prd-to-action

Universal harness that turns any PRD, spec, or requirements document into directly executed development tasks enforced via Claude Code tool calls — not /goal text generation, but real Bash-gated TDD loops. Core mechanism: Adversarial evaluation. Instead of checkbox compliance ("coverage ≥ 80%"), the harness tries to break things. The BREAK step deletes implementation and demands tests fail with AssertionError — not ImportError. GREEN includes GENERALIZE: a second test with different inputs must pass without changing code. No hardcoded thresholds — coverage floors and retry ceilings derive from baseline measurements. Flow: PRE-FLIGHT classifies projects (greenfield/brownfield/resume) → INGEST reads the PRD, measures baseline, generates CLAUDE.md and .memory/project.md → ROUTE manages milestone state machine → EXECUTE runs adversarial TDD (RED → GREEN+GENERALIZE → BREAK → REFACTOR → bash-gated GATE) → LOOP. Each step is blocked on exit code evidence. No evidence → no progress. Optional GitNexus integration for brownfield projects: pre-change impact analysis on existing symbols, post-commit architecture drift detection. Advisory only — tests remain authoritative. Progressive disclosure: SKILL.md (~540 lines) with 5 reference files loaded on demand. Shell translation table for PowerShell/Bash compatibility. Human escalation rules: only ask when genuinely blocked, never for routine decisions. 6 files, zero hardcoded numbers, one principle: if you can't delete the code and watch the test die for the right reason, you haven't proven anything.