loop-orchestrator

Loop Orchestrator

Move one canonical contract between focused skills. Own sequencing and handoff state only.

Installed dependencies

Install methodology-selector, goal-engineering, loop-readiness-score, methodology-skill-creator, and loop-run-record beside this skill. The orchestrator coordinates those siblings and is not a standalone replacement for them. Stop with the missing skill names when the set is incomplete.

Inputs

• intent: loop-goal, methodology-skill, or record-run;
• Task, Audience, Context, Output Format, Validation, and Human Review / Stop;
• the control envelope and explicit applicability declarations when intent: loop-goal;
• skill_name and skill_summary when intent: methodology-skill (both required); optional invocation aliases when supplied.

Read references/handoff-contract.md.

Routing

1. Invoke methodology-selector.

2. Preserve its canonical eight-field contract exactly.

3. If selection returns none, stop and return the missing fact or negative-fit evidence. Do not call another component.

4. Route by intent:

   • loop-goal: pass the canonical contract plus the caller's control envelope to goal-engineering, then pass both goal_input and goal_result to loop-readiness-score.
   • methodology-skill: if skill_name or skill_summary is missing, stop and return the missing field; do not call the creator. Otherwise pass the canonical contract, selection metadata, skill_name, skill_summary, and optional invocation to methodology-skill-creator.
   • record-run: first verify the exact goal/readiness handoff. Preserve blocked or supervised; only then pass complete external observations to loop-run-record.

5. Bind the readiness result to the exact goal input, rendered goal, answers, risk, and applicability envelope used by the checker.

6. Return both downstream artifacts, score and band, failed gates, the single recommended correction, and the next owner.

Verdict handling

• blocked: stop; route the single recommended correction to the component or human that owns the missing field.
• supervised: stop; return the incomplete supervision gate to a human.
• ready: return the prepared goal and evidence. ready is not execution permission.
• generated methodology skill: return it for independent and human review; do not run it or call loop-readiness-score, whose contract applies to rendered goals.
• run record: return review-required; do not execute, retry, schedule, or update another skill.

Responsibility boundaries

• Selector owns method choice and negative-fit veto.
• Goal engineering owns rendering and character-limit evidence.
• Readiness owns gating and independently recounts the executable goal.
• Methodology skill creator owns one task-scoped SKILL.md proposal.
• Run record owns normalization and post-run evidence identity.
• This orchestrator owns only order, faithful transport, and stop routing.

Do not restate method rules, render goals, recalculate readiness, repair a downstream artifact, execute the task, schedule a loop, or approve an action.

Stop

Stop after one terminal handoff: none, blocked, supervised, ready, or one generated methodology skill/run record awaiting review.