What counts as evidence in ai-agents and how to produce it. Covers the TESTING-RIGOR pos+neg+edge bar, test layout and collection reality, coverage proof commands, runtime-contract tests with negative controls, and ADR-034 QA skip semantics at session end. Use when you say `what counts as evidence`, `how do I test this change`, `run skill tests`, `can I skip QA`. Do NOT use for CI failure triage (use `ai-agents-debugging-playbook`) or measurement tooling (use `ai-agents-diagnostics-toolkit`).
What counts as evidence in ai-agents and how to produce it. Covers the TESTING-RIGOR pos+neg+edge bar, test layout and collection reality, coverage proof commands, runtime-contract tests with negative controls, and ADR-034 QA skip semantics at session end. Use when you say `what counts as evidence`, `how do I test this change`, `run skill tests`, `can I skip QA`. Do NOT use for CI failure triage (use `ai-agents-debugging-playbook`) or measurement tooling (use `ai-agents-diagnostics-toolkit`).
version
1.0.0
license
MIT
ai-agents Validation and QA
This repo runs verification-based governance: a claim is true when a gate, test, or command output says it is, not when an agent asserts it. This skill defines the evidence bar for code changes, where tests live and how they are collected, the anti-pattern canon (each entry paid for by a real incident), and the QA gate semantics at session end.
Audience: a zero-context contributor (human or model) about to write, run, or skip tests in this repository.
Probe an external tool's runtime behavior empirically
ai-agents-empirical-probe-toolkit
Understand which gates a change class triggers
ai-agents-change-control
Session-end mechanics beyond the QA row
session-end skill
Process
Phase 1: Internalize the evidence bar
.agents/governance/TESTING-RIGOR.md is BLOCKING for code changes (its own Status line, TESTING-RIGOR.md:5). The bar, per function you add or change:
Requirement
Concretely
Positive test
Valid input produces the expected output
Negative test
Invalid input produces the language's idiomatic error, asserted
Edge tests
Whitespace, empty, None, wrong type
Every error branch
Each raise / error-return path exercised
Every conditional branch
Including branches that only change user-facing strings
Mocked I/O
No live subprocess, API, or file dependencies in unit tests
CLI contract
argv-failure exits, exit codes, stdout vs --output tested
Coverage proof
100% block coverage on changed files (see Phase 4)
Origin story: in PR #1756, the original 20 unit tests gave 24% block coverage; adding negative, edge, and branch tests raised it to 100% and caught real defects the happy-path tests missed (whitespace handling on verdict matching, conditional OTHER-hint emission, type validation). That review created TESTING-RIGOR.md (TESTING-RIGOR.md:3-16). The rule exists because bots (Copilot, CodeRabbit, Gemini) reliably catch what happy-path tests skip, at roughly 10x the cost of writing the tests up front (TESTING-RIGOR.md:81).
Coverage targets by risk tier (AGENTS.md Standards; TESTING-ANTI-PATTERNS.md:112-118): 100% security-critical, 80% business logic, 60-70% docs/glue. "Security-critical" includes secret handling, input validation, command execution, path sanitization, auth checks.
Quality trumps quantity: .agents/governance/TESTING-ANTI-PATTERNS.md bans coverage theater (assertion-free tests), brittle mocks for impossible scenarios, unit-tests-as-only-testing, quantity over quality, and testing entirely after the fact. A test that never failed during development may not be testing anything (TESTING-ANTI-PATTERNS.md:91).
Phase 2: Know where tests live and how they are collected
Tests for a skill's Python scripts (importable via conftest sys.path)
uv run pytest tests/skills/github/ -x
.claude/skills/NAME/tests/
NO
Per-skill structure tests, colocated with the skill (claude-agents.md MUST 3)
uv run pytest .claude/skills/NAME/tests/ -q
Two skill-test locations is a real fork, not a typo: structure tests colocate under .claude/skills/NAME/tests/ and are NOT collected by uv run pytest tests/ -x; behavior tests for skill scripts land in tests/skills/NAME/ and are. CI pins only the second location (.github/workflows/pytest.yml:214-222 runs tests/skills/github/test_wait_for_unresolved_zero.py); no workflow collects .claude/skills/*/tests/ at all, so those files run only when you name them explicitly (issue #3593, .agents/retrospective/2026-07-28-autopilot-issue-burndown.md:72-74). When you add a skill, you usually need both (Phase 6).
