Build mode research: auto-detection, pipeline improvements, contextual awareness
Three parallel research agents completed:
1. Auto-detection (NOVEL — no existing tool does this):
- Simple keyword scoring algorithm (0-3 task, 4-7 task+plan, 8+ build)
- Signal categories: request analysis, codebase impact, historical patterns, risk
- Zero cost for 90% of requests (regex), LLM fallback for 10% ambiguous
- UX: auto-switch high confidence, suggest medium, stay low
2. Pipeline improvements (from Devin, SWE-agent, Aider, Cursor):
- File-level isolation (worktrees per batch)
- Inline verification after every edit (not just batch boundaries)
- Contract testing between batches (highest-value verification)
- Selective rollback + failure-aware re-planning after 2 failures
- Three-phase execution: Foundation → Implementation → Integration
- Trajectory recording for learning
- Dynamic batch sizing (1-3 files default, 80%+ success rate)
3. Contextual awareness:
- Hybrid keyword+LLM classification
- Mode transitions mid-task (upgrade/downgrade with confirmation)
- Context signals: codebase, request, conversation, user preferences
- Progressive disclosure: start light, escalate if complex
Key research insights:
- Agent reliability: 96% → 34% from 8K to 128K tokens (LOCA-bench, ICML 2026)
- 56.2% of agent turns are exploration, not execution (FastContext, 2026)
- Single-file changes: ~80% success vs <50% for 4-6 files (SWE-bench)
- Keep detection simple — gauntlet's complexity was its downfall
Pinned findings updated at research/build-mode/PINNED-FINDINGS.md
Nick
9997dac413