epic: Claude Code integration optimization #707
Description
Overview
Analysis of Claude Code internals reveals patterns we can adopt. After honest review, only 4 issues survived — the rest were over-engineering.
Surviving Issues
Verified value
- feat: role-based tool sets for agent execution #701 — Role-based tool sets (fewer tools = fewer tokens per API call)
- feat: effort-aware auto-compact and budget caps #702 — Effort-aware auto-compact and budget caps (needs env var verification)
- feat: PreToolUse hooks for agent guardrails #704 — PreToolUse hooks for agent guardrails
- feat: transcript normalization for squad conversations #705 — Transcript normalization for conversations
- feat: memory staleness caveats #721 — Memory staleness caveats (3 lines of code, quick win)
Deprioritized
- feat: parallel scanner execution in squad conversations #706 — Parallel scanner execution (real value but large effort, premature)
Closed as noise
- feat: prompt cache optimization — structure context for cache hits #703, feat: topic-based memory files with frontmatter #716-720, epic: memory & context injection overhaul #722 — Over-engineered patterns from Claude Code that don't fit our single-shot execution model
Lesson
Analyzing another system's architecture is valuable research. But filing issues for every pattern found is busywork disguised as planning. Each issue must answer: what real failure does this prevent that we've actually seen?