Useful invocations, all verified (as of 2026-07-03):
uv run pytest tests/ -x # full default suite, stop on first failure
uv run pytest tests/test_ai_review.py -x # one file
uv run pytest tests/ -m unit # by marker
uv run pytest .claude/skills/prose-self-check/tests/ -q # one skill's colocated tests
uv run pytest .claude/skills/NAME/tests/ --collect-only -q # prove tests are discoverable without running
Markers (pyproject.toml [tool.pytest.ini_options].markers): unit, integration, safe_push_transport, security, smoke, windows_path. safe_push_transport means the test touches a non-local transport and is excluded from pre-push. smoke means real-CLI tests needing auth/credits, nightly only; the smoke gate asserts they were not skipped (issue #2231 item 4). windows_path means the test exercises Windows path handling and must run on a Windows runner. Always uv run pytest, never bare pytest or python3 -m pytest outside the venv: PyYAML and friends live in the uv venv (see ai-agents-build-and-env).
Stale doc warning: .agents/governance/test-location-standards.md still describes a Pester/*.Tests.ps1 layout. Zero *.Tests.ps1 files exist (as of 2026-07-03; ADR-042 Python migration). Trust the table above and pyproject.toml, not that file.
Phase 3: Write tests that count
Beyond pos+neg+edge, this repo demands five specific disciplines:
1. Isolation from the real repo. The root conftest.py (repo root, lines 315-386) fails any test that moves the REAL repo HEAD (issue #2316): every git mutation must run in a tmp_path repo with cwd= that repo. Supporting fixtures in tests/conftest.py: GIT_CONFIG_COUNT injection neutralizes host commit.gpgsign so tmp-repo commits work in signing environments (issue #2548, tests/conftest.py:26-61), and AI_AGENTS_PROJECT_REPO=1 defaults identity for guards that check the origin remote (issue #2610, tests/conftest.py:64-76). Consumer-repo simulation tests override that env var to "0".
2. Runtime-contract tests for generated artifacts (FM-11). A generated artifact that has never been executed is not done (FAILURE-MODES.md:28, index row 11; the #2205 incident wedged every plugin customer for 33 days). .claude/rules/generated-artifacts.md:67-73 requires: execute the shipped artifact under the host's real contract (foreign cwd, host-set env vars), assert the intended effect, and include a negative control proving the test CAN fail (a bare relative path must fail the same harness). The exemplar is tests/build_scripts/test_generate_hooks_runtime_contract.py: see test_negative_control_bare_relative_path_fails and test_anchor_is_load_bearing_when_no_plugin_root_var_set.
3. Negative controls beat self-reference. A test that string-matches the generator's own output passes when the generator is consistently wrong. The first #2205 fix shipped exactly this (retro 2026-06-02-pr-2205-customer-wedge-incident.md:49,83) and it hid two more defects. Every contract test needs a case where the wrong artifact fails.
4. Threshold detectors need calibration. Any threshold-based signal (rework count, thread count, file count) must be replayed against roughly the last 5 real merged PRs and shown to fire correctly before shipping. PR #1989's M4 warning shipped with threshold 6 in a repo whose PRs maxed at 4 file edits: it could never fire (retro 2026-05-10-pr-1989-recursive-failure.md:151-157). A detector that cannot fire on recent real work is not calibrated.
5. Guards run on their own branch. A PR shipping a pre-push or pre-commit guard must show that guard running against its own branch, terminal output in the PR description (retro 2026-05-10-pr-1989-recursive-failure.md:131-137; M5 was never applied to the PR that shipped it).
