test-plan-generate-test-file

Test File Generator Sub-Agent

Generate a complete test file for test cases assigned to one file from the file_mapping.

Called in parallel by test-plan-case-implement (one sub-agent per file in file_mapping).

IMPORTANT: This sub-agent does NOT write files to their final destination. It:

1. Assembles test file content in memory
2. Validates syntax using bash (echo > /tmp/...)
3. Writes result to /tmp/test_plan_results/file_{i}.json
4. Returns tiny confirmation JSON
5. Parent reads files and writes to target repositories

Write tool is intentionally excluded from allowedTools. Uses context: fork for clean return.

Inputs (from Agent prompt)

Parent passes all data via Agent prompt as JSON block:

{
  "file_index": 0,
  "file_path": "tests/upgrade/test_odh_cli.py",
  "test_cases": [
    {"test_case_id": "TC-CLI-001", "title": "...", "objective": "...", ...},
    {"test_case_id": "TC-CLI-003", ...}
  ],
  "function_names": ["test_tc_cli_001", "test_tc_cli_003_pre_upgrade_readiness"],
  "framework": "pytest",
  "conventions_file": "/path/to/conventions.md",
  "pattern_guide": "...content or null...",
  "repo_instructions": "...content or null...",
  "common_setup_requirements": [...],
  "code_repo_path": "/path/to/repo",
  "feature_dir": "/path/to/feature"
}

Parse this JSON from prompt to extract all needed variables.

Process

Step 0: Extract Data from Prompt

Extract JSON data block from the Agent prompt. Parse to get:

• file_index - Index for temp file naming
• file_path - Target file path in code repo
• test_cases - Array of TC objects for this file (already filtered)
• function_names - Array of function names for TCs
• framework - Test framework (pytest, go, jest)
• conventions_file - Path to conventions file
• pattern_guide - Pattern guide content (or null)
• repo_instructions - Repo instructions content (or null)
• common_setup_requirements - Array of shared preconditions
• code_repo_path - Repository path
• feature_dir - Feature directory path

Store test_cases as tcs_for_this_file for use in later steps.

Step 1: Check Existing Test File

IMPORTANT: When analyzing target repository: Read code files and use grep/bash. Do NOT import target repo dependencies (not in test-plan venv) or use inspect.signature().

Get full file path: <code_repo_path>/<file_path>

If file exists, list existing functions:

(cd $(git -C ${CLAUDE_SKILL_DIR} rev-parse --show-toplevel) && uv run python scripts/list_test_functions.py "$full_file_path")

Returns JSON: {"functions": [{"name": "...", "line": 42, "docstring": "..."}]}

Semantic matching: For each TC and expected function name, check if semantic match exists:

• Compare TC title/objective with function name/docstring
• If match: Mark TC as already_implemented, store function details
• Use LLM judgment for semantic equivalence

If already-implemented tests found, ask user via AskUserQuestion:

✓ Found {N} already-implemented test(s) in {file_path}:
  - TC-API-001: test_notebook_creation (line 42)

Options:
1. Skip them - Only implement missing tests
2. Re-generate - Overwrite existing functions
3. Review - Show me existing code first

Set mode: "create" (new file), "append" (add to existing), or filter out implemented TCs.

Step 2: Generate File Header

If mode == "create" (new file), generate header with:

1. Module docstring - Brief description
2. Import statements - Based on framework:
   • Read pattern_guide examples to see common imports
   • Read repo_instructions for import style
   • Add framework imports (pytest, unittest, etc.)
   • Add repo-specific utilities from pattern examples
3. Common fixtures - If common_setup_requirements has items used in this file:
   • Generate fixture functions for shared preconditions
   • Use framework's fixture decorator
   • Reference in test function parameters

Step 3: Generate Test Functions

For each TC in tcs_for_this_file (sequential):

Generate function code using TC spec + conventions + patterns with this priority:

1. CONSTITUTION.md (in repo_instructions) - Hard constraints that MUST be followed
2. CLAUDE.md / AGENTS.md (in repo_instructions) - Repo testing guidelines and conventions
3. odh-test-context conventions (in conventions_file) - Repo-specific patterns (markers, file patterns, fixtures, test commands)
4. Tiger Team pattern guides (in pattern_guide) - Framework-specific patterns (pytest-tests.md, go-tests.md with code examples)

Function structure:

Decorators/Markers:

• Read available markers from conventions_file (odh-test-context provides marker list)
• Map TC priority → repo markers (P0 → tier1 or p0)
• Map TC category → repo markers (API → api)
• Only use markers defined in conventions

Function Signature:

• Use provided function_name exactly
• Add fixture parameters based on preconditions
• Follow conventions import style

Docstring:

• Format: {tc_id}: {objective}
• Example: TC-API-001: Verify metadata returned correctly

Implementation (AAA pattern):

• Arrange: Setup from preconditions
• Act: Execute test steps
• Assert: Validate expected results with messages

Collect each generated function in functions array.

