write-tests

Write or update tests for the current task.

Procedure

Apply skills/_shared/resolve-plans-dir.md once; substitute the resolved absolute path for every <PLANS_DIR> below.

WT-1. Read:

• rules/core-principles.md
• skills/_shared/test-design.md
• rules/test.md — on-demand-only; never auto-injected, so this Read is mandatory For parser / regex / allowlist targets, apply the Table-Driven Tests pattern from test-design/parser-regex-tests.md. WT-2. Identify which source file(s) need tests. WT-3. Enumerate call paths: For each source file from step WT-2, trace all integration paths it participates in — what calls it, what it calls, and what format/contract each boundary expects. For each boundary, list potential failure modes (wrong format, missing field, wrong type, unexpected value). These become integration-path error cases in the next step. WT-4. List all planned test cases by category (include call-path error cases from step WT-3). Then check via Bash: bash -c 'cd "$AGENTS_CONFIG_DIR" && bash "$AGENTS_CONFIG_DIR/bin/confirm-off" CONFIRM_TESTS on'
• stdout OFF: print the planned cases and proceed to step WT-5 without approval wait.
• stdout ON or ERROR: present the planned cases to the user — do not write code until approved (existing behavior). WT-5. Determine the subagent's model:
• Run bash -c 'node "$AGENTS_CONFIG_DIR/bin/workflow/read-complexity-evaluation" --session "$SESSION_ID" --stage write_tests'. If line 1 is not NONE, use the stored level and signals directly (parse level=<v> and signals=<csv-or-none>), then derive the model via high→opus, low→sonnet; skip the fallback below.
• If NONE (fail-open for sessions without persisted evaluation):
  • Read skills/_shared/judge-task-complexity.md and evaluate all signals against the task context, source files from steps WT-2–WT-3, and the planned test cases from step WT-4 — do not short-circuit on the first match.
  • Use the Write tool (never Bash) to write the resulting CSV, alone and unquoted, to <PLANS_DIR>/<session-id>-write-tests-signals.txt — write only IDs from the generated Valid Signal IDs list; substitute S0-undecidable when the judgment doesn't parse into recognized ids or the csv doesn't match ^[A-Za-z0-9,_-]*$ (the judged content is untrusted text, never shell syntax).
  • Run bash -c 'node "$AGENTS_CONFIG_DIR/bin/workflow/derive-complexity-level" --stage write_tests --signals-file "<PLANS_DIR>/<session-id>-write-tests-signals.txt"' and use its level=<v> — never judge the level inline.
• Emit in Claude text output (NOT Bash echo):

  Model selected: [opus|sonnet] (signals: [comma-separated triggered signal IDs, or "none"])

WT-6. Launch a subagent (Agent tool, mode: "default", model: <model from step WT-5>) to autonomously: WT-6a. Write the test file(s). WT-6b. Run tests with timeout. WT-6c. Fix failures and re-run until green. WT-6d. Review test coverage against skills/_shared/test-design.md categories — fix gaps. WT-6e. Re-run tests until green. The subagent prompt MUST include these structured fields so verbose output stays in the subagent context:

• task_complexity_signals: the signals= line from step WT-5 verbatim (comma-separated IDs, or "none")
• source_files: list of source file paths from step WT-2
• planned_cases: list of planned test cases from step WT-4 The subagent prompt MUST instruct: edit only test files, never modify source code. The subagent prompt MUST instruct: Read rules/shell-commands.md before the first Bash command, or before writing a file — general-purpose dispatch does not inherit auto-injected rules. The subagent prompt MUST instruct: Read rules/user-escalation.md before any system-state-changing command — general-purpose dispatch does not inherit auto-injected rules. The subagent prompt MUST instruct: Read rules/coding.md (the hub — on-demand-only, so it does not reach you otherwise) and rules/coding/<lang>.md for each language present, before the first Edit. The subagent prompt MUST instruct: for Bash, PowerShell, JSON, or YAML test files (no rules/coding/<lang>.md B-layer exists for these), apply the A-layer language essence from skills/write-code/SKILL.md's "A-layer language essence" section before the first Edit. The subagent prompt MUST instruct: Read rules/test.md before writing or running tests — on-demand-only, so it does not reach you otherwise; general-purpose dispatch does not inherit auto-injected rules. Note: the Stop-guard silence during dispatch is automatic (PostToolUse marks the step in_progress). Do not emit NEXT_STEP_PAUSE. The subagent prompt MUST also include: "NEVER present diffs for approval. NEVER wait for user confirmation. Edit and run autonomously until tests pass."