Phase 4: Prove coverage
100% block coverage on changed files, with only defensive-unreachable exclusions (# pragma: no cover) plus written justification (TESTING-RIGOR.md:53).
uv run pytest tests/test_ai_review.py tests/test_verdict.py tests/test_quality_gate.py --cov=scripts.ai_review_common.verdict --cov-branch --cov-fail-under=100
Two traps, both fossilized in .github/workflows/pytest.yml comments:
Trap
Rule
Evidence
File-set sensitivity
A --cov-fail-under=100 pin must run EVERY test file that exercises the module. After a test-file split, running one file alone reported 63% and tripped the gate
Issue #1963; pytest.yml:196-205
Coverage target form
Use the module-name form (--cov=wait_for_unresolved_zero), never the file-path form. File paths produce "Module never imported" + 0% with pytest-cov 7.x on Python 3.14
Issue #2063, tested in PR #2078; pytest.yml:207-222
Related discipline, FM-10: there is no neutral default for a missing signal (FAILURE-MODES.md:387). The neutral default is UNKNOWN, not PASS (extract_verdict on no-match, merge_verdicts on empty; get_verdict fails safe to CRITICAL_FAIL). PR #1965 still took 3 fix rounds to make UNKNOWN blocking everywhere it flowed: the workflow verdict list, the action parser allowlist, and the action exit-code mapping. When testing parsers or gates, always include the missing-signal case and assert it raises or blocks, never that it silently passes.
Phase 5: QA gate semantics at session end
Session protocol Phase 2.5 makes QA validation BLOCKING before committing feature code (SESSION-PROTOCOL.md:738-756). Skip classes (ADR-034, .agents/architecture/ADR-034-investigation-session-qa-exemption.md):
Evidence string
When valid
Staged files limited to
SKIPPED: docs-only
Strictly editorial doc edits: no code, config, tests, workflows, or code blocks changed
Documentation files
SKIPPED: investigation-only
No code/config changes at all
The 8 patterns in scripts/modules/investigation_allowlist.py (single source of truth per its docstring): .agents/sessions/, .agents/analysis/, .agents/retrospective/, .serena/memories/, .agents/security/, .agents/memory/ (incl. episodes/), .agents/architecture/REVIEW-*, .agents/critique/. ADR-034:78-87 lists the same 8 after the 2026-07-08 amendments (issues #831 and #732), so the ADR and the code agree. SESSION-PROTOCOL.md:754 still lists only the first 5, so treat the module and ADR-034 as authoritative
Explicitly NOT skippable: .agents/planning/, .agents/architecture/ADR-* (ADRs; REVIEW-* files ARE allowlisted by the enforcement module), .github/, .claude/agents/, src/, scripts/ (ADR-034 "Not Allowed" table, which predates the module's wider allowlist). Session logs, analysis artifacts, and memory updates are audit trail, not implementation: they are filtered out automatically when deciding whether QA is required, so they can ride along with implementation commits (SESSION-PROTOCOL.md:758).
Mixed session (investigation turned into code)? Split it: commit investigation work with the skip evidence, then start a new session for the code change with real QA (SESSION-PROTOCOL.md:798-800). Abusing skip markers is a trust failure with precedent; escape-hatch history lives in ai-agents-config-catalog and ai-agents-failure-archaeology.
Full evidence line in the session log looks like "qaValidation": { "level": "MUST", "Complete": true, "Evidence": "SKIPPED: investigation-only" } or a QA report path under .agents/qa/ (SESSION-PROTOCOL.md:996).
Phase 6: Add tests for a new skill, hook, or script
Skill: create .claude/skills/NAME/tests/test_skill_structure.py (frontmatter, sections, size, dash checks; copy the shape from .claude/skills/prose-self-check/tests/). Required by .claude/rules/claude-agents.md:18 (MUST 3).
Skill scripts with behavior: add tests/skills/NAME/test_SCRIPT.py so the default suite collects them; add __init__.py and a conftest.py for sys.path if importing the script by module name.
Hook: add tests/hooks/test_NAME.py; use tests/hook_test_helpers.py; cover exit 0 (allow, stdout context), exit 2 (block, stdout message), and malformed-stdin input.
Validation/build script: add tests/test_NAME.py or tests/build_scripts/test_NAME.py; test exit codes per ADR-035 (0 ok, 1 logic, 2 config, 3 external, 4 auth).
