
Pattern mining and proposal generation for Canon's learning loop. Analyzes codebase patterns, review history, build execution data, and conventions to produce structured improvement proposals. Covers manual analysis and auto-trigger modes. Loaded by the learner agent.
Requirements sharpening for build requests. Classifies request fuzziness into tiers (trivial, clear, fuzzy) and applies proportionate PM effort -- from pass-through to full creative divergence. Produces a sharpened-request artifact for architect hand-off. Used by the PM orchestrator.
Procedural how-to for inter-wave handoff coordination: analyze execution reports, push back on weak verdicts, draft next-wave prompts.
Principle, convention, and agent-rule authoring. Covers creation, editing, and applying accepted learner proposals. Handles interview, examples, conflict detection, format validation, and save. Loaded by the writer agent.
Mechanical runbook synthesis from canonical step vocabulary. Composes a step-by-step runbook from the design using only canonical step IDs. Validates step composition, enforces mandatory tail, and emits per-signal confidence. Used by the architect agent.
Strategic analysis and planning brief production. Evaluates build requests, challenges assumptions, considers alternatives. Used by the architect agent as a requirements interview fallback when PM conversation left gaps.