Strict TDD — launch the chain

The TDD chain is the run-lot workflow (workflows/run-lot.js, shipped by the plugin): cycles read from the sheet's ## Cycles block (design only when absent), one tdd-cycle agent per cycle — RED, a checkpoint the TDD scope hook checks on the tree, then GREEN + REFACTOR + COST + handler CLAUDE.md —, global green beside the pre-audit gate (a haiku agent runs what cctoolkit suites-plan resolves, a repair agent opens only on a red suite), audit (ddd-tdd-auditor). A batch whose files touch no authorisation, migration nor public contract closes light (cctoolkit closing-level): one review of the diff on Correctness and Scope instead of the audit; the next sheet is never reviewed by the run, its own design checks it; the report's Closing: line says which and why. The script holds the sequence; agent definitions, .claude/rules/*.md and references/ hold the method. Never run the chain by hand, never commit by hand: the workflow commits each DONE batch itself (cctoolkit batch-baseline commit).

$ARGUMENTS

When

• Validated plan from /plan-implementation, batch sheet todo/<code>/<CODE>-PLAN-FX.md present → this skill.
• No plan → no code: route to /plan-implementation.
• /effort medium before launching (hooks/implement-tdd-guard.sh refuses high).

Launch

• batch FX <global plan path> → sheet = <plan without .md>-FX.md; Skill cctoolkit:run-lot, argument = that sheet path as text (never an @ mention).
• <global plan path> alone, or "the whole plan" → Skill cctoolkit:run-plan, argument = the plan path, optionally --until FX / --only FX.
• No path → ls todo/*/*-PLAN.md; several → ask.
• batch FX — correction: manual finding → one Edit of the sheet: a new cycle in its ## Cycles block (behaviour = Correction Cn — [finding], wave after the last, contract fields per references/design.md), then cctoolkit sheet-cycles <sheet> --index exit 0. Sheet without the block → sub-step Correction Cn — [finding] under the affected step, the run's design writes the block. Finding changing scope, an RM/CU or a design decision → stop, /business-spec or /plan-implementation. Then launch run-lot on the sheet.
• replay batch FX <global plan path>, or replay <global plan path> for every batch ✅ DONE → Skill cctoolkit:run-replay, argument = the sheet (or the plan) path, then the options given: --red (each cycle's tests rerun in a worktree without the batch's production code: they must fail), --no-audit, --only FX (plan), --base <rev> (a batch committed without its Kit-Batch trailer). Read-only: nothing fixed, committed nor ticked; no plan-next-batch check, the batch is delivered.

Before launching run-lot or run-plan: cctoolkit plan-next-batch <plan> --outside FX exit 0, cctoolkit doctor without FAIL, test containers reachable when the batch touches Infrastructure.

During the run

The run never asks nor waits: an ambiguity is decided by the agent that meets it and listed under ## To validate; what no retry settles ends the batch at once, BLOCKED. Never resume it (resumeFromRunId): launch a new run, the saved cycles are not redone.

End

Relay the workflow result: status, reason, verdict, report path (run/batches/FX.md; for a plan, run/<stamp>.md, its ## To validate first: the decisions taken without the user, to read after the plan). A plan result carrying merge → say its state and lines, then Skill cctoolkit:knowledge-merge with the plan path: it shows the diff and writes the base on the user's confirmation only. BLOCKED → fix the cause, record Hn marked user decision in the sheet, launch a new run.

A replay result: its status (PASSED, FAILED, SKIPPED with the reason — no commit carries the batch's trailer), the cycles red, the suites red, the audit gaps and toCheck, the report path (run/batches/FX-replay.md; for a plan, one row per batch). FAILED → the report's ## Next holds one batch FX — correction: … per finding: propose them, apply one only on the user's word (the correction flow above, then run-lot).

Batch run DONE → end with → Batch FX complete — manual validation required. Run /clear before any other batch.