Decision Template

Decision extraction is forced into a fixed shape for stable retrieval quality.

Required Sections

Every decision memory content is normalized to include:

Summary:
Why:
Alternatives:
Impact:
Evidence:

Formatting Rules

Summary: 1-2 lines
Why/Alternatives/Impact: 1-3 bullets each
Evidence: include commit/raw evidence references whenever available

Enforcement

LLM extractor prompt asks for strict structured output.
Server-side normalization fills missing sections automatically.
Even when data is missing, section headers are preserved.

Example

Summary:
Switch memory search to hybrid ranking with explicit debug scores.

Why:
- Keyword-only ranking missed high-quality decisions.
- Teams needed consistent quality controls.

Alternatives:
- Keep keyword-only retrieval.
- Use semantic-only retrieval.

Impact:
- Recall quality improves for cross-session workflows.
- Debug mode can explain ranking details.

Evidence:
- commit_sha: 8c3f0a12d9
- raw_event_id: 8d53cceb-2aab-4f08-9bf4-b038b5f76f33