Atlas / Skills / google / Mantis Meta Agent

Mantis Meta AgentSAFE

skills/google/mantis-meta-agent

A modular, stack-agnostic toolkit for AI coding agents to autonomously find, reproduce, and patch vulnerabilities.

Verdict
SAFE
Grade
B
Trust score
89 /100
Version
—
Hosts
2 documented
License
Apache-2.0
Stars
2,539
01

Overview

A modular, stack-agnostic toolkit for AI coding agents to autonomously find, reproduce, and patch vulnerabilities.

Read from source at commit 9702f6eef495OBSERVED · 2026-10-26
Advertisement
02

Host compatibility

What the documentation claims. We have not run a compatibility test.

HostStatusNotes
claude-codementioned
gemini-climentioned
03

What it tells the agent

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: mantis-meta-agent
description: >-
  Acts as the persistent supervisor, launching and monitoring the automated review campaign.
  Use when running a long-running, continuous security review campaign that needs autonomous coordination.
  Don't use for executing individual review stages directly.
---

# Meta-Agent Orchestrator (/mantis-meta-agent)

## System Goal

Autonomous Campaign Manager. Supervises the continuous review loop, ensures
resilience, and monitors long-running security pipelines.

## Command Definition

- **Command:** `/mantis-meta-agent`
- **Description:** Acts as the persistent supervisor, launching and monitoring
  the automated review campaign.
- **Parameters:**
  - `--sync`: OPT-IN boundary sync + snapshot pinning. When present (or when the
    user explicitly instructs a sync for this pass), synchronize the target with
    its upstream at the START of the pass (Block C) and pin an immutable
    per-pass snapshot (Block D). When ABSENT (the default), do NOT sync and do
    NOT pin: leave `active_snapshot`/`snapshot_history` unwritten, leave
    `snapshot_pinned` unset, and run the pass against the live tree exactly as
    today (single static snapshot; `--target_root` semantics unchanged).
  - `--snapshot_keep=<N>`: retention for pinned snapshots under
    `<state_root>/.mantis_snapshots/` (Block D step 6 GC). DEFAULT 2 (this pass
    \+ previous), safe because correctness never re-reads an older snapshot;
    raise it for patch/PoC rebasing against a finding's discovery snapshot.
  - `--state_root=<absolute path>`: directory that contains `workspace/` and
    `.mantis_snapshots/`. DEFAULT: the absolute path of the current working
    directory at meta-agent launch (i.e., the parent of `workspace/`). If
    `state_root` resolves inside `CODE_ROOT` (the colocated case — the default
    when launching from inside the target repo), Step 3 (PIN) MUST (a) HALT and
    yield to the user with a message to pass `--state_root` outside `CODE_ROOT`
    — this is the DEFAULT and the only safe behavior without explicit opt-in.
    Auto-relocate (option b) is opt-in ONLY via an explicit
    `--auto_relocate_state` flag (or an equivalent user instruction to
    auto-relocate for this pass); without it, never move user state. When the
    user has explicitly opted in, auto-relocate by moving `workspace/` and
    `.mantis_snapshots/` to a sibling directory outside `CODE_ROOT` and updating
    `state_root` for the remainder of the pass. This is the "colocated state"
    transition that `mantis-pipeline-adapter` Scenario 1 requires. The
    relocation is a MOVE (not copy-then-delete) and is skipped if
    `workspace/.workspace_edit.lock` is held (INV-5: no user data loss). The
    MOVE MUST be atomic on the same filesystem: move into a fresh temp dir under
    the sibling's parent, then `rename(2)` to the final sibling path; refuse and
    HALT if the sibling already exists or the temp dir cannot be created (never
    clobber an existing directory). If `rename(2)` fails, leave the source
    untouched and HALT — do NOT fall back to copy-then-delete.
  - `--auto_relocate_state`: OPT-IN flag. When `state_root` resolves inside
    `CODE_ROOT` (colocated state — the default when launching from inside the
    target repo), the default behavior is to HALT and yield to the user. Passing
    this flag (or giving an equivalent user instruction) authorizes the
    atomic-MOVE auto-relocate path described under `--state_root` above. Without
    it, never move user state.

