goal-progress-interval <N> — append a persistent goal-progress message every N goal turns (default 20; 0 disables)

Headless: pi -p "/goal <objective>" runs the goal to completion, then exits.

Security note: a goal is an unattended loop that edits files and runs commands, and agent-created goals start immediately by default. If you run pi-goal against untrusted content or simply want a human checkpoint, start pi with --goal-confirm.

How it works

• Continuation loop — at pi's agent_before_settle boundary (after retries, auto-compaction and your queued messages are done), if the goal is still active, a hidden continuation carrying the objective, the live turn count and the audit rules is appended and pi makes one more model call. No timers. An exhausted error pauses the goal.
• Prompt-cache friendly — every goal message the model sees is persistent and appended at the end of the conversation: the goal reminder at start, status notes (paused/cancelled), continuations, and periodic goal-progress pings. Nothing is injected per call, edited or removed, so creating, pausing, resuming, completing, cancelling or replacing a goal only extends the cached prefix — including codex response chaining. Mid-run notes are queued for the turn boundary, never steered in.
• Compaction-proof — goal state lives in session custom entries (never enters LLM context). After a compaction the goal reminder is appended again (free: compaction already resets the cache), and every continuation repeats the objective anyway.
• Branch-aware — goal state is read from the current session branch, so /tree navigation and forks see the goal as it was at that point.
• Honest endings — the prompt wording (adapted from kimi-code's iterated audit language) pushes the agent to verify against current evidence before complete, and to only declare blocked after the same blocker repeats for 3 consecutive goal turns.
• Interrupts — Esc pauses the goal; resuming a session or switching branches restores an active goal as paused; explicit /goal resume continues it.
• Visibility — a footer status (goal: active (turn N)), inline lifecycle markers (started/paused/complete/blocked), hidden continuation messages.

Development

npm install
npm test          # node:test with a mock ExtensionAPI — no real pi process
npm run check     # tsc --noEmit
npm run lint      # biome check

The extension is plain TypeScript loaded by pi via jiti — no build step. @earendil-works/pi-* and typebox are peer dependencies (pi bundles them for extensions); they are pinned in devDependencies only for tests and type-checking.

Design provenance

The design is ported from kimi-code's goal feature (packages/agent-core-v2/src/features/goal, MIT), with oh-my-pi (MIT) as a cross-reference. Deliberately dropped: token/turn/wall-clock budgets and the /goal next queue — the goal is the goal. Re-add them if real usage asks.

License

MIT © 2026 huahuadeliaoliao