Step 4: Assemble Complete File

• mode == "create": complete_file = header + "\n\n".join(functions)
• mode == "append": complete_file = existing_content + "\n\n".join(functions)

Step 5: Validate Syntax

cat > /tmp/test_file_${file_index}.py << 'PYEOF'
$complete_file
PYEOF
python -m py_compile /tmp/test_file_${file_index}.py 2>&1

If syntax error: Fix once, retry. If still invalid: save as .draft, skip scoring.

Step 6: Score Quality

For each function, invoke in parallel:

/test-plan-score-test-function \
  --test-code-file /tmp/func_${tc_id}.py \
  --tc-file <tc_file_path> \
  --conventions-file <conventions_file> \
  --framework <framework> \
  --output-file /tmp/test_scores/${tc_id}_score.md

Parse score, handle verdicts:

• 7-10: Accept
• 4-6: Auto-revise (re-generate with feedback)
• 0-3: Save as .draft

MAX REVISIONS: 1 per function

Step 7: Return Result to Parent

7a: Write Result to File and Return Confirmation

Build metadata JSON and write to persistent location:

# Create results directory
mkdir -p /tmp/test_plan_results

# Build metadata JSON
cat > /tmp/file_metadata_${file_index}.json << EOF
{
  "file_index": ${file_index},
  "file_path": "${file_path}",
  "tc_ids": ${tc_ids_for_file_json},
  "functions": ${functions_json},
  "quality_summary": ${quality_summary_json},
  "draft_files": ${draft_files_json},
  "errors": ${errors_json}
}
EOF

# Format and write complete result (reads /tmp/test_file_${file_index}.py, embeds content)
(cd $(git -C ${CLAUDE_SKILL_DIR} rev-parse --show-toplevel) && uv run python scripts/format_file_result.py /tmp/file_metadata_${file_index}.json) > /tmp/test_plan_results/file_${file_index}.json

# Output tiny confirmation (NOT full JSON - keeps context clean)
echo '{"status": "complete", "file_index": '${file_index}', "result_file": "/tmp/test_plan_results/file_'${file_index}'.json"}'

CRITICAL: Output ONLY the tiny confirmation JSON. Do NOT add:

• Narrative summaries ("Sub-agent complete...")
• Full file content or quality data
• Success messages
• Any text after the JSON

The parent reads the full result from the file. This keeps the return value small (~100 bytes) and avoids polluting parent context.

7b: Clean Up Temp Work Files

Clean up temporary work files (keep result file for parent):

# Remove temp work files (parallel-safe - indexed by file_index and unique tc_ids)
rm -f /tmp/test_file_${file_index}.py
rm -f /tmp/file_metadata_${file_index}.json

# Remove per-TC temp files
for tc_id in ${tc_ids_for_file[@]}; do
    rm -f /tmp/func_${tc_id}.py
    rm -f /tmp/test_scores/${tc_id}_score.md
    rm -f /tmp/test_scores/${tc_id}_score_revised.md
done

# DO NOT remove /tmp/test_plan_results/file_${file_index}.json - parent will read it

Why keep result file: Parent needs to read the full result after all sub-agents complete. Parent will cleanup /tmp/test_plan_results/ after reading all files.

CRITICAL: After Step 7b cleanup completes, this sub-agent is DONE. Do NOT output any additional text, summaries, or explanations. The tiny confirmation JSON from Step 7a is the complete return value. The parent skill will read the full result from the file and proceed automatically.

---

Style Decision Priority

When making code generation decisions, follow this priority:

1. CONSTITUTION.md (in repo_instructions) - Hard constraints, never violate
2. CLAUDE.md / AGENTS.md (in repo_instructions) - Repo-level testing requirements
3. odh-test-context conventions (in conventions_file) - Repo-specific test patterns from Tiger Team's repository analysis (markers, fixtures, test commands, linting)
4. Tiger Team pattern guides (in pattern_guide) - Framework-specific patterns with code examples (pytest-tests.md, go-tests.md, jest-tests.md)
5. General framework conventions - Standard framework idioms

odh-test-context and Tiger Team guides are both from the Tiger Team - one analyzes repos, one provides framework patterns.

Error Handling

• Syntax error after fix: Save as .draft, continue
• Scorer fails: Accept without score, log warning
• Generation fails: Log error, continue

Philosophy: Partial success > total failure.

$ARGUMENTS