claude-tdd-pro

Write a Markdown specification for a non-trivial feature BEFORE any code. Loads the spec-first skill explicitly.

Write an Architecture Decision Record (MADR format) in the project's adr/ directory. Loads adr-writer skill.

Cold-start audit of the current codebase against Google's published engineering standards (eng-practices, JS/TS style, Python style). Produces COMPLIANCE-REPORT.md (ranked findings with citations) and LANDMINES.md (high-risk files to refactor with care). Read-only — no edits.

Verify claude-tdd-pro is wired up correctly on this machine and against this project. Smoke-tests every hook, runs the rubric runner against a known-bad fixture, runs the eval suite. Reports green / yellow / red per primitive. Read-only — no edits.

Extract a component / module from a god-file using the strict 9-step test-first refactor pattern.

Example customer-facing agent template demonstrating the §2.26 v1.11 safety contract. Operators copy this file and replace fields. The agent answers product-catalog questions grounded in the operator's RAG store; it refuses ungrounded queries and routes its output through the safety classifier.

Architecture elicitor subagent for W-1. Decomposes a feature description into decision points, enumerates grounded options per S/L/C source, asks the operator per decision, declines un-grounded options.

Independent-verifier subagent for changes touching paths listed in `.claude-tdd-pro/critical-paths.txt`. Reads the candidate diff on stdin and emits `verdict: agree` or `verdict: disagree <reason>` so the F-3 `/agent-verify` command can gate the commit. Runs as a separate Claude instance with no shared context with the editing instance, per the Karpathy/Cursor independent-verifier discipline.

Use as a final-pass reviewer with NO prior conversation context. Catches drift the active session is blind to (assumptions accumulated, scope creep, naming inconsistencies invisible from inside the work). Pattern from 2389-research/claude-plugins. Read-only; returns a punch list to the parent.

W-11 coordinator subagent that fans out work to multiple §2.3 subagents within a single CL, serializing on §2.7 overlapping lock sections, deduplicating findings by rule_id+file+line, and reporting per-subagent token cost with subagent_id preserved for H-12 telemetry. Refuses parallel execution when total estimated tokens exceed `userConfig.parallel_budget_tokens`.

Use when the user explicitly invokes /adr or asks to "record an architecture decision" / "write an ADR." Produces an MADR-format Architecture Decision Record. Pattern from adr.github.io and the AgDR (Agent Decision Record) extension at me2resh/agent-decision-record. Lightweight, in-repo, future-engineer-readable.

H-13 long-running agent harness continuity per §26 v1.11. When a CL exceeds a single agent context window, the initializer agent writes a continuation artifact at `.claude-tdd-pro/agent-continuations/<session_id>.json` and the incremental successor agent reads it at startup to resume from `next_action` without losing the §2.15 workflow state envelope.

Interactive architecture elicitation. Decomposes a feature description into decision points, enumerates grounded options per S/L/C source, prompts the operator per decision, and writes ADRs.

Use when the user reports a BUG (existing behavior is wrong) — phrases like "X breaks when Y", "this used to work", "got an error/exception/500", "regression", "this is a bug". NOT for new features (use tdd-feature-build). Enforces bug-as-failing-test — capture the broken behavior FIRST, confirm red, fix, confirm green. The fix never lands without the test that would have caught it.

Opt-in concurrent-CL execution (W-10 / §2.23). Reads active session envelopes from `.claude-tdd-pro/active-sessions/<session_id>.json` and gates `/spec` / `/feature` / `/architect` invocations against the §2.23 five-dimension disjointness contract (phase set, workflow-state subsections, lock sections, source-folder ownership, commit branch). Default is sequential per §20; activated via `userConfig.allow_concurrent_cls: true` or per-invocation `--concurrent`.

5 hooks across 4 events

Output style that biases prose toward citing Google-style rule IDs (RUBRIC.yaml) and upstream Google anchors. Use when working in code review, /analyze, /remediate, and /pr flows. Suppresses purple prose; favors terse, citation-anchored statements.

Output style for opening a PR. Test-Plan-first prose voice. Strips marketing language, requires concrete commands and observed output, and cites rule IDs the diff closes.

Output style for the red-green-refactor cycle. Each step is announced as RED, GREEN, or REFACTOR with the test or behavior it covers. Suppresses anything that breaks the cadence — no exposition between cycles.