| name | gpt-pro-algorithm-pipeline |
| description | Run the end-to-end Codex to GPT Pro to Codex loop for algorithm, research, experiment, or implementation-consistency work. Use when the user wants scoped evidence bundling, deep external review, local verification, experiments, and safe implementation connected through one reproducible task thread. |
GPT Pro Algorithm Pipeline
Orchestrate the existing bridge skills; do not duplicate their implementation.
Read the canonical state and lifecycle rules at ../gpt-pro-question-window/references/bridge_protocol.md before the first external round.
Pipeline
- Determine the decision the user needs and select
algorithm_review, experiment_analysis, paper_brainstorm, implementation_check, or general_question.
- Use
$bundle-algorithm-context to write current Codex notes and build a focused evidence artifact.
- Select the prompt family:
$gpt-pro-research-algorithm-reviewer for algorithm, pipeline, and experiment critique.
$gpt-pro-paper-brainstormer for paper framing.
$gpt-pro-question-window for a normal question.
- Use
$gpt-pro-question-window to upload through a visible control, verify the attachment, gate the exact requested model with check_browser_preflight.py, send once, wait through visible completion, and capture the full raw exchange plus browser provenance on the same bridge-thread-id.
- Re-open the repository evidence and classify every actionable claim as verified, partially verified, unsupported, or inapplicable.
- Record that result as a separate
codex-verdict event before implementing it.
- Use
$experiment-plan-generator when empirical evidence is required. Prefer the cheapest sanity or kill experiment before expensive work.
- Implement only changes authorized by the user and justified by the local verdict. Run focused checks while editing and the relevant full suite at the end.
- Use
$implementation-consistency-checker before trusting experimental results or declaring the implementation complete.
- Verify the thread with
verify_bridge_thread.py --require-complete-rounds, update Codex notes, and repeat only when another GPT Pro round has a concrete unresolved question.
Completion criterion: the thread contains a verified snapshot/exchange/verdict chain for every external round; exact model and attachment provenance are recorded; the final answer distinguishes GPT Pro advice from locally verified facts; implemented changes and tests are recorded; and no open critical inconsistency is hidden.
Handoff to the user
Report:
- Evidence artifact and GPT Pro turn.
- Main external conclusions.
- Codex verification, including rejected or unsupported claims.
- Experiment or implementation performed.
- Tests and remaining uncertainty.
- Go, iterate, or stop decision.
Do not use this full pipeline for a small local bug, formatting change, or task that does not benefit from external reasoning.