Test Deep — Context-Aware Test Orchestration

Supplementary Agent

When test failures occur, dispatch background triage:

Agent({ description: "Analyze test failure root cause and suggest fixes", subagent_type: "verify-app", prompt: Analyze the following test failure: <failure output> Identify root cause and suggest minimal fix. })

Trigger

• Keywords: context-aware testing, smart test, test orchestration, test triage, test failure analysis, test deep, run relevant tests

When NOT to Use

Scenario	Alternative
Writing new tests	/post-dev-test
Reviewing test coverage	/codex-test-review
Generating unit tests	/codex-test-gen
Full manual test run	/verify
Code review	/codex-review-fast

Prohibited Actions

❌ git add | git commit | git push — per @rules/git-workflow.md

This skill runs tests and may apply safe fixers, but does not commit. To commit, the user must invoke /smart-commit --execute separately.

Workflow

flowchart TD
    U[User: /test-deep] --> S[Phase 0: Test Selection]
    S --> |git diff mapping| T[Test Targets]
    S --> |no mapping| F1[Framework --changedSince]
    S --> |low confidence| F2[Full Suite]
    T --> L[Phase 1: Progressive Ladder]
    F1 --> L
    F2 --> L
    L --> |unit| R1[Run Unit Tests]
    R1 --> |pass| R2[Run Integration Tests]
    R1 --> |fail| TR[Phase 2: Triage Pipeline]
    R2 --> |pass| R3[Run E2E Tests]
    R2 --> |fail| TR
    R3 --> |pass| DONE[All Pass]
    R3 --> |fail| TR
    TR --> P[Parser: Structured Tags]
    P --> LLM[LLM: Root Cause + Action]
    LLM --> FC[Phase 3: Fixer Catalog Lookup]
    FC --> SG{Safety Gate}
    SG --> |safe| AUTO[Auto-run Fix]
    SG --> |side-effect| ASK[AskUserQuestion]
    SG --> |destructive| BLOCK[Manual Only]
    AUTO --> L
    ASK --> |approved| L

Phase 0: Test Selection

Map code changes to test targets. See references/test-selection.md for full strategy.

Strategy priority:

Priority	Method	When
1	Git diff filename mapping	Default — map changed files to test files via Glob
2	Framework native	Jest --changedSince, Vitest --changed
3	Full suite fallback	Config change, no mapping, --all flag

Steps:

1. Collect changed files: union of unstaged + staged + untracked (--branch uses merge-base)
2. Apply filename mapping rules to generate candidate test paths
3. Glob-confirm each candidate exists
4. Classify confirmed tests by layer (unit / integration / e2e)

Full suite escalation triggers: config file changed, CI/CD file changed, package dependency changed, no test files mapped, --all flag.

Phase 1: Progressive Ladder

Execute tests layer-by-layer with fail-fast.

Layer	Directory Pattern	Timeout	Fail Behavior
Unit	test/unit/**, test/scripts/lib/**, unclassified	60s	Fail-fast → triage
Integration	test/integration/**	300s	Fail-fast → triage
E2E	test/e2e/**	600s	Enter triage

Fail-fast rule: Unit fail → skip integration + e2e. Integration fail → skip e2e.

Override: --no-fail-fast disables this behavior (run all layers regardless).

Layer detection: Classify test files by directory prefix. Files not matching any integration/e2e pattern → treat as unit.

Execution: Run tests using project's configured test command (from package.json scripts or CLAUDE.md). Capture stdout/stderr and exit code.

Phase 2: Failure Triage Pipeline

When failures occur, analyze and classify. See references/triage-pipeline.md for full spec.

Step 1: Output Parser

Extract structured tags from test output (no classification — just structure):

Tag	Description
exit_code	Process exit code
error_signatures[]	Regex-matched error patterns
failing_tests[]	Failed test names
failing_files[]	Failed test file paths
env_hints[]	Environment clues (testnet, localhost, etc.)
stack_depth	Stack trace line count

Step 2: LLM Root Cause Analysis

Feed parser tags + compressed output to LLM for classification:

Classification	Definition	Typical Action
code_bug	Logic error in code	Fix code
infra	Infrastructure issue (port, dependency)	Restart / reinstall
environment	External precondition unmet	Fixer catalog action
flaky	Non-deterministic failure	Retry + quarantine tag

Mandatory secret redaction before LLM analysis — per @rules/logging.md:

• API keys, private keys, tokens, passwords, mnemonics, URLs with credentials
• See references/triage-pipeline.md for redaction patterns

Step 3: Safety-Gated Action

Route suggested_fixer through safety gate:

Tier	Auto-run?	Confirmation	Examples
safe	Yes	None	retry, clear_cache
side-effect	No	AskUserQuestion	reinstall_deps, restart_server
destructive	Blocked	Manual only	Reset state, drop tables

Default-deny: Unknown fixer → side-effect tier (require confirmation).

Phase 3: Fixer Execution

Lookup fixer from catalog, check tier, execute or prompt. See references/fixer-catalog.md for full catalog.

Core fixers (plugin-shipped):

Fixer	Tier	Description
retry	safe	Re-run failing tests
clear_cache	safe	Clear build/test cache
reinstall_deps	side-effect	Remove node_modules + reinstall
restart_server	side-effect	Kill + restart dev server
port_cleanup	side-effect	Kill process on conflicting port

Host extensions: Project-specific fixers in .claude/test-deep/fixers.md. Schema validation at load — missing required fields or invalid tier → rejected with warning.

Fixer loop: After safe/approved fixer runs, re-enter progressive ladder for failed tests only. Max 1 fixer retry per failure to prevent infinite loops.

Session Artifacts

Write per-run artifacts for comparison and debugging.

Location: .claude/cache/test-deep/<runId>/

Run ID format: <timestamp>-<shortSHA>-<pid>

File	Content
metadata.json	Run config: selected tests, ladder config, changed files
results.json	Per-layer results: pass/fail, duration, exit codes
triage.json	Parser tags, LLM classification, fixer chosen, outcome
output.log	Compressed test output (secret-redacted)

TTL: Keep last 5 runs. Prune older on new run start.

latest symlink: Points to most recent run directory.

Arguments

Flag	Default	Description
--all	false	Force full test suite
--layer <unit|integration|e2e>	all	Run only specified layer
--no-fail-fast	false	Run all layers regardless of failures
--no-fix	false	Triage only, skip fixer execution
--focus <path>	—	Limit test selection to path
--branch	false	Use merge-base diff instead of working tree

Output

## Test Deep Report

### Test Selection
- Changed files: N
- Mapped test files: N (N unit, N integration, N e2e)
- Selection method: git diff mapping | framework native | full suite

### Results

| Layer | Tests | Passed | Failed | Skipped | Duration |
|-------|-------|--------|--------|---------|----------|

### Failure Triage

| # | Test | Classification | Root Cause | Fixer | Tier |
|---|------|---------------|------------|-------|------|

### Actions Taken
- [N] fixer_id: outcome

### Gate
✅ All Pass | ⛔ N failures pending resolution

Verification Checklist

• Test selection maps changed files to test targets
• Progressive ladder respects fail-fast
• Triage pipeline produces structured classification
• Safety gate enforces tier-based execution
• Secret redaction applied before LLM analysis
• Session artifacts written to cache
• No git add / git commit / git push executed

References

• references/test-selection.md — Git diff mapping strategy + full suite escalation
• references/triage-pipeline.md — Parser tags + LLM prompt + safety gate
• references/fixer-catalog.md — Core fixers + host extensions + safety tiers
• @rules/logging.md — Secret redaction policy
• @rules/docs-writing.md — Output format conventions