Paths: File paths (references/, ../ln-*) are relative to this skill directory.

Manual Evidence Auditor (L3 Worker)

Type: L3 Worker

Specialized worker auditing whether manual test scripts produce reproducible, useful evidence.

Purpose & Scope

• Audit Manual Evidence (Category 7: Medium Priority)
• Evaluate bash test scripts in tests/manual/ against quality dimensions
• Emit REWRITE_MANUAL_EVIDENCE or KEEP_MANUAL_EVIDENCE
• Calculate compliance score (X/10)

Inputs

MANDATORY READ: Load references/audit_worker_core_contract.md.

Receives contextStore with: tech_stack, testFilesMetadata (filtered to type: "manual"), codebase_root, output_dir.

Manual test metadata includes: suite_dir, has_expected_dir, harness_sourced.

Workflow

Detection policy: use two-layer detection (candidate scan, then context verification); load references/two_layer_detection.md only when the verification method is ambiguous.

1) Parse Context: Extract manual test file list, output_dir, codebase_root from contextStore 2) Discover Infrastructure: Detect shared infrastructure files:

• tests/manual/config.sh -- shared configuration
• tests/manual/test_harness.sh -- shared test framework (if exists)
• tests/manual/test-all.sh -- master runner
• tests/manual/TEMPLATE-*.sh -- test templates (if exist)
• tests/manual/regenerate-golden.sh -- golden file regeneration (if exists)

3) Scan Scripts (Layer 1): For each manual test script, check 7 quality dimensions (see Audit Rules) 3b) Context Analysis (Layer 2 -- MANDATORY): For each candidate finding, ask:

• Is this a setup/utility script (e.g., 00-setup/*.sh, tools/*.sh)? Setup scripts have different requirements -- skip harness/golden checks
• Is this a master runner (test-all.sh)? Master runners orchestrate, not test -- skip all checks except fail-fast
• Does the project not use a shared harness at all? If no test_harness.sh exists, harness adoption check is N/A

4) Collect Findings: Record violations with severity, location (file:line), effort, action, recommendation 5) Calculate Score: Count violations by severity, calculate compliance score (X/10) 6) Write Report: Build full markdown report in memory per references/templates/audit_worker_report_template.md, write to {output_dir}/ln-636--global.md in single Write call 7) Return Summary: Return minimal summary to coordinator (see Output Format)

Audit Rules

1. Harness Adoption

What: Test script uses shared framework (run_test, init_test_state) instead of custom assertion logic

Detection:

• Grep for run_test, init_test_state in script
• If absent AND script contains custom test loops/assertions -> custom logic
• If test_harness.sh does not exist in project -> skip this check entirely

Severity: HIGH (custom logic = maintenance burden, inconsistent reporting)

Recommendation: Refactor to use shared run_test from test_harness.sh Action: REWRITE_MANUAL_EVIDENCE

Effort: M

2. Golden File Completeness

What: Test suite has expected/ directory with reference files matching test scenarios

Detection:

• Check if suite directory has expected/ subdirectory
• Compare: number of test scenarios (grep run_test calls) vs number of expected files
• If test uses diff against expected files but expected dir is missing -> finding

Layer 2: Not all tests need golden files. Tests validating HTTP status codes, timing, or dynamic data may legitimately skip golden comparison -> skip if test has no diff or comparison against files

Severity: HIGH (no golden files = no regression detection for output correctness)

Recommendation: Add expected/ directory with reference output files Action: REWRITE_MANUAL_EVIDENCE

Effort: M

3. Config Sourcing

What: Script sources shared config.sh for consistent configuration

Detection:

• Grep for source.*config.sh or . .*config.sh
• If absent -> script manages its own BASE_URL, tokens, etc.

Layer 2: If script is self-contained utility (e.g., tools/*.sh) -> skip

Severity: MEDIUM

Recommendation: Add source "$THIS_DIR/../config.sh" for shared configuration Action: REWRITE_MANUAL_EVIDENCE

Effort: S

4. Fail-Fast Compliance

What: Script uses set -e and returns exit code 1 on failure

Detection:

• Grep for set -e (or set -eo pipefail)
• Check that failure paths lead to non-zero exit (not swallowed by || true everywhere)

Severity: HIGH (silent failures mask broken tests)

Recommendation: Add set -e at script start, ensure test failures propagate Action: REWRITE_MANUAL_EVIDENCE

Effort: S

5. Template Compliance

What: Script follows project test templates (TEMPLATE-api-endpoint.sh, TEMPLATE-document-format.sh)

