write-tests

Write or update tests for the current task.

Procedure

WT-0. Read the session facts once, before the pre-launch steps that consume them: node "$AGENTS_MAIN_ROOT/bin/workflow/read-session-facts" --session "$SESSION_ID"

• PLANS_DIR= — substitute this absolute path for every <PLANS_DIR> below.
• GATE_CONFIRM_TESTS= — the WT-5 pre-action gate (ON / OFF / ERROR).
• COMPLEXITY_LEVEL_write_tests=, COMPLEXITY_MODEL_write_tests= and COMPLEXITY_SIGNALS= — the WT-6 level, model and signals.
• If the command exits non-zero, or PLANS_DIR=NONE, stop — do not proceed with any step that uses <PLANS_DIR>; report via /supervisor-report; never construct a path like NONE/<session-id>-....

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 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 WT-4. WT-4. List all planned test cases by category (include call-path error cases from WT-3). WT-5. Decide the destination of each planned case — append to an existing test file, or create a new one.
• For each planned case, state its complete source set S (the # Tests: tokens the case protects).
• Run bash "$AGENTS_MAIN_ROOT/bin/find-tests-for-source.sh" --with-language --sources <comma-joined S> with every distinct S in one call, repeating --sources (Bash, standalone command with timeout 600000; split rule and language columns in skills/_shared/test-design/append-vs-new.md; rows come back in query order; read-only, so hooks/block-tests-direct.js does not apply). On exit 4 or a tool timeout, HALT and surface the stderr.
• Record each row's verdict/reason/target as the case group's destination. append is mandatory; the sole permitted new file when a target exists is the size-hard-limit case, per skills/_shared/test-design/append-vs-new.md — do not decide by eye.
• If GATE_CONFIRM_TESTS is ON or ERROR, present the planned cases together with their destinations and wait for user confirmation before WT-6. WT-6. Determine the subagent's model:
• If COMPLEXITY_LEVEL_write_tests from WT-0 is not NONE, use COMPLEXITY_MODEL_write_tests as the model and COMPLEXITY_SIGNALS directly; skip the fallback below.
• If NONE (fail-open for sessions without persisted evaluation): follow skills/_shared/jev-two-stage-judge.md with point=complexity-judge stage=write-tests; use its model= and SIGNALS_CSV — never judge the level inline.
• Emit in Claude text output (NOT Bash echo):

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

WT-7. Launch a subagent (Agent tool, mode: "default", model: <model alias from WT-6>) to autonomously: WT-7a. Write the test file(s). WT-7b. Run tests with timeout. WT-7c. Fix failures and re-run until green. WT-7d. Review test coverage against skills/_shared/test-design.md categories — fix gaps. WT-7e. 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 WT-6 verbatim (comma-separated IDs, or "none")
• source_files: list of source file paths from WT-2
• planned_cases: list of planned test cases from WT-4 — each entry states the source set it protects
• test_destinations: one entry per case group, keyed by that group's complete source set — append <path> or new. Each source set's new verdict generates its own independent new file — one new file per source set. A new test file's path MUST be tests/<category>/<file> (category = first path segment of the source it protects; valid categories: hooks bin skills agents install tests) — never a flat tests/<file>; the commit gate rejects a newly-added flat test (#1834, #2392). Name a new file by a supported entry's patterns in hooks/lib/test-language-registry.json. Follow the language column of the WT-5 output: a new test file MUST load that registry entry's helperLibrary — bin/check-test-frontmatter.sh rejects new files without it. If the Tests: header lists 2 or more paths, every case MUST carry a case marker written per that entry's notationDoc; bin/check-case-markers.sh flags multi-path files without markers as HIGH. On append: never rewrite the target's # Tests: line; # Tags: may only be added to. append is mandatory when the verdict is append; skills/_shared/test-design/append-vs-new.md is the SSOT for when a new file is warranted instead. 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 include the canonical line of the rules/shell-commands.md Subagent Dispatch Prompts section, verbatim as a standalone line, with <AGENTS_MAIN_ROOT> replaced by the literal absolute path. 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."
• (Optional) Follow agents/lib/nfr-severity-calibration.md to obtain the PROJECT NFR block and use it as a test constraint when relevant.