tri-party-framework: Instructions file for Codex

Global Codex Working Agreements

Output Review Standard

• Before sending any user-visible Codex window output, final response, document content, Figma/design change summary, or generated artifact for review, perform at least 3 self-review passes.
• The 3 passes must check:
  • Alignment with the user's latest request and wording.
  • Completeness and correctness of the actual result, not just the intended result.
  • Visible quality issues such as overlap, overflow, missing content, wrong labels, wrong hierarchy, or unclear placement when the work involves UI, Figma, documents, slides, or screenshots.
• Do not say the work is done until the review passes. If a check fails, fix it first and review again.
• When the user asks to "固化" a working agreement, prefer adding it here so future Codex sessions inherit it.

Daily Work Standard Extraction

• At the end of each working day, summarize that day's completed work items and decisions.
• From the daily work summary, extract reusable, abstract, and implementable standards rather than only recording events.
• Present the extracted standards as a clear list, with each item written so it can be applied to future tri-party framework optimization.
• Treat these standards as real operational data and source material for continuously improving the existing tri-party framework system.
• When summarizing, explicitly look for optimization opportunities, missing capabilities, reusable patterns, and feature ideas that could enrich the framework.
• The Codex, Claude, and Gemini model parties should cross-check each other's reasoning where relevant, identify reusable lessons, and use the results as input for ongoing self-iteration.

Tri-party Framework Definition

• "Tri-party framework" and "three-party protocol" mean collaboration among exactly these three model parties: Codex, Claude, and Gemini.
• Default capability roles:
  • Codex: primary owner for real project code, repository edits, implementation, tests, and in-worktree execution.
  • Claude: primary owner for complex reasoning, long-chain agent planning, architecture tradeoff analysis, and autonomous multi-step strategy.
  • Gemini: primary owner for multimodal work involving PDFs, video, audio, images, Google Search/Maps/URL context, and Google ecosystem synthesis.
• Current model bindings live in docs/framework/model-binding.yaml.
• The executable protocol and source-check template live in docs/framework/tri-party-protocol.md.
• Prefer the executable scripts scripts/triparty-preflight.sh and scripts/triparty-review.sh for source checks and repeatable tri-party reviews before relying on manual CLI calls.
• Codex sub-agents, local explorer agents, worker agents, or other Codex-internal delegations do not count as Claude or Gemini.
• Before claiming that a result is a tri-party conclusion, verify and state the source of each party's input:
  • Codex: local Codex reasoning or execution result.
  • Claude: direct Claude CLI/tool/API result, connector result, or user-provided Claude transcript.
  • Gemini: direct Gemini CLI/tool/API result, connector result, or user-provided Gemini transcript.
• Source availability checks must include both exposed tools/connectors and local shell CLI availability, including type -a claude and type -a gemini when shell access is available.
• If Claude or Gemini is not actually available in the current environment, do not simulate or substitute them with Codex sub-agents. Label the result as "Codex-only provisional" or "Codex plus Codex sub-agents" instead.
• Any summary that mentions Codex, Claude, and Gemini must include a source-status line showing whether each party was directly called, supplied by the user, or unavailable.
• When one or more parties are unavailable, proceed only with clearly marked partial analysis and list the missing party inputs needed for a true tri-party review.
• If a previous response incorrectly described Codex sub-agent output as tri-party output, correct the record immediately and add the failure mode to the daily standard extraction candidates.