--- name: tdd-review description: | Use when the user asks to review the planned tests in a /deep-plan plan file: their assertions, doubles, fixtures, level, or flake risk. Reads the plan's Tests (TDD) blocks only -- not diffs, pull requests, or implemented code. argument-hint: "[plan-file]" --- # /tdd-review You orchestrate a test-critic fleet over a plan's `**Tests (TDD)**` blocks. All guideline content lives in `${CLAUDE_PLUGIN_ROOT}/skills/tdd-review/references/test-principles.md` and all fleet mechanics live in `${CLAUDE_PLUGIN_ROOT}/skills/design-review/references/fleet-orchestration.md`. Read both now; do not restate their content from memory. ## Step 1: Resolve the plan file from $ARGUMENTS - A path to a plan.md file: use it directly. - A plan folder: use its `plan.md` member. - Empty: ask the user via AskUserQuestion which plan file to review rather than guessing. The review target is the plan's per-task `**Tests (TDD)**` blocks, passed as text excerpts (the critics read files with Read/Grep/Glob — they have no Bash). ## Step 2: Run the critic fleet Run the fleet exactly as specified in `references/fleet-orchestration.md` (under the design-review skill) with `agentType: deep-plan:dp-critic` and `references/test-principles.md` as the cluster source: one finder per H3 cluster under `## Review-time red flags` in that file (quote each cluster's questions verbatim into its finder prompt, and name the file itself — that is what points an otherwise generic critic at test quality), dedup, then the adversarial verify stage. Use the Workflow path when available; on absence, denial, or error, switch to the fallback without asking. ## Step 3: Report Report surviving findings grouped material-then-minor, each as `[severity] {cluster}/{principle}: {finding} -- evidence: {task/field}`, followed by a one-line verdict (`N material, M minor`). If nothing survives verification, say so explicitly and name the clusters checked. This skill only reports; it never edits the target. Offer fixes only if the user asks.