Verification

Reference skill read by /work, /ship, /design-build, /hypothesis-debugging, and /quiver:review at their verification step. It is never invoked and never runs a command itself: the consumer resolves the command by the rules here, runs it, and quotes the evidence.

Command Resolution

test: <command> | none (<reason>)
build: <command> | none (<reason>)
source: docs | stack | none

The consumer resolves once, before the first task runs, records these three lines, and passes the resolved command to every subagent as a literal. none is a value, not a gap to fill: it carries its reason in the parentheses and is never promoted to a guess. Test and build resolve independently under the same rules; source names the rule that produced the test command.

Apply the rules in order. The first rule that yields a command decides source.

1. Project instructions. A command named in the project's CLAUDE.md, AGENTS.md, or equivalent agent-instruction file, in a Testing or Build section. When present it wins and source is docs.
2. Stack detection. The table below, matched on files at the project root. source is stack.
3. None. source is none. Write the reason in the parentheses; never substitute a guess.

A command from rule 1 or 2 is confirmed before it runs: the script or task is declared in the manifest, and the tool is on PATH (command -v). A command that fails confirmation resolves to none with the reason, not to the next rule.

Cross-cutting rules

1. Exit 0 is not evidence. The summary line must show at least one test executed; otherwise the outcome is skipped.
2. Compare exit codes as zero versus nonzero, never == 1. pytest and unittest exit 5 when nothing ran, the Dart test runner exits 79, cargo exits 101, and mocha exits the failure count.
3. Never let a runner wait. CI=true on every Node script, vitest run in place of a bare vitest script, and the consumer's own wait rules for cold builds.
4. A command is runnable only when the tool is on PATH and the script or task is declared. Otherwise it is none.
5. A failure whose first error line names a service, port, or environment variable (a docker-compose.yml, .env.example, or DATABASE_URL in test config is the usual signal) is reported verbatim and flagged as environmental. It is not counted as a code failure and not retried.

Evidence Rule

• Pass is this run's exit code plus the runner's summary line, from a run made after the last edit. When the runner prints no summary, the last output line stands in.
• Fail is the first failing test name plus the first error line. More output belongs in a report file, not in the evidence line.
• Skipped carries its reason: none from resolution, a failed confirmation, or a zero-test run. It never counts as a pass; a consumer that requires a pass treats skipped as not passed.
• No evidence means not run. "Should pass", "ran earlier", and "the change is safe" are not outcomes.
• A command that has not returned is not a pass. Waiting for it is the consumer's business.

Consumers quote the outcome in one of these line formats:

pass:    <command> -> exit 0: <summary line>
fail:    <command> -> exit <code>: <first failing test> -- <first error line>
skipped: <reason>

For Skill Authors

• Read this file at the verification step, not at skill start.
• Resolve once per run in the controller and pass the resolved command into any subagent prompt as a literal (Test command: npm test), never as "detect the test command".
• Require the evidence line format back from a subagent, and paste the ### Subagent restatement below into its prompt verbatim. It is the only text from this file that is copied anywhere, and tests/skills/test-verification-contract.sh compares the copies to it.
• Do not restate the resolution table. When a stack is missing, add a row here.

Subagent restatement

Run the test command you were given exactly as written, and report a pass only by quoting this run's exit code and the runner's summary line -- a run that executed zero tests, a command that has not returned, or a result remembered from an earlier run is not a pass. On failure, quote the first failing test name and the first error line.