Runtime verification

Runtime Verification

Domain skill for verifying code works at runtime, beyond static analysis and unit tests.

Iron Law

NO COMPLETION UNTIL THE CODE BUILDS, RUNS, AND BEHAVES CORRECTLY. If you cannot verify it yourself, build the infrastructure to verify it.

A green test suite is necessary but not sufficient. Runtime verification proves code actually works. "No test framework" is a problem to solve, not a reason to skip.

Skip Whitelist (Enumerated — No Subjective Exemptions)

Runtime verification is MANDATORY for every change. The only permitted skips are the three categories below. Any skip outside this whitelist is forbidden and must be escalated via the Proactive-Autonomy protocol (see below).

Skip Category	Definition	Required Evidence to Claim
markdown-only	The diff touches only .md, .markdown, .txt, or .rst files. Zero code, config, or data files.	git diff --name-only origin/$DEFAULT_BRANCH..HEAD output showing only doc extensions.
config-only	The diff touches only configuration files (.json, .yaml, .yml, .toml, .ini, .env.example, dotfiles) with no executable code path changes. Config syntax must still be validated (lint/schema check, dry-run apply).	Full file list plus syntax validation output.
dependency-bump-only	The diff touches only lockfiles and manifest version strings (e.g., package.json version fields, package-lock.json, poetry.lock, Gemfile.lock, go.sum, Cargo.lock) with no source code, no config semantics, and no new dependencies. Build must still succeed.	Full file list plus successful build output.

If the diff mixes any whitelisted category with anything else (a single .py or .ts file, a new dependency, a config value change that alters behavior), the skip is disallowed. Run full runtime verification.

If In Doubt, Run It

If you are uncertain whether the change qualifies for a whitelist skip — run runtime verification. Uncertainty is never a reason to skip. The cost of an extra verification run is small; the cost of shipping unverified code is large.

Forbidden reasoning patterns include "small change", "CI-only edit", "tests already cover this", "I read the diff and it looks safe", "just a refactor", and similar. None of those are whitelist categories. If your reasoning does not map cleanly to markdown-only, config-only, or dependency-bump-only with the explicit evidence shown in the table above, you MUST run verification.

Escalation Protocol for Out-of-Whitelist Skips

If you genuinely believe a skip outside the whitelist is warranted (e.g., infrastructure-only change, generated-code-only change, or something the whitelist does not yet cover), you MUST NOT proceed silently. Raise a Proactive-Autonomy escalation per references/escalation-format.md using AskUserQuestion and receive an explicit approval before proceeding. The Situation field MUST cite the specific change and why the whitelist does not cover it; the Tried field MUST list the standard paths attempted (fast-path verify script, build, smoke tests). Blanket "always skip for this repo" authorization is never valid — each out-of-whitelist skip requires its own escalation.

Fast-Path Verification

Check for a project-level verify script first:

[ -x "verify.sh" ] && echo "FAST_PATH: verify.sh found"
[ -x "scripts/verify.sh" ] && echo "FAST_PATH: scripts/verify.sh found"

If found, run it and return results. Skip remaining steps.

Build Verification

Mandatory build step for all project types. Build failure IS the finding — do NOT skip to runtime checks.

# Node.js / TypeScript
[ -f "package.json" ] && npm run build 2>&1

# Python
[ -f "setup.py" ] || [ -f "pyproject.toml" ] && pip install -e . 2>&1

# Go
[ -f "go.mod" ] && go build ./... 2>&1

# Rust
[ -f "Cargo.toml" ] && cargo build 2>&1

# Ruby
[ -f "Gemfile" ] && bundle install 2>&1

If build fails → iterate: read errors, fix, rebuild (up to closedLoop.maxBuildIterations, default 5). Do NOT proceed until the build passes.

LSP Diagnostics Verification

Pre-check: Read lsp.enabled from settings (default true). If false, skip this section entirely.

When LSP is available and lsp.diagnosticsAsQuality is enabled in settings, collect language server diagnostics as an additional quality signal. This complements — never replaces — CLI-based quality commands.

Process

1. Identify all files changed on the branch:

   git diff --name-only origin/$DEFAULT_BRANCH..HEAD
   

2. For each changed source file, use LSP(documentSymbol) to confirm the language server recognizes the file, then collect any diagnostics reported.

3. Map diagnostics to findings:

   LSP Severity	Finding Priority	Action
   Error	P1	Must fix before proceeding
   Warning	P2	Should fix
   Information/Hint	P3	Consider

4. Deduplicate against CLI tool output — if the same issue is reported by both LSP and a CLI tool (e.g., tsc and TypeScript LSP), keep only one entry.

Timeout Handling

Each LSP operation must complete within lsp.timeout (default 5000ms from settings). If an operation times out:

• Mark that file's LSP check as "Timeout — skipped"
• Continue with remaining files
• Note timeout in output table
• Never block the workflow on a slow LSP server

Graceful Fallback

If no LSP server is available, skip this section entirely with note: "LSP diagnostics: N/A — no language server configured." This is not an error and does not affect the verification outcome.

Ad-Hoc Verification

For projects without formal test frameworks, verify by running the code:

Project Type	Verification Approach
Backend/API	Start server, curl endpoints, verify responses, check logs
CLI tools	Build, run with --help, run with sample input, check exit codes
Libraries	Write temporary test script, exercise public API, verify outputs, delete script
Static sites	Build, serve locally, verify pages load
Config-only	Validate config syntax, apply dry-run if supported

"No test framework" is a problem to solve, not a reason to skip verification.

