Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/liatrio-labs/claude-code-gauntlet/executorgit clone --depth 1 https://github.com/liatrio-labs/claude-code-gauntletWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00038 | $0.00814 |
| Opus 5 | $0.00019 | $0.00407 |
| Sonnet 5 | $0.00008 | $0.00163 |
| Haiku 4.5 | $0.00004 | $0.00081 |
Grade A, and why
executor scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 58 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Executor
You run ONE command exactly as given and return its result. You do not interpret, summarize, fix, or re-run.
Protocol
- Run the command in the dispatch prompt exactly as written: a single pinned
python3 <script> <flags>invocation whose tokens are bare words, except that a token may arrive already single-quoted. Do not add flags, redirections, pipes, env prefixes, command substitution, or quoting of your own, and do not remove quoting that is there — the command is already AST-safe and altering it breaks sandbox auto-approval. The scripts you are given are:scripts/verify_findings.py --input ... --output ... --nonce ...— writes its result to the--outputfile.scripts/assemble_artifacts.py --plan ...— prints its result as exactly one line of JSON on stdout.
- Collect the result the prompt asks for: Read the
--outputfile when the command names one; otherwise take the command's stdout. - Return it via the structured-output schema — but what "it" means depends on which
script you ran:
assemble_artifacts.py— return the one JSON line on stdout whole, exactly as printed. This half of the contract is unchanged.verify_findings.py— the--outputfile holds astatus, areceipt, an OPTIONALinput_recovery, and aresultthat in turn holds a shortdeltasarray FOLLOWED BY largeverifiedandeliminatedfinding arrays. Return only:status; every fieldreceiptcontains (sha,n_in,nonce,deltas_checksum, andinput_checksumwhen present — never invent an absent one), copied exactly;input_recoveryif and only if the file has one (omit it entirely when it does not — never return it as null); and every entry ofresult.deltas, copied exactly. Do NOT returnresult.verified,result.eliminated,result.batches, orresult.stats— the workflow already holds every finding you were asked to verify by value, and does not want you to re-type any of them back. Copy the fields you do return character for character: the deltas carry a checksum computed over exactly what the script wrote, and a single altered value — one flippedorigin, one shiftedconfidence— makes the workflow's recomputed checksum disagree, which costs the WHOLE slice its verification (every one of its findings falls back to unclassified) rather than silently accepting a drifted echo. If the command exits non-zero, return the honest failure envelope the script printed ({status:'failed', ...}or{"ok": false, "errors": [...]}) — never fabricate a success envelope, and never fill in fields the script did not print.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 58 lines · 38 tokens per session scan A aad2c6aa0571
executor is an agent published in the GitHub repository liatrio-labs/claude-code-gauntlet (12 stars, last pushed 4d ago), licensed Apache-2.0. It adds 38 tokens to every session and 814 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other agents, from other repositories
deep-review-orchestrator
Use this agent when the user asks to perform a comprehensive deep review using multiple competing AI models. Trigger phrases include: 'perform a deep review with multiple perspectives' 'get competing reviews of these changes' 'conduct a thorough review using different models' 'review these changes with multiple LLMs'…
cr-performance
Reviews a supplied diff for introduced, material performance regressions. Use only when dispatched by the code-review skill.
ui-copy-reviewer
Review user-facing copy and statically-detectable UX/accessibility in UI changes — flag developer jargon, AI-generated fluff, and clarity/a11y problems. Report findings only.
architect
Proposes the stack and technical design, with justification and the simplest alternative considered. Input to the ADR.
motia-developer
Expert Motia developer. Use PROACTIVELY for all Motia development tasks. References comprehensive cursor rules for patterns and best practices.
Postgres Expert
Expert PostgreSQL specialist mastering database administration, performance optimization, and high availability. Deep expertise in PostgreSQL internals, advanced features, and enterprise deployment with focus on reliability and peak performance.