Learning Feedback Loop — decisions.jsonl + timeline summaries (no PII)

[DevVig]

Summary
Capture decisions and outcomes to enable lightweight learning over time. Persist as local JSONL and post concise, human-readable summaries to the timeline. Exclude PII and keep data in-repo.

Why

• Improves consistency and speed on repeated patterns
• Maintains privacy and portability

Scope

• JSONL decision log + summarizer that posts to docs/timeline/

Tasks

• Define schema for decisions: {context_keys, decision, outcome, ts}
• Implement writer utility a2a/trace.py to append JSONL: docs/timeline/decisions.jsonl
• Add summarizer that produces docs/timeline/decisions-summary.md (periodic or per N entries)
• Ensure no PII included; document governance
• Hook calls in orchestrator after key actions (gate, model selection, adaptations)
• Add tests for writer and summarizer

Acceptance Criteria

• JSONL log is created and updated with each key decision
• Summary file present with compact sections and links to artifacts
• Verified that entries contain no PII

Risks & Mitigations

• Data sprawl -> keep minimal schema and rotate/prune as needed
• Privacy -> document and enforce no PII policy

References

• suggestion.md (Feedback Loop Integration)
• a2a/trace.py, a2a/orchestrator.py