Iterative Debug Loop

When any verification fails:

1. Read the FULL error message and stack trace — don't skim
2. Identify root cause (not just the symptom)
3. Fix the root cause
4. Re-verify

If the same error persists after a fix attempt: re-read code paths, try a different approach. Max closedLoop.maxDebugIterations (default 5) iterations, then escalate to user.

The user should NEVER have to provide logs or tell you what went wrong. You have access to the same errors — read them yourself.

Discovery (dev server, port, E2E framework)

The capability-discovery skill (invoked at the start of every verify-relevant command) already detects tech stack, dev server scripts, and E2E frameworks. Re-running discovery here is redundant; consume the existing output.

When operating standalone (no prior capability-discovery invocation), use these probes:

# Dev server: CLAUDE.md hints, package.json scripts, framework config files
[ -f ".claude/CLAUDE.md" ] && grep -iE "(dev|server|start|serve):" .claude/CLAUDE.md
[ -f "package.json" ] && python3 -c "import json; d=json.load(open('package.json')); [print(f'{k}: {v}') for k,v in d.get('scripts',{}).items() if k in ('dev','start','serve')]"

# Port: running listeners on common ports
lsof -i -P -n 2>/dev/null | grep LISTEN | grep -E ':(3000|4000|5000|8000|8080)' | head -5

# E2E framework: config files
[ -f "playwright.config.ts" ] || [ -f "playwright.config.js" ] && echo "Playwright"
[ -f "cypress.config.ts" ] || [ -f "cypress.config.js" ] && echo "Cypress"

Smoke Tests

If a dev server is running, perform basic health checks:

curl -s -o /dev/null -w "%{http_code}" http://localhost:$PORT/health
curl -s -o /dev/null -w "%{http_code}" http://localhost:$PORT/

A non-200 response is a P1 finding emitted via the canonical schema in references/finding-schema.md with category=runtime.

Visual Verification (extracted to its own skill)

Visual verification has been extracted to skills/visual-verification/SKILL.md. That skill owns the screenshot-analyze-verify loop, the browser-tool priority cascade (Playwright MCP → Chrome DevTools MCP → CLI → external skill fallback), the responsive viewport checks, and the result vocabulary (PASS / FAIL / SKIP / SKIP_WARN / SKIP_USER_APPROVED / MANUAL / BLOCKED).

When a diff is UI-relevant (UI file extensions OR acceptance criteria with UI keywords — see the visual-verification skill for the detection rules), the consumer (commands/start.md Phase 4, commands/pr.md Phase 4) invokes BOTH skills in parallel:

Skill(runtime-verification)  # build + server + smoke + E2E + LSP diagnostics
Skill(visual-verification)   # browser-rendered UI loop

The two skills coordinate via the dev server URL: if runtime-verification cannot start the dev server, visual-verification returns SKIP with reason "dev server unavailable" and the completion gate treats the dev-server failure as the primary finding rather than emitting a SKIP_WARN that obscures the root cause.

Acceptance Criteria Verification

Map each acceptance criterion to a verification method:

Criterion Type	Verification
API behavior	curl/fetch endpoint, check response
UI rendering	Screenshot-analyze-verify loop (see Visual Verification)
UI responsive	Multi-viewport screenshot verification
Data processing	Run with test data, check output
Configuration	Verify config loads without error

Completion Gate

Runtime verification is complete only when:

• Every testable acceptance criterion has a verification result (Pass/Fail/N/A with reason)
• "N/A" is justified (e.g., no dev server for a CLI tool), never used as a shortcut
• Failed checks are reported as P1 findings, not silently noted

If the dev server won't start, that IS the finding. Report it.

Output Format

### Runtime Verification Results

| Check | Status | Details |
|-------|--------|---------|
| Dev server | {Running/Not found} | Port {N} |
| Health check | {Pass/Fail/N/A} | HTTP {status} |
| E2E tests | {Pass/Fail/N/A} | {framework} |
| Smoke tests | {Pass/Fail/N/A} | {details} |
| LSP diagnostics | {Pass/Fail/Skip/N/A} | {errors: N, warnings: N, files checked: N} |

### Acceptance Criteria Verification
| # | Criterion | Verified | Method |
|---|-----------|----------|--------|

When the consumer also invokes Skill(visual-verification), the visual rows (Visual check, Responsive, Console errors, Visual Evidence table) are produced by that skill — see its Output Format section. The consumer renders the runtime-verification table and the visual-verification table side by side rather than merging them, so each skill's output remains attributable.

Timeout Configuration

From settings.json:

• timeouts.devServerStartup: Max seconds to wait for dev server (default: 30)
• timeouts.e2eTest: Max seconds for E2E suite (default: 120)

visualVerification.* settings (including maxIterations) are owned by Skill(visual-verification); see that skill's documentation when both run together.

Active Problem Solving

Do NOT silently skip verification. Actively solve problems:

Problem	Action
No dev server	Attempt to start one. Report P1 if no start command exists and no alternative verification is possible.
No E2E framework	Run ad-hoc smoke tests (curl endpoints, run CLI, exercise API)
Server won't start	Read the error, fix the code, retry (up to closedLoop.maxServerRetries, default 3)
Port already in use	Try alternative ports
LSP server unavailable / times out	Skip LSP diagnostics; do NOT block on them. Note "LSP diagnostics: N/A — {reason}" in the output table.

Browser-tool problems (no Playwright, no Chrome DevTools, no npx) are handled by the visual-verification skill — see its "Active Problem Solving" section.