## Input/Output Contract

- **Reads**:
  - `workspace/.mantis_state.json` (to track current loop pass; and to read
    `active_snapshot`/`snapshot_history` for snapshot provenance and Block E
    PREV-reference).
  - Individual JSON findings in `workspace/findings/` (to verify states between
    subagent transitions, including each finding's `patch_status` for the
    pre-sync backup-hygiene gate).
  - `workspace/.workspace_edit.lock` (GATE A of the pre-sync backup-hygiene
    step: must be FREE before deleting stale `*.bak-*` files).
  - `CODE_ROOT/.mantis_snapshot_id` sentinel (the snapshot reuse/STOP check in
    Block D step 0).
- **Writes**:
  - Creates archive directory `workspace/archive/findings_pass_N/` and moves
    finding JSON files and `.trash/` to it; and (Stage 15) COPIES
    `workspace/.mantis_state.json` and `workspace/kb/` (incl. `THREAT_MODEL.md`)
    into it as the pass boundary reference.
  - Updates `workspace/.mantis_state.json`: increments `pass_number`, updates
    `last_updated`, and (always) refreshes `vcs_info`. When `--sync` was
    requested this pass, also writes `active_snapshot`
    (`{root, snapshot_id, snapshot_pinned, pass, vcs_type}`) with
    `snapshot_pinned` reflecting reality (`true` when pinned, `false` in HALT);
    `snapshot_history` is appended by Block D step 5 (one entry per pass:
    `{pass, snapshot_id, snapshot_pinned, timestamp}`) — Step 4 RECORD below
    does NOT re-append. These new keys are defined in `schema.json`
    `#/$defs/state`; they are optional and absent in MODE-OFF (no `--sync`).
  - When `--sync` pins a snapshot (Block D): materializes an immutable snapshot
    copy under `<state_root>/.mantis_snapshots/pass_<N>/` (deliberately OUTSIDE
    any `/workspace/` path segment, so mantis-patch's existing state-vs-code
    path guard still treats it as CODE, not state), writes the
    `CODE_ROOT/.mantis_snapshot_id` sentinel, and `chmod -R a-w` the copy.
- **Preconditions**:
  - Target project must be identified. Campaign orchestration tools/subagents
    must be ready.
- **Idempotency Guarantee**:
  - Skips the finding move if `workspace/findings/` is empty or missing (the
    state/KB boundary copy still runs). Determines pass number dynamically once
    from disk if the state file is missing to avoid overwriting existing
    archives during new runs.
  - Snapshot pinning is crash-safe: Block D reuses an already-materialized
    snapshot whose sentinel matches t
04

Trust audit

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

LayerWhat it checksResult
L0Provenance & inventoryPASS
L1Static analysis of the codeNA
L2Instruction surface (what it tells the agent)PASS
L3Class-specific surfacePASS
L4Behavioural (sandbox)SKIPPED

What the source does

Filesystem
none-observed
Network
none-observed
Shell
none-observed
Dependencies
pinned
Secrets in source
none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.

Audited 2026-10-26 · audit v0.4.1 · source sha 9702f6eef495full audit observations/trust-audit/skill/google__mantis-meta-agent.json · Report an issue / request a re-scan
05

Audit history

Every audit this skill has had.

DateSourceVerdictGradeScoreChange
2026-10-269702f6eef495SAFEB89first audit
06

Questions

What does the mantis-meta-agent — google/mantis skill do?

A modular, stack-agnostic toolkit for AI coding agents to autonomously find, reproduce, and patch vulnerabilities.

Is mantis-meta-agent — google/mantis safe to install?

The audit found nothing in the source that contradicts what it says it does, and graded it B (89/100). Grade A is held back for packages that have also passed a sandboxed behavioural run, which is why a clean skill reads B.

What can mantis-meta-agent — google/mantis access on my machine?

The audit observed no filesystem, network or shell use at all in its source.

Which assistants does mantis-meta-agent — google/mantis work with?

Its documentation mentions claude-code and gemini-cli. That is what the text claims, not a compatibility test we ran.

How current is this page?

The grade is for one exact copy of the source (9702f6eef495), read on 2026-10-26. The repository is watched, and a new audit runs when it changes — this is the first audit.

Add this badge to your README

Shows this page's audit result and links back to it. It updates when the audit does.

ModelsAtlas audit badge

Markdown
[![ModelsAtlas audit: Mantis Meta Agent](https://modelsatlas.com/skills/google/mantis-meta-agent/badge.svg)](https://modelsatlas.com/skills/google/mantis-meta-agent/)

Not your listing, or something out of date? Claim or update your listing.

Advertisement