# Verify

> Have Grok independently verify a change in an existing worktree (no source edits, hermetic)

- **Type:** Skill
- **Install:** `agentstack add skill-sfourdrinier-grok-skills-verify`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [sfourdrinier](https://agentstack.voostack.com/s/sfourdrinier)
- **Installs:** 0
- **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [sfourdrinier](https://github.com/sfourdrinier)
- **Source:** https://github.com/sfourdrinier/grok-skills/tree/main/plugin/skills/verify

## Install

```sh
agentstack add skill-sfourdrinier-grok-skills-verify
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

## How to run (transparent)

1. Take the absolute **Base directory for this skill** from the Skill tool
   (the folder that contains this skill's `SKILL.md` and `run.mjs`).
2. Set `SKILL_BASE` to that path. Do **not** invent versioned cache paths.
3. Always invoke the companion **only** through this skill's runner:

```bash
SKILL_BASE=''
# Required for completion notifications (plugin/references/execution-context.md):
export GROK_COMPANION_EXECUTION_CONTEXT=foreground   # or background
node "$SKILL_BASE/run.mjs"  [args...]
```

`run.mjs` finds the plugin install from its own location and runs
`scripts/grok-companion.mjs`. No `CLAUDE_PLUGIN_ROOT` / `PLUGIN_ROOT` required.

If the host already exported `CLAUDE_PLUGIN_ROOT` or `PLUGIN_ROOT`, you may call
`node "$CLAUDE_PLUGIN_ROOT/scripts/grok-companion.mjs"` instead; prefer
`"$SKILL_BASE/run.mjs"` whenever the Skill tool loaded this skill.

Return companion **stdout verbatim**. Never put free-text in `--task "..."`;
use `--task-file -` with a single-quoted heredoc.

Run a Grok `verify` through the hardened wrapper and relay its result envelope.
`verify` inspects and tests a change in an EXISTING worktree (typically one a
prior `code` run produced), has no source-editing tools, and always ends with a
machine-readable `pass` / `fail` / `inconclusive` verdict plus evidence.

Raw slash-command arguments:
`$ARGUMENTS`

Required wrapper flags (copy exactly, substitute only placeholder values):
- `--worktree ` is required.
- Exactly one of `--task ` or `--task-file ` is required.
- Preserve the user's arguments exactly. Do not strip, add, or reorder flags.
  Do not invent a flag that is not in the argument-hint.
- Shell-injection safety for `--task `: the task is free text you must NEVER
  place in a shell-evaluated position. `$(...)`/backticks inside a double-quoted
  `--task "..."` run locally BEFORE the wrapper validates them. When the arguments
  carry a `--task `, deliver that text on STDIN with `--task-file -` and a
  SINGLE-QUOTED heredoc so the shell passes it byte-for-byte; the companion stages
  it into a temp file for the wrapper.
- Shell-injection safety for flag VALUES (`--worktree`, a `--task-file `,
  `--model`, `--timeout`, `--max-turns`, and EVERY other value you substitute from
  `$ARGUMENTS`): wrap each substituted value in SINGLE quotes, for example
  `--worktree ''`. Single quotes stop the shell from evaluating
  `$(...)`/backticks, so a hostile value reaches the companion as one literal argv
  token and the wrapper validates it (worktree path resolution + escape guards).
  An unquoted OR double-quoted value would be command-substituted locally BEFORE
  the wrapper ever sees it -- the same injection class as an unsafe `--task "..."`.

NO web access:
- `verify` NEVER accepts `--web` - independent verification stays hermetic by
  design. If the user asks for `--web` on a verify run, do NOT add it; explain
  that verify is hermetic and offer `reason --web` or `review --web` instead.

Run it as one Bash call and relay the result. When the arguments carry a
`--task `, route that text through STDIN so it is never shell-evaluated:
```bash
export GROK_COMPANION_EXECUTION_CONTEXT=foreground
node "$SKILL_BASE/run.mjs" verify --worktree '' [other non-task flags from $ARGUMENTS, each substituted value single-quoted] --task-file - 
GROK_TASK
```
When the arguments already use `--task-file `, drop the heredoc and pass
every flag as single-quoted argv tokens:
```bash
export GROK_COMPANION_EXECUTION_CONTEXT=foreground
node "$SKILL_BASE/run.mjs" verify --worktree '' --task-file '' [other non-task flags from $ARGUMENTS, each substituted value single-quoted]
```
- Return the command stdout envelope VERBATIM. Do not paraphrase, summarize, or
  add commentary before or after it. Preserve the exit status. The verdict lives
  in the envelope's `verifier.verdict`; do not restate it as your own conclusion
  in place of the raw envelope.

If the companion prints an actionable "could not locate the Grok wrapper"
message instead of an envelope, tell the user to run `/grok:setup`.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [sfourdrinier](https://github.com/sfourdrinier)
- **Source:** [sfourdrinier/grok-skills](https://github.com/sfourdrinier/grok-skills)
- **License:** MIT

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-sfourdrinier-grok-skills-verify
- Seller: https://agentstack.voostack.com/s/sfourdrinier
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.