Help wanted — verify ripwire's openclaw integration on a real openclaw

ripwire ships openclaw support as of dddd5171. **Nobody maintaining this repo runs openclaw**, so every
claim below is verified against docs.openclaw.ai and openclaw's own source, and **none of it has been run
against a live openclaw**. This issue exists so that gap is stated in public rather than discovered by
someone installing.

If you have openclaw, this costs one command:

```bash
bash scripts/verify-agent-integration.sh openclaw
```

Phase 1 is automatic and installs nothing into your real home — it runs the recipe, feeds the installer
the exact flag the recipe printed, and checks the skills land and every symlink resolves, all inside a
throwaway `HOME`. Phase 2 prints two prompts to paste into openclaw and a filled-in markdown block to
paste back here.

**A FAIL is the useful outcome.** It is the only way we learn that a recipe we publish does not work.

## The four claims least likely to survive contact with reality

1. **Context file.** We tell users to paste the wiring into `~/.openclaw/workspace/AGENTS.md` — openclaw's
own workspace bootstrap file, NOT the repository's `AGENTS.md`. An earlier draft of the row said the
repo's, which would have sent people to edit a file openclaw never reads. Is the workspace path right?
2. **Skills root and its condition.** We say skills go to `~/.agents/skills` — the same root Codex uses,
which openclaw's docs call a "compatibility skill root" — and that openclaw reads it **only** when
`OPENCLAW_STATE_DIR` is the default `~/.openclaw`. We print that condition alongside the path and a
gate fails if we ever stop printing it. Is the condition stated correctly today?
3. **MCP registration spelling.** We emit
`openclaw mcp add ripwire --command --arg --mcp`. That is read off `src/cli/mcp-cli.ts` (Commander,
`--arg ` with a required value), **not** run. If `openclaw mcp doctor ripwire --probe` disagrees,
we are shipping a command that silently registers nothing.
4. **Hook slot.** We deliberately emit **no** hook line for openclaw, on the grounds that its
`before_tool_call` is an in-process plugin API rather than a shell-command slot like Claude Code's or
Codex's. If openclaw does have a shell hook slot, we are under-claiming and should port ours.

## What we do NOT need

Not asking anyone to advocate for ripwire, file a PR, or change anything in openclaw. Just: does the
recipe we print actually work? Paste the block, including failures.

Run bash scripts/verify-agent-integration.sh openclaw with a real openclaw installation and review the Phase 1 output. Paste the two prompts into openclaw, check the context path, skills-root condition, MCP command, and hook behavior, then paste the filled markdown block back into the issue; done means the claims are confirmed or failures are documented.

Written by the indexing model from the issue text.