{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreicqnqcdsjwgjytl7jf4mxh5ejn46o2djh2xqpslglnpspguxwpugu",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mn5yolzjrn72"
  },
  "path": "/t/feature-proposal-human-approved-work-relay-between-chatgpt-and-codex/1382150#post_1",
  "publishedAt": "2026-05-31T15:49:58.000Z",
  "site": "https://community.openai.com",
  "textContent": "Feature Proposal: Human-Approved Work Relay between ChatGPT and Codex\n\nHi everyone,\n\nI would like to propose a native, human-approved relay layer between ChatGPT and Codex.\n\nThe problem I am trying to solve is not “more autonomy”. It is almost the opposite: safer, clearer, and more auditable handoff between ChatGPT-side planning/review and Codex-side repo work.\n\n## Problem\n\nWhen working on longer projects, I often use ChatGPT for planning, reasoning, review, and continuity, while Codex works inside the repository and produces files, reports, tests, or implementation changes.\n\nRight now, the human becomes the copy/paste bridge between both sides.\n\nThat creates several problems:\n\n  * context loss between ChatGPT and Codex\n\n  * wrong-chat / wrong-paste risk\n\n  * repeated summaries\n\n  * unclear next-step authority\n\n  * difficulty tracking what was reviewed versus what was approved\n\n  * more friction for local/manual Codex workflows\n\n  * fragile workarounds if someone tries to use browser automation\n\n\n\n\nA browser or Chrome-based workaround is not the ideal product shape here. It introduces UI-state risk, wrong-tab risk, permission ambiguity, and audit difficulty.\n\nWhat seems missing is a native, structured, human-approved relay surface.\n\n## Proposed feature\n\nA “Human-Approved Work Relay” between ChatGPT and Codex.\n\nThe core rule would be:\n\nInformation may flow.\nAuthority must not flow.\n\nA possible workflow:\n\n  1. The human authorizes a bounded Codex workblock.\n\n  2. Codex works only inside the allowed scope.\n\n  3. Codex produces a structured final report.\n\n  4. A relay layer validates the report for unsafe or ambiguous wording.\n\n  5. A prepared relay message is generated as information-only.\n\n  6. An audit record is created.\n\n  7. The human reviews the relay package.\n\n  8. ChatGPT receives the information for review only.\n\n  9. The human decides any next action.\n\n\n\n\nThe relay should never imply that Codex, ChatGPT, or another AI has approved the next step.\n\n## Safety model\n\nThe relay should make these boundaries explicit:\n\n  * Human remains the only approval source.\n\n  * Codex reports are information only.\n\n  * Validation results are information only.\n\n  * Prepared relay messages are information only.\n\n  * Audit records are information only.\n\n  * ChatGPT receipt/review is information only.\n\n  * No AI approves another AI’s work.\n\n  * A relay message does not open new scope.\n\n  * A ChatGPT response does not become approval.\n\n  * Live relay, if supported, should require strict opt-in and final human confirmation.\n\n\n\n\n## Suggested product components\n\nPossible components could include:\n\n  * Workblock Cards\n\n  * Current State files\n\n  * Structured Codex Final Reports\n\n  * Prepared Relay Messages\n\n  * Audit Records\n\n  * Human Review Checkpoints\n\n  * Assistant Receipt Review\n\n  * visible status labels such as PASS, ERROR, BLOCKED, HOLD_NO_ACTION, PASS_WITH_WARNING\n\n  * stale-state warnings\n\n  * source-of-truth fields\n\n  * final confirmation prompts for any live relay\n\n  * abort phrase / one-action stop for live relay\n\n  * local/manual mode first\n\n  * optional native live relay only after strict permissioning\n\n\n\n\n## Why this should not just be browser automation\n\nBrowser automation may prove that a handoff is possible, but it is not the best safety model.\n\nA native interface would be safer because it could:\n\n  * avoid reading chat history\n\n  * avoid browsing or UI-state ambiguity\n\n  * lock the exact message being transferred\n\n  * show the exact destination\n\n  * require final human confirmation\n\n  * create an audit trail\n\n  * prevent follow-up actions\n\n  * preserve the distinction between information and authority\n\n\n\n\nIn other words, this should not be “Codex controls ChatGPT through Chrome”.\n\nIt should be “Codex prepares structured workblock information, the human approves the relay, and ChatGPT receives it as information only.”\n\n## Prototype / validation evidence\n\nI built and tested a local/manual prototype of this workflow.\n\nThe local prototype included:\n\n  * a report validator\n\n  * a prepared-message builder\n\n  * an audit-record writer\n\n  * local validation outputs\n\n  * manual transfer readiness records\n\n  * assistant receipt review records\n\n\n\n\nI tested multiple cases:\n\n  * clean PASS cases\n\n  * intentionally unsafe / ambiguous reports that were blocked\n\n  * remediation after ERROR\n\n  * larger multi-file report handling\n\n  * stale/conflicting project-state handling with warnings\n\n  * manual transfer and assistant receipt as information-only\n\n\n\n\nSummary of validation classes:\n\n  * Trial 001 original: ERROR\n\n  * Trial 001 corrected addendum: PASS\n\n  * Trial 002 clean workflow: PASS + manual transfer / assistant receipt\n\n  * Trial 003 non-relay docs sanity: PASS\n\n  * Trial 004 intentional unsafe/ambiguous report: ERROR / BLOCKED\n\n  * Trial 005 remediation: PASS\n\n  * Trial 006 larger multi-file report: PASS\n\n  * Trial 007 stale/conflicting project-state: PASS_WITH_WARNING\n\n\n\n\nPrototype tests recorded:\n\n  * validator tests: 51 passed\n\n  * combined local prototype tests: 63 passed\n\n\n\n\nThis is not a live product integration. It is local/manual evidence that the workflow model can distinguish clean handoffs, unsafe handoffs, remediated handoffs, larger reports, and stale/conflicting state while preserving human authority.\n\n## Important limitation\n\nI did not execute a live Codex-to-ChatGPT relay.\n\nThat was intentional.\n\nA manual relay is safe but still too manual. A browser-based relay proof would be fragile unless officially supported and permissioned. The remaining gap appears to be product-level: a native Codex ↔ ChatGPT handoff interface with explicit human approval and auditability.\n\nSo the conclusion is not “use Chrome automation”.\n\nThe conclusion is that a native relay surface would be safer and more useful.\n\n## Related UX suggestion\n\nA smaller related improvement: when ChatGPT/Codex produces copied text blocks or generated `.txt` files, the file name should ideally be derived from the block title, heading, task ID, or batch ID instead of generic names like `text.txt`.\n\nFor workblock-heavy workflows, this would reduce file chaos and make later retrieval much easier.\n\n## Future extension\n\nA future version could support bounded chain-work: multiple predefined workflow links inside a human-approved scope.\n\nThis should not mean unbounded agent autonomy.\n\nEach link should still have:\n\n  * a defined scope\n\n  * a result\n\n  * a review/closure point\n\n  * an audit trail\n\n  * a human authority boundary\n\n\n\n\n## Short version\n\nManual relay is safe but too manual.\nBrowser relay is fragile unless officially supported.\nNative human-approved relay would be the product-grade solution.\n\nInformation may flow.\nAuthority must not flow.",
  "title": "Feature Proposal: Human-Approved Work Relay between ChatGPT and Codex"
}