Verify — evidence before "done"

"Verified" is not a feeling; it's output you can point at. Never claim passing without showing it (R5 + verification-before-completion).

When this fires

"verify" / "is this done / shippable / ready to merge?" / just before a commit or PR.

Runtime neutrality

The verification runner and evidence standard are client-neutral. Never infer the active agent from this skill's install path: portable canonical content may be copied into a required client adapter such as .claude/skills. Use canonical AGENTS.md when an instruction is relevant.

Fast path — if the Agentsmith CLI is available

1. Run agentsmith verify (or the installed-project shim .agentsmith/agentsmith verify on macOS/Linux and .agentsmith\\agentsmith.cmd verify on Windows). It runs every phase in .harness/verify.conf and stops at the first failure). --list shows the phases; --only <tag> iterates just one. Add --record <directory> --tree-class <clean-clone|disposable-fixture|linked-worktree|operator-worktree> when deterministic command evidence needs a durable local receipt. The explicit class keeps a dirty operator tree from masquerading as clean-checkout evidence; clean-clone refuses a dirty tree. The destination must be new; relative paths resolve from the project target. Record mode redacts secret-shaped console and sidecar output before either is emitted.
2. On a failure: read the label + command it printed, explain in plain language what broke, and point at that phase's line in .harness/verify.conf to fix or refine.
3. Report the actual pass/fail output — not a summary of intent.

The v1 receipt covers deterministic commands only. Keep screenshots, videos, manual assertions, published results, and real runtime or visual checks as separately referenced artifacts. A green receipt does not satisfy an end-to-end gate that requires those observations.

Fallback — no runner or no conf

1. Say so plainly. With the CLI but no conf, agentsmith verify discover --save <plan> proposes phases from the project; review the plan, then agentsmith verify apply --plan <plan>.
2. Run the obvious checks for this project directly and show their output: build, test, lint, link/render check, a dry-run. For non-code work, "verify" = open the artifact and confirm it renders / the numbers reconcile / the links resolve.
3. If nothing is wired yet, propose the phases this project needs and offer to write .harness/verify.conf — but still run the checks by hand this time.

Output you can quote as proof

• Run it raw. An output-rewriting hook (a token-saving shell wrapper on Bash, for example) can summarise, empty or re-code test, lint, search and git output, and report its own exit status. Run any command whose output you quote through the wrapper's bypass (for rtk: rtk proxy <command>). agentsmith doctor reports such hooks.
• Absence needs an untruncated, count-based check. "No match" from a truncated or summarised search is not proof that something is absent.
• CI is green when pending == 0 and failing == 0 on the matching head SHA, read from step conclusions. A cancelled, missing or not-yet-registered check is not green; don't wait for a fixed number of checks.

Report

State each check and its evidence: "build ok, 42 tests green, lint clean" — with the output, not "should pass". Anything skipped is "deferred: reason", never silence.