tmdgusya/roach-pi/agentic-goal

Agentic Goal

Work only through the durable /goal runtime.

Core Rules

1. Start by reading /goal status when a goal may be active.
2. Work only on the active goal or active subgoal shown by /goal status.
3. When /goal <request> is used as a new entrypoint, triage the request first: answer simple investigation/question/explanation requests like normal prompts, but route complex, ambiguous, or verifier-worthy work into deep clarification before durable goal activation.
4. When /goal is invoked without a specific target, continue the entire active goal across subgoals until the goal itself receives verifier PASS.
5. Track immediate work with todoread and todowrite.
6. Add evidence with /goal evidence <targetId> <evidence> before requesting completion.
7. Never claim a goal or subgoal is complete until the verifier subagent returns PASS.
8. If the verifier returns FAIL, continue working on the blockers and gather new evidence.
9. If a subgoal verifier returns PASS, continue to the next runtime-provided subgoal; stop only after the active goal itself receives PASS or user intervention is required.

New Request Triage

For /goal <request>, silently decide whether the request needs durable goal runtime:

• Simple investigation, lookup, explanation, or read-only question: answer directly as a normal prompt.
• Complex implementation, multi-step work, ambiguous scope, or work needing completion evidence/verifier PASS: begin deep clarification and produce a Goal Contract before activation.
• If uncertain, prefer clarification for complex or ambiguous work.

Who Implements

For goals started via autostart (flagged with the worker→validator gate), the runtime implements each subgoal through an isolated worker→validator loop: the durable runtime dispatches an isolated worker subagent to implement the subgoal and an information-isolated validator subagent to judge it, then records the receipt and completion itself. In that mode, do not implement subgoals yourself; advance the runtime by running /goal (no arguments) until the goal completes or the runtime escalates.

For small tasks, prefer the manual /goal create → activate → complete path instead of autostart.

Workflow

1. Inspect /goal status.
2. Identify the active goal/subgoal objective, success criteria, constraints, evidence required, and blockers.
3. Create or update todos for the immediate work.
4. Implement the required changes.
5. Record evidence with /goal evidence.
6. Request completion with /goal complete <targetId>.
7. Follow the verifier outcome:

• Subgoal PASS: continue with the next runtime-provided subgoal.
• Goal verifier PASS: for review-gated goals the runtime opens the security/qa review panel; the goal completes only when security and qa both PASS, and any review FAIL recycles into fix subgoals driven by the worker→validator loop.
• FAIL: address blockers, record new evidence, and request completion again.

Durable State Handoff

The /goal runtime is canonical. Do not use external planning documents as the source of truth. Do not route to legacy workflow skills as user-facing next steps.