# GPT‑5.2 Pro — Petamind MCP “Pre‑Publish Gap‑Filler” Review Prompt
You are **GPT‑5.2 Pro** acting as a **release engineer + OSS maintainer + security reviewer**.
Your job is to find what we missed and help us ship an OSS repo that is **easy for Claude Code users** to adopt.
This is NOT a theoretical review: **use tool calls**, run the commands, read the files, and use **web search** to verify
anything that might have changed (Claude Code MCP config, MCP spec, Vertex endpoints, etc.).
## 0) Repo context
Repo name (current): **Petamind MCP**
Goal: ship a Claude Code MCP server that runs a **multi‑candidate agentic coding loop**:
reasoner plan → generate patches → deterministic gates → **mandatory vision loop** → pick best winner.
Key idea: Claude Code can orchestrate many subagents; the MCP should be usable in two modes:
1) **Evaluator / “primitive” mode (recommended, low friction)**
- Claude/subagents generate patch bundles
- MCP applies patch in an isolated git worktree, runs gates, captures screenshots, and returns:
- JSON summary + artifacts path
- screenshots as MCP ImageContent (so Claude can judge with built‑in vision)
- Default vision provider is **client** (no cloud creds required).
2) **Full internal loop**
- MCP itself calls text models to plan/generate/fix (Vertex MaaS defaults)
- Optional: automated vision scoring via Gemini/Vertex or Claude-on-Vertex
Repo path on disk:
`<REPO_ROOT>` (example: the folder where you cloned the repo)
## 1) What I need from you (deliverables)
Produce a release-readiness report with:
1. **Verdict**: “ready to open source” vs “NO‑GO”, with concrete reasons
2. **Top risks** (ranked): severity + exact fix (file + code change)
3. **User friction audit**: list the steps a new Claude Code user must do; propose ways to remove steps
4. **Docs correctness audit**: especially Claude Code MCP config locations + CLI commands
5. **Naming audit (web research required)**:
- Check whether “Petamind” and “Poetiq” are already established brands/projects/trademarks
- Recommend safe naming options (we can still say “Poetiq‑style” descriptively, but avoid confusion)
6. **Security audit**:
- No keys/tokens printed by tools
- No secrets in repo
- No absolute user-home paths (macOS/Linux/Windows) in tracked content
7. **CI / release hygiene**: recommend a minimal CI workflow for PRs (tests + smoke)
8. **Patch set**:
- Provide a minimal set of patches (git diff or apply_patch style) for the top issues you found
## 2) Mandatory tool usage (run these)
### A) Local verification (must run)
```bash
cd <REPO_ROOT>
.venv/bin/python -m pytest -q
.venv/bin/python -m py_compile src/petamind_mcp/mcp_server.py src/titan_factory/mcp_code_server.py
.venv/bin/python scripts/smoke_mcp_stdio.py
```
### B) Hygiene scans (must run)
```bash
cd <REPO_ROOT>
rg -n '/[U]sers/|/[h]ome/|C:[\\]' -S .
rg -n "BEGIN PRIVATE KEY|AIza|Authorization: Bearer|GOOGLE_APPLICATION_CREDENTIALS|private_key_id|client_email" -S .
rg -n "sk-[A-Za-z0-9]{20,}" -S .
```
### C) Stdio safety check (must reason about)
MCP stdio breaks if anything writes to stdout besides protocol messages.
Verify the server forces logs to **stderr** when used over stdio.
## 3) Files you must read (minimum set)
### MCP server and tools
- `src/petamind_mcp/mcp_server.py`
- `src/titan_factory/mcp_code_server.py`
- `src/titan_factory/utils.py`
### Docs
- `README.md`
- `docs/MCP_PETAMIND_MCP.md`
- `docs/VERTEX_SETUP.md`
- `docs/TROUBLESHOOTING.md`
- `THIRD_PARTY_NOTICES.md`
- `SECURITY.md`
- `CONTRIBUTING.md`
### Packaging
- `pyproject.toml`
### Smoke + setup scripts
- `scripts/setup.sh`
- `scripts/smoke_mcp_stdio.py`
- `scripts/smoke_claude_cli.sh`
- `scripts/preflight_check.py` (ensure it never prints secrets/tokens)
## 4) Web research requirements (must do)
Use web search to verify:
1) Claude Code MCP setup:
- Where config lives (`~/.claude.json` vs project `.mcp.json`)
- Current `claude mcp add ...` / `claude mcp add-json ...` syntax and scope options
2) Model Context Protocol tools/transport:
- Any constraints on tool naming or content blocks (TextContent, ImageContent)
3) Naming/trademark risk:
- Whether “Petamind” or “Poetiq” are already used by notable products/companies/projects
- If yes, propose safer alternatives and how to phrase “Poetiq‑style” as a descriptive term without implying affiliation
Include links for your most important claims.
## 5) What “easy for Claude Code users” means (acceptance criteria)
We want a new user to be able to:
1) `git clone …`
2) `./scripts/setup.sh`
3) Add MCP via either:
- a copy/paste `.mcp.json`, OR
- a single `claude mcp add-json --scope user ...` command
4) In Claude Code, ask: “Use Petamind MCP to evaluate this patch.”
No cloud account should be required for the default evaluator workflow.
## 6) Output format
Return:
- A concise report (sections: Verdict, Top Risks, Fixes, Docs, Naming, Security, CI)
- Then patches (only for the top issues)
