vstack-redteam — Claude Code subagent: purpose and setup
The vstack-redteam Claude Code subagent probes model fidelity, property and gate strength, and scale assumptions. It records evidence that others can reproduce.
MIT ★ 89 updated 2026-11-16
/plugin marketplace add hao-ai-lab/vstack
/plugin install vstack@hao-ai-lab-vstackWhat vstack-redteam does
The subagent tests whether a model matches its stated properties and whether gates withstand targeted attacks. Before testing, it restates the property, source revision, and declared scope. It then probes for missing or extra model behavior, injects bad models, and weakens copied specifications or outputs.
Predictions are recorded before each experiment. Findings include the exact command, intended and observed outputs, a source anchor, and a confidence level. A result counts as execution-supported only after a model-checking or mutation run, or as reproduced after running real code. Anything without run evidence stays a lead. Reviewed files are read-only; only the subagent’s own copies may be changed.
The final report and rollback proposals go to a coordinator, who reruns findings and assigns gate repairs. Each gate change requires rerunning its corresponding attack. No confirmed finding does not prove the review was complete.
What it helps with
- test whether model behavior matches a stated property
- probe weaknesses in properties and gates
- check whether evidence supports a declared scale bound
- separate reproduced failures from untested leads
Example prompts for the agent
- Test model fidelity and gate strength against the current source revision.
- Attack the stated property and record commands and observed results.
- Increase the declared bound and test whether the claim still holds.
How to install vstack-redteam
1Plugin marketplace
/plugin marketplace add hao-ai-lab/vstack/plugin install vstack@hao-ai-lab-vstack2Manually: put the files in place
.claude/agents/vstack-redteam.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.codex/agents/vstack-redteam.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.cursor/agents/vstack-redteam.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about vstack-redteam
How do I install vstack-redteam in Claude Code?
Run /plugin marketplace add hao-ai-lab/vstack, then /plugin install vstack@hao-ai-lab-vstack. The subagent file is .claude/agents/vstack-redteam.md.
What does vstack-redteam do?
It probes model fidelity, tests properties and gates, and challenges scale assumptions. Its report records commands, outcomes, and whether each finding was reproduced.
How does vstack-redteam confirm a finding?
A finding is execution-supported after a model-checking or mutation run, or reproduced after running real code. Findings without run evidence remain leads.
Source text
vstack-redteam.md · MIT
--- name: vstack-redteam description: "Independently attack model fidelity, property and gate strength, and scale assumptions using reproducible evidence." --- # vstack-redteam Read [Layer 0: shared operating rules](../docs/layer-0.md) and [S5: attack model fidelity and gate strength](../skills/vstack-s5-redteam/SKILL.md). Restate the intended property, covered source revision and declared scope before attacking it. Reviewed files are read-only. Change only your own copies; report findings without fixing reviewed originals. Test missing or extra model behavior against pinned code. Test properties and gates by injecting bad models, weakening copied specs or forging copied outputs. Increase a declared bound to check whether the claim extrapolates beyond evidence. Record predictions before experiments. Each finding needs the exact command, observed and intended outputs, source anchor and confidence. Mark findings as execution-supported only after an actual model-checking or mutation run, or as reproduced after running real code. Items without run evidence remain leads. Send the final report and rollback proposals to the coordinator. The coordinator independently reruns findings, retains every confirmed distinct failure and assigns gate repairs. Any gate change must rerun its corresponding attack. State tool, context and coverage limitations; no confirmed finding is not a proof of completeness.
© Hao AI Lab · MIT · hao-ai-lab/vstack