Architecture Decision Records

• 1. Record architecture decisions
• 2. Use sentinel file and argument-shape heuristic for CodeGraph nudge hook
• 3. Document ADR tooling workflow as a skill
• 4. Pre-dispatch model tier resolution enforced by a PreToolUse hook
• 5. On-demand knowledge indexing with four freshness gates
• 6. Per-increment trunk integration topology for agent work
• 7. Eval confidence-pyramid tier vocabulary
• 8. Use effort bands instead of model names in agent frontmatter
• 9. Human Consent Gate for Learning Loop
• 10. Per-Session Analysis Granularity
• 11. Enforce the context ceiling with a transcript-measured PreToolUse hook
• 12. Auto-bootstrap the CodeGraph index per clone
• 13. LLM-driven orchestration over deterministic workflow scripts
• 14. Python for cross-OS scripts
• 15. Bash removal complete — plugins/dev-team is Python-native
• 16. Rely on harness-native compaction; the plugin performs structured summarization only
• 17. Single build cadence — remove the Classic TDD opt-in
• 18. Native sub-issues determine the epic label
• 19. Orchestrate the phase pipeline, choreograph the safety layer
• 20. Fold the ACI mutation pipeline into the plugin as scripted mechanics
• 21. Claude-only model routing — no multimodal or cross-provider tier
• 22. Reject the delegation-only sweep — delegation must earn its measured overhead
• 23. Calibrate agent effort bands from the #1184 eval baseline
• 24. Keep the single model-routing map keyed by canonical model IDs
• 25. Allow versioned (dated-snapshot) model IDs in the routing map
• 26. Adopt native model:/effort: agent frontmatter, retire the band resolver
• 27. Require and mechanically derive color on every agent
• 28. Require skills: preload and memory: project on matching agents
• 29. Code-modifying agent personas remain separate from software-engineer
• 30. Use ancestor-check instead of exact-SHA comparison for mutation baseline reuse
• 31. Raise the shipped Python floor from 3.8 to 3.10
• 32. Shipped-script path-resolution taxonomy
• 33. Quote every ${CLAUDE_PLUGIN_ROOT} expansion in a shell fence
• 34. Do not build a shared-context pre-pass for duplicate full-file reads (#1611)
• 35. Defer two-lane generation/validation concurrency in mutation-kill --all (#1909)
• 36. The two session extractors stay forked, for now
• 37. Block by default at the context ceiling
• 38. Raise the absolute context ceiling from 150K to 350K
• 39. Only Skill loads are worth blocking at the context ceiling
• 40. Evaluate splitting orchestrator.py: no-go
• 41. Autoship's per-round agent dispatch is not ADR 0022's rejected sweep
• 42. Unify the session extractors