verify · mag123c — Claude Code skill: what it does, how to install
Runs a self-healing Rust verification loop using cargo test, clippy, and fmt. It prevents false passes by enforcing actual test execution before review.
npx skills add mag123c/toktrack --skill verify -a claude-codeWhat verify does
The skill automates Rust project checks in a strict sequence: cargo test --quiet, cargo clippy --all-targets --all-features -- -D warnings, and cargo fmt --all -- --check. When a step fails, the agent inspects compiler or linter errors, modifies the source code, and retries. If the exact same failure occurs 3 times in a row, the loop stops to notify the user.
Verification honesty rules prevent false green reports. Skipped tests marked #[ignore] or runs blocked by missing external services count as unresolved blockers instead of passes. Mock-only coverage must be explicitly flagged in the report as unconfirmed integration.
The workflow depends entirely on an active Rust toolchain with Cargo. Once all checks run honestly without blockers, the agent immediately triggers /review without asking for confirmation.
What it helps with
- run Cargo test, Clippy, and formatting checks before committing
- fix compiler and linter errors automatically in a retry loop
- stop unrun or ignored tests from masquerading as verified passes
- trigger the
/reviewstep right after all verification gates clear
Example prompts for the agent
- Run the verify loop on this crate before making a commit
- Check the Rust codebase and fix any clippy warnings automatically
- Verify test execution and report unresolved blockers
How to install verify
1Skills CLI (Vercel)
npx skills add mag123c/toktrack --skill verify -a claude-code2Manually: put the files in place
~/.claude/skills/verify/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/verify/
1Skills CLI (Vercel)
npx skills add mag123c/toktrack --skill verify -a codex2Manually: put the files in place
.agents/skills/verify/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add mag123c/toktrack --skill verify -a cursor2Manually: put the files in place
.cursor/skills/verify/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install mag123c/toktrack verify2Manually: put the files in place
.github/skills/verify/SKILL.md1Skills CLI (Vercel)
npx skills add mag123c/toktrack --skill verify -a gemini-cli2Platform CLI
gemini skills install https://github.com/mag123c/toktrack --path .claude/skills/verifyManually: put the files in place
.devin/skills/verify/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add mag123c/toktrack --skill verify -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about verify
How to install verify in Claude Code?
Run npx skills add mag123c/toktrack --skill verify -a claude-code in your terminal.
What does verify do?
It executes cargo test, cargo clippy, and cargo fmt --check in order. When tests or linter checks fail, the agent repairs the code and retries the process.
What happens when an error persists during self-healing?
After 3 identical failures, the agent halts retries and escalates the issue directly to the user.
Source text
SKILL.md · MIT
---
name: verify
description: Self-healing verification loop (test → clippy → fmt)
required_context: []
---
# Verify
## Flow
```
cargo test → cargo clippy → cargo fmt --check
│ │ │
└── On fail: fix and retry (notify user after 3 same failures)
```
## Commands
```bash
cargo test --quiet
cargo clippy --all-targets --all-features -- -D warnings
cargo fmt --all -- --check
```
## Verification Honesty (MUST — PASS gate)
Verification trustworthiness lies not in "what passed" but in **"what was *actually executed* and passed"**. Enforced before any PASS verdict.
### Skip ≠ Pass
- Tests marked `ignored` / `#[ignore]` are **unverified** → do not count as green.
- Tests skipped due to environment (missing external service, platform constraints, etc.) are a **BLOCKER**. If unfixable, escalate to user as "verification incomplete + reason" rather than PASS.
### Do not cite unrun tests as verification evidence
- "Confirmed via test X" is only valid when **that test actually ran and passed in this execution**.
- Do not use a merely written test as evidence of "confirmed" (false-green = false reassurance).
### Coverage honesty (mock ≠ real)
- Mock-based unit tests verify **only the logic inside the mock boundary**. Behaviors invisible to mocks — file I/O, concurrency, simd-json real parsing — are **confirmed only by integration verification**.
- If a path is covered only by mocks, report it as **"mock coverage only — integration unconfirmed"** rather than "verified".
### Acceptance/DoD cross-check (carry PLAN gates)
- Read the PLAN Acceptance/DoD and any `[unverified-gate]` passed from implement; cross-check whether each item **actually ran and passed in this execution**.
- If unrunnable (prod smoke / external service / CI-only environment) → no green-wash. Tag as `[unverified-gate: delegated, owner/due/probe]` and **carry to wrap Plan Reconciliation** (`../wrap/references/reconciliation.md`).
### Report format
- ✅ Actually executed and passed
- ⛔ Skipped/environment-blocked + **reason** (unresolved = BLOCKER, not PASS)
- ⚠️ Mock coverage only, integration unconfirmed
- ⏳ Delegated to wrap due to unverifiable Acceptance item
## Self-Healing
- Fail → analyze error → fix code → retry
- Same error 3 times → notify user
## Rules
- Required before commit
- Order: test → clippy → fmt
- All must pass to proceed
## Next Step
**Only when the Verification Honesty gate passes** (actual execution passed, no unresolved BLOCKERs) → call `/review` immediately. Do not ask "should I review?".
If there are unresolved BLOCKERs, **do not proceed to /review** — report "verification incomplete + reason + required environment actions" to the user.
© mag123c · MIT · mag123c/toktrack
Similar skills
All skills for Testing and QA →verify by other authors comes first.