Detection:

• If TEMPLATE files exist in tests/manual/, check structural alignment:
• Header comment block with description, ACs tested, prerequisites
• Standard variable naming (THIS_DIR, EXPECTED_DIR)
• Standard setup pattern (source config.sh, check_jq, setup_auth)
• If NO templates exist in project -> skip this check entirely

Layer 2: Older scripts written before templates may diverge. Flag as MEDIUM, not HIGH

Severity: MEDIUM

Recommendation: Align script structure with project TEMPLATE files Action: REWRITE_MANUAL_EVIDENCE

Effort: M

6. Idempotency

What: Script can be rerun safely without side effects from previous runs

Detection:

• Grep for cleanup patterns: trap.*EXIT, rm -f, cleanup functions
• Check for temp file creation without cleanup
• Check for hardcoded resource names that would conflict on rerun (e.g., creating user with fixed email without checking existence)

Layer 2: Scripts that only READ data (GET requests, queries) are inherently idempotent -> skip

Severity: MEDIUM

Recommendation: Add cleanup trap or use unique identifiers per run Action: REWRITE_MANUAL_EVIDENCE

Effort: S-M

7. Documentation

What: Test suite directory has README.md explaining purpose and prerequisites

Detection:

• Check if suite directory (NN-feature/) contains README.md
• If missing -> finding

Layer 2: Setup directories (00-setup/) and utility directories (tools/) may not need README -> skip

Severity: LOW

Recommendation: Add README.md with test purpose, prerequisites, usage Action: REWRITE_MANUAL_EVIDENCE

Effort: S

Scoring Algorithm

MANDATORY READ: Load references/audit_scoring.md.

Severity mapping:

• Missing harness adoption (when harness exists), No golden files (when expected-based), No fail-fast -> HIGH
• Missing config sourcing, Template divergence, No idempotency -> MEDIUM
• Missing README -> LOW

Output Format

MANDATORY READ: Load references/templates/audit_worker_report_template.md.

Write JSON summary per references/audit_summary_contract.md. In managed mode the caller passes both runId and summaryArtifactPath; in standalone mode the worker generates its own run-scoped artifact path per shared contract.

Write report to {output_dir}/ln-636--global.md with category: "Manual Evidence" and checks: harness_adoption, golden_file_completeness, config_sourcing, fail_fast_compliance, template_compliance, idempotency, documentation.

Return summary per references/audit_summary_contract.md.

When summaryArtifactPath is absent, write the standalone runtime summary under .hex-skills/runtime-artifacts/runs/{run_id}/evaluation-worker/{worker}--{identifier}.json and optionally echo the same summary in structured output.

Report written: .hex-skills/runtime-artifacts/runs/{run_id}/audit-report/ln-636--global.md
Score: X.X/10 | Issues: N (C:N H:N M:N L:N)

Critical Rules

Apply the already-loaded references/audit_worker_core_contract.md.

• Do not auto-fix: Report only
• Effort realism: S = <1h, M = 1-4h, L = >4h
• Skip when empty: If no tests/manual/ directory exists, return score 10/10 with zero findings
• Exclude non-test files: Skip config.sh, test_harness.sh, test-all.sh, regenerate-golden.sh, TEMPLATE-*.sh, files in tools/, results/, test-runs/
• Context-aware: Setup scripts (00-setup/) have relaxed requirements (no golden files, no harness needed)
• Unique angle: Audit only manual test evidence. Do not judge automated test value, E2E priority, product behavior, missing coverage, trustworthiness, oracle strength, or structure.
• Action required: Use REWRITE_MANUAL_EVIDENCE for findings. Summarize compliant scripts as KEEP_MANUAL_EVIDENCE in the report summary, not as findings.

Definition of Done

Apply the already-loaded references/audit_worker_core_contract.md.

• [ ] contextStore parsed successfully (including output_dir)
• [ ] Manual test infrastructure discovered (config.sh, harness, templates)
• [ ] All 7 checks completed per test script
• [ ] Layer 2 context analysis applied (setup/utility exclusions)
• [ ] Findings collected with severity, location, effort, action, recommendation
• [ ] Score calculated using penalty algorithm
• [ ] Report written to {output_dir}/ln-636--global.md (atomic single Write call)
• [ ] Summary written per contract

Reference Files

• Audit output schema: references/audit_output_schema.md

Version: 1.0.0 Last Updated: 2026-03-13