create-goals

name	create-goals
description	Turn a freeform idea, brief, or discussion into a durable sequential goal plan under .agent/goals/ for later execution by slop-janitor goals run.

Create Goals

Create a minimal, ordered goal plan from the current conversation. This skill is interactive planning only. Do not implement code, run tests, or start executing the goals.

Goal

Turn the user's idea into a durable plan under .agent/goals/<id-slug>/:

• brief.md
• goals.json
• ledger.jsonl

Also point .agent/goals/active at the plan directory so the user can run or invoke the approved plan without retyping its path.

The artifact is the source of truth. Codex thread goals are only the execution context used later by slop-janitor goals run.

Workflow

1. Distill the user's objective, constraints, non-goals, assumptions, risks, stop conditions, and validation expectations.
2. Ask concise clarifying questions only when the next goal sequence would be materially wrong without the answer. If a question can be answered by reading the repo, read the repo instead.
3. Create .agent/goals/<id-slug>/.
4. Write brief.md as stable human-readable context.
5. Write goals.json with an ordered goals array.
6. Initialize ledger.jsonl with a plan_created event.
7. Create or update .agent/goals/active as a symlink to <id-slug>.
8. Stop and ask the user to approve or edit the plan before execution.

Artifact Model

goals.json must be a JSON object:

{
  "id": "2026-05-03-my-goal-plan",
  "title": "My goal plan",
  "status": "active",
  "active_goal_id": "goal-1",
  "created_at": "2026-05-03T12:00:00Z",
  "updated_at": "2026-05-03T12:00:00Z",
  "goals": [
    {
      "id": "goal-1",
      "title": "First goal",
      "objective": "Concrete objective for Codex to pursue.",
      "rationale": "Why this goal comes first.",
      "scope": ["Included work"],
      "non_goals": ["Explicitly excluded work"],
      "stop_condition": "Concrete condition that ends this goal.",
      "acceptance_criteria": ["Observable success criterion"],
      "validation": ["Exact command, artifact, or inspection expected"],
      "depends_on": [],
      "status": "ready",
      "result_summary": null,
      "evidence": [],
      "risks": ["Known risk"],
      "assumptions": ["Planning assumption"]
    }
  ]
}

Allowed plan statuses: active, blocked, completed, abandoned.

Allowed goal statuses: ready, active, blocked, completed, failed, skipped.

Set active_goal_id to the first executable goal. Keep dependencies simple: prefer ordered goals and use depends_on only for real blockers.

Use a relative symlink for the active pointer:

ln -sfn <id-slug> .agent/goals/active

If the platform cannot create symlinks, stop and tell the user to pass the explicit plan path to slop-janitor goals run .agent/goals/<id-slug>.

Brief Shape

Keep brief.md compact:

• title
• objective
• constraints
• non-goals
• ordered goal summary
• acceptance and validation expectations
• assumptions and risks

Anti-Patterns

• executing the plan while creating it
• invoking complete-goals before the user approves the artifacts
• hiding important criteria only in prose
• creating a DAG or pipeline engine in JSON
• making every goal depend on every previous goal without a real dependency
• dumping transcript history into artifacts