test-gate

Run the full verification stack against a project: verify command, golden principles, quality checks, and regression detection. Short-circuits on first failure. Returns structured per-gate results.

This skill is stateless and read-only — it never modifies files. The caller (build-loop or operator) handles file management like baseline promotion, progress.yaml updates, and commit decisions.

Parameters

• project — path to the project root (absolute, or relative to cwd; required)
• verify — the shell command to run for verification (optional; if omitted, read from AGENTS.md ## Verify section)
• phase-goal — path to phase-goal.md for golden principles extraction (optional; if omitted, Gate 2 is skipped)
• baseline — path to the previous test baseline file (optional; if omitted, Gate 4 is skipped)
• current-out — path where current verify output should be saved by the caller (optional; used only by Gate 4 as the "after" snapshot — if omitted, Gate 4 is skipped)
• suppressions — comma-separated list of quality check names to skip in Gate 3 (optional)

How it works

1. Resolve inputs.
   • Resolve project to an absolute path.
   • If verify is not provided, read the project's AGENTS.md and extract the verify command from ## Verify. If no verify command is found, report FAIL with "No verify command configured" and stop.
2. Gate 1 — Verify.
   • Run the verify command from the project root.
   • Capture stdout+stderr.
   • Determine pass/fail from exit code (0 = pass, non-zero = fail).
   • If FAIL: record the gate result, output the structured report, stop (short-circuit).
   • If PASS: hold the captured output — it is used by Gate 4 as the "current" snapshot.
3. Gate 2 — Golden Principles.
   • If phase-goal is not provided: record SKIP. Proceed to Gate 3.
   • If provided: read the file, extract bullet points under ### Golden principles (phase-relevant).
   • For each principle, check whether the current uncommitted changes (from git diff and git diff --cached) violate it.
   • If any violation is found: record FAIL with the specific principle violated. Output the structured report, stop.
   • If no violations: record PASS. Proceed to Gate 3.
4. Gate 3 — Quality Checks.
   • Read the project's AGENTS.md for a ## Quality checks section. If no such section exists: record SKIP. Proceed to Gate 4.
   • Always run root-agents-immutable first (non-optional). If it fails, record FAIL, stop.
   • For each configured check: if the check name appears in suppressions, skip it.
   • Run each non-suppressed check following the verify-gate check library (same logic as /sabs:verify-gate). Do not duplicate the check definitions — follow the procedures described in the verify-gate skill (${CLAUDE_PLUGIN_ROOT}/skills/verify-gate/SKILL.md).
   • If any check fails: record FAIL with the specific check name and findings. Output the structured report, stop.
   • If all checks pass: record PASS. Proceed to Gate 4.
5. Gate 4 — Regression Detection.
   • If baseline is not provided, or current-out is not provided, or the baseline file does not exist: record SKIP.
   • Otherwise: compare the baseline file against the Gate 1 verify output using the regression-detect logic (same procedure as /sabs:regression-detect). Parse both files for test results, diff the maps, check for regressions.
   • If regressions are found: record FAIL with the regressed test names. Output the structured report, stop.
   • If no regressions: record PASS.
6. Output. Produce the structured result (see Output format below).

Output format

## Test Gate Results — <project name>

Gates run: N
Passed: N
Failed: N
Skipped: N

### GATE 1: Verify — PASS | FAIL
<if FAIL: first 20 lines of verify output showing the failure>
<if PASS: summary line — "Exit code 0. N tests passed." or first summary line from output>

### GATE 2: Golden Principles — PASS | FAIL | SKIP
<if FAIL: "Violated: <principle text>" with brief explanation>
<if SKIP: "No phase-goal provided — skipped.">
<if PASS: "No violations detected.">

### GATE 3: Quality Checks — PASS | FAIL | SKIP
<if FAIL: per-check results — "FAIL: <check-name>" with findings>
<if SKIP: "No quality checks configured in AGENTS.md — skipped.">
<if PASS: "N checks passed." with check names listed>

### GATE 4: Regression Detection — PASS | FAIL | SKIP
<if FAIL: list of regressed tests — "was PASS, now FAIL">
<if SKIP: "No baseline provided — skipped." or "No baseline file exists — skipped.">
<if PASS: "No regressions. N tests compared.">

### Verdict: PASS | FAIL
<if FAIL: "Failed at Gate N: <gate name>. Subsequent gates not run.">
<if PASS: "All gates passed.">

Gates that were never reached (due to short-circuit) are omitted from the output entirely — they are not marked SKIP.

Rules

• Stateless. Never modify files. Never update progress.yaml. Never commit. The caller is responsible for all state management.
• Short-circuit on first failure. Gates run in order 1->2->3->4. On first FAIL, remaining gates are not run.
• Verify-gate reuse. Gate 3 follows the same check library and procedures as /sabs:verify-gate. If verify-gate evolves (new checks, changed behavior), Gate 3 inherits the changes by reading the same skill file.
• Regression-detect reuse. Gate 4 follows the same diff logic as /sabs:regression-detect. Same parsing rules, same verdict criteria.
• No file writes. The verify output captured in Gate 1 is held in memory for Gate 4. The caller saves it to disk if needed (e.g., build-loop saves to test-baseline-current.txt on pass).

Integration with build-loop

The build-loop's Branch A steps 4-7 currently contain inline verification logic. Test-gate consolidates this into a single call:

Build-loop invocation (future):

/sabs:test-gate project=<PROJECT_ROOT> phase-goal=docs/plan/phase-goal.md baseline=docs/plan/test-baseline.txt current-out=docs/plan/test-baseline-current.txt

Failure-to-fix-task mapping (build-loop reads the structured output):

The build-loop passes suppressions from quality_check_suppressions in progress.yaml. The repeated-failure counting (3 consecutive quality check failures -> suggest suppression) remains in the build-loop since it is orchestration state.

Standalone usage

Minimal — runs Gate 1 and Gate 3 only (Gates 2 and 4 skipped):

/sabs:test-gate project=<path>

Full — all four gates:

/sabs:test-gate project=<path> verify="npm test" phase-goal=docs/plan/phase-goal.md baseline=docs/plan/test-baseline.txt current-out=docs/plan/test-baseline-current.txt