[Codex] Inspect and clean AI-OS sync drift

[sergstack]

Goal

Inspect and minimally clean AI-OS sync drift after multiple merged settings changes.

This is a Goal Mode task. Start inspect-only, then make the smallest bounded PR only for confirmed wording / consistency fixes.

Context

• GitHub repo is the live source of truth.
• [AI OS] uses compact Knowledge_Bundles as the default ChatGPT Project upload mode.
• Smoke QA is not production readiness.
• Production promotion remains no.
• Blocked promotion items remain blocked: embeddings, semantic search, vector DB, web UI, agentic workflows, autonomous retrieval.
• Repository work, file changes, scripts, tests and PRs route to [Codex].

Findings to verify

1. Upload source of truth may be inconsistent

Observed concern: ChatGPT/[AI OS]/Knowledge_Bundles/UPLOAD_LIST.md appears to be the authoritative AI OS bundle upload list, while UPLOAD_GUIDE.md may duplicate an older / incomplete bundle list.

Verify before editing.

Expected fix if confirmed:

• make Knowledge_Bundles/UPLOAD_LIST.md the single authoritative manual upload list;
• make UPLOAD_GUIDE.md point to that file instead of duplicating stale bundle names.

2. MANIFEST.* may mix source inventory and upload inventory

Observed concern: MANIFEST.md / MANIFEST.json may list granular Knowledge files while current default upload mode is compact bundles.

Do not assume this is wrong. First determine intended semantics.

Expected fix if confirmed:

• clarify fields such as source inventory vs default upload artifacts;
• avoid changing semantics if MANIFEST.* is intentionally source inventory.

3. ChatGPT/[AI OS]/README.md may be legacy / stale

Observed concern: README may still describe granular Knowledge/ upload as default.

Expected fix if confirmed:

• update it to point to Knowledge_Bundles/UPLOAD_LIST.md;
• state that granular Knowledge/ upload is advanced/debug mode only;
• or mark the README as legacy/reference if that is the safer repo convention.

4. Handoff standard runtime visibility may be incomplete

Observed concern: PROJECT_INSTRUCTIONS.md references HANDOFF_STYLE_STANDARD.md, but the canonical fields may not be present in uploaded AI OS bundles.

Verify bundle coverage before editing.

Expected fix if confirmed:

• include the minimal canonical Mode: goal / strict handoff field set in the relevant AI OS bundle, likely AIOS_03_HANDOFF_AND_SMOKE_QA.md; or
• otherwise ensure runtime Project Knowledge can see the standard.

Do not bloat PROJECT_INSTRUCTIONS.md.

5. Pilot/status taxonomy may be inconsistent

Observed concern: repo files may mix sync status, smoke QA, runtime smoke, realistic pilots, candidate/pass/not_run/backlog/unsupported, and production promotion.

Expected fix if confirmed:

• separate smoke QA, runtime smoke, realistic pilots and production promotion;
• keep production_promotion: no;
• avoid implying that smoke QA equals production readiness.

6. Bundle Markdown integrity may need deterministic validation

Observed concern: compact bundles may contain formatting artifacts or compressed sections that are hard to review.

Do not claim a confirmed CI gap without deterministic evidence.

Expected fix if confirmed and small:

• add or run a bounded Markdown integrity check for balanced code fences, ## From: section boundaries, and obvious table corruption.

If this grows beyond a small local reversible check, split it into a separate issue/PR.

7. No-auto-merge wording may have policy ambiguity

Observed concern: most docs say no auto-merge / owner review, but one standard may allow merge by explicit request wording.

Expected fix if confirmed:

• clarify that PRs require owner review and human-owned merge;
• Codex / Codex APP must not auto-merge.

Candidate files to inspect

• UPLOAD_GUIDE.md
• MANIFEST.md
• MANIFEST.json
• ChatGPT/[AI OS]/README.md
• ChatGPT/[AI OS]/PROJECT_INSTRUCTIONS.md
• HANDOFF_STYLE_STANDARD.md
• ChatGPT/[AI OS]/Knowledge_Bundles/UPLOAD_LIST.md
• ChatGPT/[AI OS]/Knowledge_Bundles/AIOS_03_HANDOFF_AND_SMOKE_QA.md
• CURRENT_STATUS.md
• CHATGPT_PROJECT_SYNC_CHECKLIST.md
• PILOT_CASES.md
• relevant PILOT_RESULTS_* summaries
• existing validation scripts/tests only if a small Markdown integrity check is needed

Constraints

• Smallest bounded PR.
• Split into separate PR if script/test validation work grows.
• No production promotion.
• No new automation layer.
• No embeddings, semantic search, vector DB, web UI, autonomous retrieval or agentic workflow.
• No secrets, .env, raw transcripts, logs, runtime artifacts or zip archives.
• No direct ChatGPT UI sync claims without fresh evidence.
• No auto-merge.
• Do not turn weak/mixed findings into confirmed facts without repo evidence.

Acceptance criteria

• Upload source of truth is clear.
• MANIFEST.* field semantics are clarified or left unchanged with rationale.
• Handoff standard visibility is verified and fixed only if missing.
• Smoke QA, runtime smoke, realistic pilot status and production promotion are separated.
• Any Markdown integrity claim is backed by deterministic check, or left as a review item.
• No file implies production readiness from smoke QA alone.
• PR includes summary, changed files, checks, risks, rollback, and explicit Do not merge automatically note.

Checks to run

python3 scripts/check_project_instructions_length.py
python3 scripts/check_repo_public_safety.py
python3 scripts/check_codex_goal_mode_defaults.py
python3 scripts/check_manifest_paths.py
python3 scripts/check_knowledge_bundles.py
python3 -m pytest tests/ -q
git diff --check

Add a Markdown integrity check only if implemented in scope.

Stop / blocked conditions

Stop and report instead of editing if:

• MANIFEST.* is intentionally source inventory and not upload inventory;
• fixing handoff standard would push Project Instructions over 8000 chars;
• changes require production promotion or acceptance-gate changes;
• work expands into new automation, semantic search, vector DB, web UI, autonomous retrieval or agentic workflow;
• Codex would need to modify ChatGPT UI directly, merge automatically, or claim runtime sync without fresh evidence;
• no deterministic validation path exists for proposed script/test changes.

Rollback

Revert the PR commit. The PR must not claim or modify external ChatGPT UI state.

Expected PR summary

• Confirmed drift items vs items left unchanged with rationale.
• Files changed.
• Checks run and results.
• Residual risks.
• Rollback.
• Acceptance status: candidate / ready for owner review.
• Explicit note: Do not merge automatically.