Generated artifact: add a runtime-contract test with a negative control (Phase 3, item 2).
All git-mutating tests isolated in tmp_path repos (the #2316 guard will fail you otherwise).
Prove collection: uv run pytest PATH --collect-only -q shows your tests; then run them; then run the coverage pin form from Phase 4 on changed files.
Run uv run python scripts/validation/pre_pr.py before the PR (the local shift-left aggregate; see ai-agents-change-control for the full gate ladder).
Anti-Patterns
Each row cost real time. Do not re-earn these lessons.
Anti-pattern
Incident
Binding rule
Self-referential test (asserts generator output against itself)
First #2205 fix shipped one; it passed while the artifact was broken (retro :49, :83)
Runtime-contract test + negative control (generated-artifacts.md:67-73)
Happy-path-only test suite
PR #1756: 20 tests, 24% coverage, bots caught the rest
TESTING-RIGOR.md pos+neg+edge, BLOCKING
Threshold detector never calibrated
#1989 M4: threshold 6, repo max 4, could never fire
Calibration table against last ~5 real PRs before commit
Guard not run on its own branch
#1989 M5: bot-cascade hook shipped but never applied to its own PR
Guard output on the shipping branch in the PR description
Test mutates the real repo
Repo-root conftest.py:315-386 (#2316)
Isolate in tmp_path, run git with cwd= the tmp repo
TESTING-ANTI-PATTERNS.md 1: each test answers a stakeholder concern
Trusting the Pester test-location doc
test-location-standards.md predates ADR-042; zero .Tests.ps1 files remain
Use Phase 2 table + pyproject.toml [tool.pytest.ini_options].testpaths
Skipping QA on a mixed session
ADR-034 allowlist exists precisely to fence this
Split the session; skip evidence only with allowlisted paths staged
Verification
Before claiming a change meets the evidence bar:
Every new/changed function has pos, neg, and edge tests; every error and conditional branch exercised
External I/O mocked in unit tests; CLI exit codes tested where a CLI changed
uv run pytest tests/ -x green locally, plus explicit runs for any .claude/skills/NAME/tests/ you touched
Coverage proven at 100% on changed files using the module-name --cov form
Generated artifacts have a runtime-contract test with a negative control
Any new threshold or guard shows calibration/self-application evidence in the PR description
QA row in the session log carries a report path or a legal SKIPPED: evidence string with allowlisted staged files
Provenance and Maintenance
Verified 2026-07-29 against the working tree (issue #3828 re-verification pass; the earlier 2026-07-03 pass had rotted for pyproject.toml, both conftest.py files, .github/workflows/pytest.yml, and .claude/rules/generated-artifacts.md). Sources: .agents/governance/TESTING-RIGOR.md:3-55,77-81, .agents/governance/TESTING-ANTI-PATTERNS.md:9-118, pyproject.toml [tool.pytest.ini_options], repo-root conftest.py:315-386, tests/conftest.py:19-76, .github/workflows/pytest.yml:196-222, .claude/rules/generated-artifacts.md:67-73, .claude/rules/claude-agents.md:18, .agents/architecture/ADR-034-investigation-session-qa-exemption.md:62-110, .agents/SESSION-PROTOCOL.md:738-800,996, .agents/governance/FAILURE-MODES.md:14-30,387, .agents/retrospective/2026-05-10-pr-1989-recursive-failure.md:110-157, .agents/retrospective/2026-06-02-pr-2205-customer-wedge-incident.md:23-83.
Re-verify volatile facts:
grep -n testpaths pyproject.toml # collection roots
sed -n '315,386p' conftest.py # #2316 HEAD guard still present
grep -n "cov-fail-under" .github/workflows/pytest.yml # coverage pins and forms
grep -n "SKIPPED" .agents/SESSION-PROTOCOL.md | head -5 # QA skip evidence stringsls tests/skills/ .claude/skills/prose-self-check/tests/ # both skill-test locations alive
git ls-files '*.Tests.ps1' | wc -l # Pester doc still stale if 0
If TESTING-RIGOR.md, ADR-034, or the pytest.yml pins change, update the matching phase here in the same PR.
