goal

name	goal
description	Run a long-form goal with judge-verified completion.
disable-model-invocation	true

/goal

Declare a durable objective with a verifiable end state. The agent plans → acts → observes → iterates across many turns until every end-state criterion is confirmed — or you /goal pause or /goal clear.

Read First

1. .github/copilot-instructions.md — repo conventions, quality gates, anti-patterns.
2. docs/ — domain-specific docs relevant to the goal.

Mental Model

A goal has three required parts and one built-in safeguard:

Part	What it is	Good	Bad
Goal	A clear, actionable objective in the imperative.	«Migrate all v1 API calls to v2.»	«Make the API better.»
End state	The checklist. Observable, binary, machine-checkable.	«rg "/api/v1" src/ returns 0 matches; npm test exits 0.»	«It should work.»
Constraints	Boundaries: scope, style, safety, non-goals.	«Only edit src/ and tests/. Conventional Commits.»	(none — agent will drift)

The end state is the judge. The agent must verify each criterion independently before declaring done. This stops both premature completion («looks good to me!») and endless spinning («just one more refactor…»).

Invocation

Type /goal followed by your objective. The agent will ask for the three parts if any are missing, then run the loop until the end state is confirmed.

Minimal form:

/goal Migrate all deprecated v1 API calls to v2.
End state: rg "/api/v1" src/ tests/ returns 0 matches, npm test exits 0.
Constraints: only src/ and tests/, Conventional Commits.

Full form: use the layered template for complex, multi-hour goals.

Subcommands

Command	What it does
/goal status	Show progress, last verification results, remaining criteria.
/goal pause	Freeze the loop. State is preserved in session memory.
/goal resume	Continue from the last checkpoint.
/goal clear	Stop and reset. Use when you need to refine the goal itself.

The Loop

After receiving a goal, the agent runs this cycle until the end state is met or a stop condition fires:

1. Plan — break the goal into concrete, verifiable steps.
2. Act — execute one step. Edit files, run commands, gather evidence.
3. Verify — check EVERY end-state criterion. Run the exact verification commands. Report which pass and which fail.
4. Iterate — if any criterion fails, go back to plan with the gap in mind. If all pass, the goal is done.

The agent never declares victory on its own say-so. Every «done» must be backed by verification output.

Repo Integration

This skill inherits the repo's quality gates. When the end state includes test or lint criteria, use the canonical commands:

npm run quality:sanity    # markdownlint + eslint + tsc + vitest
npm run quality:full      # above + Playwright E2E
npm run test:coverage     # coverage report (≥90% threshold per file)

Branch-per-goal convention:

git switch -c goal/<short-name>

Commit before launching so git reset --hard is always a clean escape.

Stop Conditions

The loop stops when:

• All end-state criteria pass on a clean verification run — goal achieved.
• The user runs /goal pause or /goal clear.
• Hard iteration cap is reached (default: 20). The agent writes BLOCKERS.md with what's left and stops — no guessing.
• A destructive or irreversible action is required. The agent stops and asks.

Completion Criterion

The goal is done when every end-state criterion has been independently verified and the verification output is presented to the user. No criterion is «close enough.» Partial passes are failures.

Reference

• Layered template, anti-patterns, and examples
• Project instructions