ceo-feedback-iteration
Use when an Agent is asked to classify and improve open repository feedback through the managed feedback-iteration capability. The Skill owns the bounded diagnosis and decision protocol; the service o
SKILL.mdskills/ceo-feedback-iteration/SKILL.md
GitHubCEO Feedback Iteration
Use only the local Console API at http://127.0.0.1:8765/api/console/feedback.
Read the batch and every selected detail before changing a Skill, configuration,
or code. Treat feedback_key as the stable identity. The supplied persisted
summary, detail references, runtime configuration, loaded revision numbers, and
SHA values are factual context; do not ask a model to regenerate a summary or
invent IDs. Direct SQLite writes are forbidden.
Bounded discussion profile
- Establish the observed behavior and loaded capability facts, then identify the likely root-cause class.
- Use brainstorming only if a material product or policy uncertainty remains. Ask one focused user question at a time, offering alternatives and a recommendation. Do not require discussion when the facts establish the path.
- Before any change, persist one typed decision through
POST /api/console/feedback/batches/{batch_id}/decisions. Include the currentfeedback_keyvalues, persisted references, root cause, acceptance scenario, expected behavior, and verification. - Use exactly one scope:
skill_onlywhen policy/procedure is deficient but data, tool, and route exist;runtime_configwhen a valid capability is incorrectly bound or disabled;codewhen a required capability is missing or defective;mixedonly when both are needed;needs_humanfor missing reusable policy, authorization, or indispensable business facts. - Execute only the persisted path.
needs_humannever resolves feedback.
Verification boundary
Read back the decision and batch history. Record path-specific evidence through
the feedback API; a later resolution must prove the selected path, including
the active runtime/load receipt for Skill or configuration work, or commit and
tests for code work. Never claim a new batch while feedback iteration is
disabled, and never change generic brainstorming to encode this protocol.