Advisory Loop

The advisory loop turns CodyMaster's execution telemetry into an operator-facing recovery workflow.

What it uses

• execution_analyses in .cm/context.db for recent task outcomes
• skill_metrics in .cm/context.db for per-skill quality signals
• the legacy cm-skill-health advisory consumer id for diagnosis payloads
• cm-skill-evolution for FIX / DERIVED / CAPTURED follow-through

Commands

cm advisory report

Use this to inspect recent execution analyses in a human-readable format.

Example:

cm advisory report --project . --limit 10

What you get:

• task title
• execution status
• recommended action
• confidence
• active skills involved in the run

cm advisory metrics

Use this to inspect aggregated per-skill metrics and current qualityWeight() output.

Example:

cm advisory metrics --project . --limit 10

What you get:

• skill name
• quality score
• selection/application/completion counts
• fallback count
• last recommended action

cm advisory handoff

Use this to build the structured note that self-healing skills consume.

Examples:

cm advisory handoff --project . --for cm-skill-health
cm advisory handoff --project . --for cm-skill-evolution --format json
cm advisory handoff --project . --for cm-skill-evolution --analysis EA-123abc --skill cm-browse

What it includes:

• consumer (cm-skill-health legacy handoff id or cm-skill-evolution)
• recommendation (FIX, DERIVED, CAPTURED, or NONE)
• confidence
• source analysis metadata
• target skill judgment
• current metric snapshot and quality_weight
• evidence summary
• suggested next step

Recommended operator flow

1. Run cm advisory report to see the latest execution outcomes.
2. Run cm advisory metrics to confirm whether the affected skill is weak or healthy over time.
3. Run cm advisory handoff --for cm-skill-health to create a diagnosis note.
4. If the issue is real, run cm advisory handoff --for cm-skill-evolution and repair the skill in exactly one mode: FIX, DERIVED, or CAPTURED.
5. Re-run validation and the repo test gate after any meaningful repair.

Notes

• Mandatory chain steps still remain mandatory. Advisory quality feedback only influences optional skill ranking.
• --format json is the stable path if another agent or script needs to consume the handoff payload.
• The loop is still operator-invoked. CodyMaster does not auto-mutate skills from advisory output.

MCP / JSON Surface

Agents that consume CodyMaster through MCP can now read the advisory loop directly without scraping CLI text:

• cm_advisory_report — recent advisory analyses as JSON
• cm_advisory_metrics — aggregated skill metrics and quality_weight
• cm_advisory_handoff — structured recovery handoff for cm-skill-health or cm-skill-evolution

These tools are exposed by src/mcp-context-server.ts and summarized in API Reference.

See also

• Engineering Pipeline
• Working Memory
• CLI Command Reference
• Orchestration Skills