What problem does it solve? Completed task runs contain valuable lessons buried in session transcripts and feedback files, but without a systematic process they are never captured. This Skill mines those sources and records immutable, time-ordered observations into docs/reflection/observations.yaml, with an ingestion ledger that prevents re-processing. ## Core Features & Use Cases - Transcript and feedback mining: Discovers candidate sessions via auto search and detects changed docs/tasks/*/feedback.md files by SHA-256 hash, then dispatches sub-agents to extract structured observations. - Immutable, time-ordered records: Mints lexically time-ordered IDs via the bundled reflect.py helper and appends observations in place with yq, scaling to thousands of records without loading whole files into context. - Ingestion ledger: Tracks processed session IDs and feedback content hashes so re-runs are incremental and idempotent. - Use Case: After finishing several tasks, run the observe stage to gather lessons from all recent sessions and feedback files into observations.yaml, then hand off to playbook-refine to distil them into reusable rules. ## Quick Start Ask the agent to mine recent task sessions and feedback files for new reflection observations and append them to docs/reflection/observations.yaml.
