rosetta-consult

# rosetta-consult Tool: `mcp__rosetta__consult` (one tool, returns assistant text only — no files, no images). ## Model selection (2026-07, GPT-5.6 "Sol" lineup) **Default is `gpt-5-6-pro` (Pro) — use it for everything.** rosetta exists precisely for Pro access; if a task doesn't merit Pro, it usually shouldn't go through rosetta at all (answer it yourself or use codex). Don't downgrade the model to "save" anything. - Non-Pro slugs (`gpt-5-6-thinking`, `gpt-5-6`, `gpt-5-5`) exist via `model`, but only pass them when the user explicitly asks for a cheaper/faster tier. - `thinkingEffort`: `"standard" | "extended" | "max"` — the web UI's 中/高/极高 lanes are `gpt-5-6-thinking` with these three values. Omit to use the model's natural default. ## When to call - Hard reasoning, math derivations, proof checks, tricky algorithm design → default Pro (30 s – 15+ min, blocks the call; the MCP server heartbeats progress so hour-long runs survive host idle timeouts). - Brainstorming, design alternatives, naming, writing feedback → default Pro too; Pro's breadth is what you're paying the latency for. - Cross-checking your own answer when stakes are high → default Pro. - Long-running research context that must survive across tool calls / MCP restarts → `recall: "<topic>"` (disk-persisted thread at `~/.rosetta/state.json`). ## When NOT to call - Anything you can answer correctly yourself — consult costs a Chrome tab + focus mutex and slows down parallel work. - Code edits, file ops, running commands — consult only returns text. - Routine questions that do not benefit from an external second opinion. ## Threading model Each MCP server process is one implicit conversation. Back-to-back `consult` calls in the same host agent keep multi-turn context automatically. - **Default (no threading flags) → continue the implicit session conversation.** Prefer this aggressively: follow-up questions, revisions, new evidence, related subtasks, alternative designs, and critique of an earlier answer all belong in the existing conversation. Do not pass `fresh` merely because the next prompt is self-contained or because this is another tool call. - `fresh: true` → start a new conversation, which becomes the new session default. Use it only when at least one of these is true: 1. the problem is genuinely unrelated and earlier context has no value; 2. an independent adversarial/zero-context assessment is required; 3. the current conversation is already very long, confused, or contaminated; 4. the user explicitly asks for a new conversation. - `recall: "<name>"` → ignore implicit session, route through a named disk-persisted thread. Different names = parallel contexts that survive restarts. - `fresh: true` + `recall: "<name>"` → reset that named thread. The implicit conversation is scoped to one MCP server process. Codex subagents/agent threads may each own a different process even when they belong to the same user task. When context must cross agent boundaries, host restarts, or separate sessions, choose one stable descriptive `recall` name and reuse it; do not compensate by repeatedly creating fresh conversations. ## Prompt patterns - **Pro reasoning**: include the full problem statement, the constraints, and what kind of answer you want (proof? plan? code sketch?). Pro thinks harder when goals are explicit. - **Cross-check**: paste your candidate answer + ask "find errors or confirm correctness, be specific." - **Recall thread**: first call sets context ("we're working on X, here's the setup"); subsequent calls assume prior turns are remembered — don't re-paste. ## Response integrity - Rosetta is expected to block until the backend Pro turn is verifiably complete. A short preamble or stage summary that plainly fails the requested output contract must be treated as **incomplete**, even if it looks polished or carries success-like metadata. - Do not send a follow-up into that conversation until the original backend turn is confirmed complete. Report the incomplete/error state first; a follow-up can otherwise branch from an internal thoughts/code node and contaminate the named thread. - This is a defensive caller check, not the completion mechanism. MCP transport and conversation-mapping verification remain responsible for returning only the full final answer. ## Prerequisite Chrome must already be running with `--remote-debugging-port=9222 --user-data-dir=$HOME/.rosetta/profile` and signed in to chatgpt.com. If consult returns `auth error [not-logged-in]`, the user needs to start/sign in to that Chrome — don't retry blindly.