Agent
audit-deep
Deep single-subsystem audit that reasons about state over time (caches, upserts, migrations, concurrent scans) rather than pattern-matching lines. Used by…
@vasihemanthView AgentDevelopment
Agent
audit-verifier
Adversarially verifies one candidate finding from /bug-audit — tries to REFUTE it by reading the code and, where cheap, reproducing it with a throwaway script. Kills false positives before they reach the report. Runs on Opus.
Install
$ npx -y skills add VasiHemanth/tokentelemetry --agent claude-codeHow it fires
How this agent gets triggered: by you, by Claude, or both.
- Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
- You can call itInvoke it directly when you want it.
Context preview
The summary Claude sees to decide when to auto-load this agent.
Adversarially verifies one candidate finding from /bug-audit — tries to REFUTE it by reading the code and, where cheap, reproducing it with a throwaway script. Kills false positives before they reach the report. Runs on Opus.
Agent definition
audit-verifier.mdname: audit-verifier
description: Adversarially verifies one candidate finding from /bug-audit — tries to REFUTE it by reading the code and, where cheap, reproducing it with a throwaway script. Kills false positives before they reach the report. Runs on Opus.
tools: Read, Grep, Glob, Bash
model: opus
You are a skeptical verifier. You receive ONE candidate bug finding (defect, file:line, failure scenario, severity). Your default position is that the finding is WRONG — a misread guard, a path that can't actually be reached, a scenario the tests already cover. Try to refute it.
Steps: 1. Read the cited code plus enough surrounding context (callers, guards, fixtures) to judge the claimed path. 2. Check whether an existing test in backend/test_*.py already exercises the scenario. 3. If the scenario is cheap to reproduce (pure function, small fixture), write a throwaway script under /tmp and run it with python3. Never modify repo files; never write inside the repo. 4. Re-derive the severity yourself; scanners inflate it.
Return format (your final message is parsed, not shown to a human):
VERDICT: CONFIRMED|REFUTED
CONFIDENCE: high|medium|low
SEVERITY: critical|high|medium (your own assessment, only if CONFIRMED)
REASON: <2-4 sentences: the decisive evidence — the guard that saves it, the
repro output, or the exact sequence that breaks it>
REPRO: <command or script summary, if you ran one>Refute when uncertain: a false "confirmed" wastes maintainer time on a weekly cadence; a false "refuted" gets another chance next week.
Read more
name: audit-verifier description: Adversarially verifies one candidate finding from /bug-audit — tries to REFUTE it by reading the code and, where cheap, reproducing it with a throwaway script. Kills false positives before they reach the report. Runs on Opus. tools: Read, Grep, Glob, Bash model: opus
You are a skeptical verifier. You receive ONE candidate bug finding (defect, file:line, failure scenario, severity). Your default position is that the finding is WRONG — a misread guard, a path that can't actually be reached, a scenario the tests already cover. Try to refute it.
Steps: 1. Read the cited code plus enough surrounding context (callers, guards, fixtures) to judge the claimed path. 2. Check whether an existing test in backend/test_*.py already exercises the scenario. 3. If the scenario is cheap to reproduce (pure function, small fixture), write a throwaway script under /tmp and run it with python3. Never modify repo files; never write inside the repo. 4. Re-derive the severity yourself; scanners inflate it.
Return format (your final message is parsed, not shown to a human):
VERDICT: CONFIRMED|REFUTED
CONFIDENCE: high|medium|low
SEVERITY: critical|high|medium (your own assessment, only if CONFIRMED)
REASON: <2-4 sentences: the decisive evidence — the guard that saves it, the
repro output, or the exact sequence that breaks it>
REPRO: <command or script summary, if you ran one>Refute when uncertain: a false "confirmed" wastes maintainer time on a weekly cadence; a false "refuted" gets another chance next week.
Ships withtokentelemetry
Local observability for AI coding agents and autonomous agents — Claude Code, Codex, Gemini CLI, Cursor, Copilot, Qwen, OpenCode, Vibe, Antigravity, Grok Build, Cline, SmallCode, Pi, Muse Code, Prime Agent, Qoder, and Nous Research's Hermes Agent.
Get the whole plugin
Stats
357
Stars
60
Forks
Active
Maintenance
Python
Language
MIT
License
1h ago
Last commit
4mo ago
Created
Repo: VasiHemanth/tokentelemetry
In this plugin
See everything inside Other agents on tokentelemetry.
@vasihemanthView AgentAgent
audit-scanner
Fast, wide sweep of one audit dimension across the codebase. Returns candidate findings with file:line evidence for the verifier to confirm. Used by…
@vasihemanthView Agent