| name | token-meter-verification |
| description | Use when Token Meter changes or claims need independent source, test, runtime, native, browser, privacy, or platform verification. |
Token Meter Verification
Purpose
Independently verify the task envelope's acceptance criteria against one exact head.
Do not modify tracked files or repair failures while acting as tester.
Build the verification matrix
Read specs/AGENTS.md, the task envelope, developer result, current diff, and
.agents/workflow/review-policy.yaml. Select checks from the surfaces the diff can
affect:
| Surface | Evidence |
|---|
| Python and contracts | Focused tests, full suite when proportionate, and py_compile |
| Dashboard | Embedded-JavaScript parse plus live wide-desktop and 1024-pixel-laptop interaction checks |
| Native menu | Swift compile, deterministic smoke output, and live menu behavior |
| Packaging | Installer syntax, manifest coverage, installed-runtime revision, and source/runtime parity |
| HTTP behavior | /health, relevant endpoint payloads, cache/privacy boundaries, and failure states |
| Cross-platform | Target-host parse, launch, interaction, update, and uninstall evidence |
Run focused checks first, then the broader checks required by risk. Synthetic fixture
passes do not replace real upstream, installed-runtime, or visible behavior where the
acceptance criteria depend on them.
Evidence rules
- Record the command or interaction, outcome, and inspected head.
- Use
passed, failed, blocked, or not-run; never turn missing evidence into a
pass or measured zero.
- Separate source checks from live installed behavior.
- Confirm source/runtime parity after installation rather than assuming the checkout
is what the user sees.
- Preserve exact target-host limitations. macOS evidence does not prove Windows or
Linux behavior.
- Stop and report a failure; return the change to the developer instead of fixing it.
Tester result
Return acceptance-criteria coverage, commands and live checks, counts where useful,
the exact head, failures, not-run checks, unverified behavior, blockers, and the
recommended next state. A later tracked code or test change invalidates the result.