mathpipe-browser-proof-ops

name	mathpipe-browser-proof-ops
description	Use when working in MathPipeProver on ChatGPT project preparation, durable source refresh, project-source inspection, or browser-backed external_agent submission. Prefer scripts/chatgpt_browser_agent.sh prepare/submit over ad hoc UI work, enforce the Extended Pro target, and keep durable sources separate from temporary attachments.

MathPipe Browser Proof Ops

Use this skill for:

• preparing a ChatGPT project before proof work
• inspecting or syncing durable project sources
• submitting external_agent request packets in the browser

Workflow

1. Read docs/browser_chatgpt.md if you need the durable-vs-temporary policy.
2. Use the existing visible Chrome/CDP session for automation. Prefer MPP_CHATGPT_CDP_URL or pass --cdp-url URL explicitly.
3. Before prepare or submit, verify the CDP endpoint is alive:

curl -s "$MPP_CHATGPT_CDP_URL/json/version"

If CDP is unavailable, stop and report the infrastructure problem. Do not switch to headless mode, do not launch a fresh profile with --profile-dir, and do not substitute an unrelated browser tool path on your own.

4. For project setup or source sync, run:

scripts/chatgpt_browser_agent.sh prepare --project-url URL [--cdp-url URL] [--add-source PATH ...] [--remove-source NAME ...]

This command prints JSON including:

• base_model
• effort_mode
• source_names
• confirmed_sources
• missing_sources
• lingering_sources

5. For one browser-backed role submission, run:

scripts/chatgpt_browser_agent.sh submit --project-url URL --request-file PATH --response-file PATH [--cdp-url URL] [--attach-file PATH ...]

6. Treat browser drift as part of the job, not as an automatic stop condition.

• Re-open the project page and verify you are still in the intended project.
• Re-check the Extended Pro target before important prompts: reasoning Pro plus model 5.5.
• If prepare reports missing durable sources, retry one file at a time and verify again.
• If a response file is missing but the chat URL exists, recover the existing chat before resubmitting (use /inspect-chat for a one-shot status read, /recover-chat to harvest).
• If ChatGPT lands on an account chooser and there is a single clear account entry, select it and continue.
• Keep natural-language analytical proof roles on ChatGPT Extended Pro via external_agent. Use subagents only for explicit coding/simulation tasks or Lean formalization proof-engineering.
• For unattended runs, /heartbeat <interval> starts an orchestrator-pace loop that wakes the orchestrator periodically and advances the pipeline on its own (this is the orchestrator-loop skill, not the old Python watcher chain which was removed).
• For the literature role specifically (prompts/soft/02_literature_soft.md), submit with --deep-research instead of Extended Pro — DR can browse the open web and academic repositories and return long-form synthesis with citations.

Durable Source Policy

• Add only stable background files.
• Typical candidates: objective or claim file, paper PDF, durable proof-state file, current stable route memo.
• Do not add packets, logs, or per-step drafts as durable sources.
• Remove stale branch-specific durable sources before adding replacements.

Temporary Attachment Policy

• Use --attach-file only for per-step inputs that should stay chat-local.

Browser Policy

• Reasoning must be Pro.
• Model must be 5.5.
• In automated repo tests, use the same visible Chrome/CDP route that the live workflow uses.
• Treat CDP attach failure as a hard stop, not as a cue to invent another browser launch path.
• Treat ambiguous login/auth state as a hard stop, but handle straightforward account-selection and source-verification repairs autonomously.

References

• docs/browser_chatgpt.md
• docs/soft_scaffolding.md
• .claude/commands/set-sources.md
• .claude/commands/submit-role.md