Reflect Digest

Turn repeated post-run reflection signals into bounded improvement work without re-diagnosing one receipt or drafting a parallel pull-request artifact.

The graph reads durable reflect events through data-store, admits explicit replay projections when supplied, applies deterministic confidence and support floors, then asks one bounded agent act to describe at most one improvement per skill. A final deterministic step verifies every cited receipt belongs to that skill's group and emits a skill-lab improve handoff.

The skill does not write a package, open a pull request, publish, or invoke the mutating improvement runner. The handoff preserves the target, objective, receipt evidence, and non-goals for later governed execution.

Composes

• data-store#list_stream_heads

Inputs

• reflect_projections: explicit projections for replay; otherwise durable state is read from data_source_ref and state_resource.
• skill_filter, since: optional read bounds.
• min_support: minimum number of admitted projections for one skill.
• min_confidence: minimum confidence for each admitted projection.

Outputs

• proposals: validated skill-specific improvement opportunities.
• skill_lab_handoffs: executable request packets naming skill-lab, runner improve, target directory, objective, primary receipt, evidence summary, and supporting receipt ids.

Agent task contracts

reflect-digest

Review each grouped reflection independently. Return at most one proposal per skill_ref, and return an empty proposals array when no group supports a bounded fix. Every proposal must contain skill_ref, repo-relative target_dir, objective, evidence_summary, supporting_receipt_ids, and boundaries. Receipt ids must come from that group. Do not diagnose a single run again, author code, draft a pull request, publish, or mutate the target package.