follow-goal

Skill: Follow Goal

Run a long, durable objective across many turns toward a verifiable stop condition — instead of stopping after one normal exchange.

Requires session storage + artifact.

1. Persist — Save goal.md to session storage. This is the durable source of truth, readable across turns.
2. Mark as artifact — After saving, call setArtifacts to mark the file as a plan artifact so the user can see live goal state in the UI.

Do not write goal.md into the workspace. Do not rely on the artifact alone — artifacts are UI-only and cannot be read back by tools.

A goal has four parts. Capture all four before starting work:

1. Objective — what to achieve, in one sentence.
2. Stop condition — an observable, verifiable signal that "done" has been reached (a command exits 0, a file matches a spec, tests pass, parity check passes, etc.).
3. Validation — the concrete commands or artifacts that prove progress (test command, build command, lint, diff against reference, review or rubber duck skills).
4. Constraints — what NOT to change, scope boundaries, files/areas off-limits.

If any of these are missing or vague, ask the user before saving the goal. A goal without a verifiable stop condition is not a goal — it's a wish.

State

The goal lives in session storage as goal.md — a single markdown file with this shape:

---
status: active        # active | paused | done | cleared
created: 2026-05-16
checkpoints: 0
max_checkpoints: 20   # hard safety budget
---

# Objective
<one sentence>

# Stop Condition
<verifiable end state>

# Validation
- `<command or check>`
- ...

# Constraints
- <what not to change>

# Progress Log
<append a one-line checkpoint after each validation pass>

max_checkpoints is a safety budget. The agent MUST stop and report when it reaches this number even if the stop condition is not met, and ask the user how to proceed.

Sub-commands

The goal skill dispatches to this skill. Behavior depends on the argument:

The Loop

When status: active, on every turn while there is work to do:

1. Read goal.md (objective, stop condition, validation, progress log).
2. Check stop condition. Run the validation commands. If the stop condition is met → set status: done, append a final checkpoint, report success, exit the loop.
3. Plan one checkpoint of work — the next small, scoped step that moves toward the stop condition. Keep checkpoints small enough that validation runs between each.
4. Do the work (edits, commands).
5. Validate — run the validation commands.
6. Append a one-line entry to the progress log in goal.md with what was done and the validation result.
7. Increment checkpoints in the frontmatter.
8. If checkpoints >= max_checkpoints → pause, report, ask the user how to proceed.
9. If status was changed externally to paused or cleared → stop. Otherwise continue the loop.

The progress log entries should be short and machine-skimmable, e.g.:

- 2026-05-16 14:02 — Migrated `auth/` to new API. Tests: 184/184 pass.
- 2026-05-16 14:09 — Migrated `billing/`. Tests: 184/184 pass. Lint clean.

Discipline

• Never continue past a failed validation without diagnosing the failure first. A failed validation is a checkpoint event — log it, fix the cause, then re-validate.
• Never silently expand scope. If the work requires touching something in Constraints, pause and surface it.
• Never invent a stop condition. If the user can't articulate one, the goal is not ready — push back.
• Keep the progress log truthful. Do not log "tests pass" without running them.
• One active goal at a time. If the user sets a new goal while one is active, ask whether to replace, pause, or clear the existing one first.

When to use vs. when not to

Use a goal when:

• The work has a clear, verifiable end state (a test suite green, a migration complete, a build artifact matches a spec).
• The work is too big for one turn but small enough that progress is measurable.
• The user wants to step away while the agent iterates.

Do not use a goal for:

• A loose grab-bag of unrelated tasks. Use a todo list instead.
• Work where "done" is subjective (design polish, prose). Use normal review skills instead.
• Anything destructive without rollback (production deploys, force pushes). The loop is not for irreversible operations.