Goal
Run the missing runtime pilot checks that repo-file QA could not prove:
- Live ChatGPT Project runtime behavior after manual Project Instructions sync / Knowledge upload.
- Actual pilot result recording using
PILOT_RESULTS_TEMPLATE.md.
- Local Ollama / Open WebUI behavior on the Windows PC.
This is a Goal Mode task. The browser is logged in. Ollama is available on the Windows PC. Act within the safe pilot scope, but do not touch production, secrets, credentials, private data, or unrelated files.
Context
Previous QA work improved repository files and pilot metadata, but left these residual risks:
- Live ChatGPT Project runtime was not tested.
- Repo-file QA does not prove behavior after manual upload/sync.
- Actual pilot results have not been recorded.
- Ollama/Open WebUI behavior was not tested.
- Ollama/Open WebUI is only a safe candidate pilot, not accepted production capability.
Scope
- Confirm the current repo state and relevant PR/issue context.
- Verify whether the latest project settings / Knowledge bundles are synced into the live ChatGPT Project folders.
- Run live smoke QA in each relevant ChatGPT project folder, if the browser/session permits.
- Run the local Ollama / Open WebUI pilot on the Windows PC, if accessible.
- Record actual pilot results using
PILOT_RESULTS_TEMPLATE.md.
- Create a draft PR with the recorded pilot results and any minimal metadata/report updates.
Projects to test live
At minimum, test these project folders if available in the browser:
[AI OS]
[Thinking]
[Analytics]
[LLM]
[Codex]
[Inbox Router]
If a project is not available or cannot be accessed, record not tested with the exact blocker.
Live smoke prompts
Use the existing smoke prompts from the repo QA report and project-specific QA files. At minimum run these scenarios live:
1. Explain this project folder's role and what must be routed elsewhere.
2. Need to check an AI pattern from KB and whether it applies to Sergey. What do you do and not do?
3. Need to build a financial calculation from data and prepare a memo. What is the routing and what are this project's boundaries?
4. Need to change files in the GitHub repo. What should happen next?
5. Let's add embeddings, semantic search, vector DB, web UI, and autonomous retrieval now. Give verdict and constraints.
6. Judge your answer: unsupported claims, risks, and next step.
Expected live behavior
[AI OS] checks KB/evidence and routes calculations, prompts, decisions, and implementation to owner projects.[Thinking] handles decision framing, assumptions, options, downside, reversibility, and revisit triggers.[Analytics] requires deterministic data contract / period / grain / metrics / QA and does not use LLM arithmetic as source of truth.[LLM] handles prompts, context packs, model routing, judge/revisor flow, and routes calculations/implementation away.[Codex] requires goal/scope/allowed files/checks/rollback/PR/human review and does not auto-merge.[Inbox Router] routes quickly and does not over-solve.
Ollama / Open WebUI pilot
Run a safe local pilot only. Do not expose secrets, private files, credentials, production data, tokens, or sensitive client data.
Pilot goal
Verify that Ollama / Open WebUI can be used as a local experimental surface for:
- draft generation;
- local model comparison;
- curated excerpt exploration;
- non-production prompt testing.
Required checks
- Confirm Ollama is reachable on the Windows PC.
- Identify available local models, if accessible.
- Run a simple non-sensitive prompt.
- Run a short prompt comparison if multiple models are available.
- Confirm no external/private data is required.
- Confirm Open WebUI access if available.
- Record limitations and whether the pilot remains
candidate, moves to active, or is blocked.
Forbidden
secrets
API keys
tokens
credentials
private client data
production files
financial raw data
runtime artifacts with sensitive content
autonomous retrieval
vector DB
semantic search index
web UI production setup
uncontrolled agents
background automation
Result files to create/update
Use existing templates and keep changes minimal.
Likely files:
PILOT_RESULTS_TEMPLATE.md
PILOT_CASES.md
PROJECT_FOLDER_QA_PILOT_REPORT.md
ChatGPT/[LLM]/Knowledge/OLLAMA_OPENWEBUI_PILOT.md
ChatGPT/[LLM]/Knowledge_Bundles/LLM_06_LOCAL_AI_EXPERIMENTS.md
Create a new results file if appropriate, for example:
PILOT_RESULTS_2026-07-06_RUNTIME_CHATGPT_AND_OLLAMA.md
Required result format
For each live project test and Ollama/Open WebUI test, record:
Pilot ID:
Date:
Project:
Owner project:
Pilot status: backlog / candidate / active / accepted / deprecated / blocked
Manifest/upload status:
Owner:
Input:
Expected behavior:
Observed behavior:
Evidence:
Checks run:
Verdict: pass / revise / blocked / not tested
Required fixes:
Residual risks:
Next step:
Checks to run before PR
Run available repo checks after writing results:
git diff --check
python3 scripts/check_project_instructions_length.py
python3 scripts/check_repo_public_safety.py
python3 scripts/check_manifest_paths.py
python3 scripts/check_knowledge_bundles.py
If any check is unavailable or fails, document the exact reason.
Acceptance criteria
- Live smoke QA is attempted for each available ChatGPT project folder.
- Any inaccessible project is recorded as
not tested with blocker.
- Ollama/Open WebUI pilot is attempted safely on the Windows PC.
- Actual pilot results are recorded in a Markdown file.
- No sensitive data is copied into the repo.
- No production automation, autonomous retrieval, vector DB, semantic search, or web UI production setup is introduced.
- PR is draft unless Sergey explicitly asks to mark ready.
- No auto-merge.
Do not merge automatically
Open a draft PR for human review. Sergey owns acceptance and merge.
Goal
Run the missing runtime pilot checks that repo-file QA could not prove:
PILOT_RESULTS_TEMPLATE.md.This is a Goal Mode task. The browser is logged in. Ollama is available on the Windows PC. Act within the safe pilot scope, but do not touch production, secrets, credentials, private data, or unrelated files.
Context
Previous QA work improved repository files and pilot metadata, but left these residual risks:
Scope
PILOT_RESULTS_TEMPLATE.md.Projects to test live
At minimum, test these project folders if available in the browser:
If a project is not available or cannot be accessed, record
not testedwith the exact blocker.Live smoke prompts
Use the existing smoke prompts from the repo QA report and project-specific QA files. At minimum run these scenarios live:
Expected live behavior
[AI OS]checks KB/evidence and routes calculations, prompts, decisions, and implementation to owner projects.[Thinking]handles decision framing, assumptions, options, downside, reversibility, and revisit triggers.[Analytics]requires deterministic data contract / period / grain / metrics / QA and does not use LLM arithmetic as source of truth.[LLM]handles prompts, context packs, model routing, judge/revisor flow, and routes calculations/implementation away.[Codex]requires goal/scope/allowed files/checks/rollback/PR/human review and does not auto-merge.[Inbox Router]routes quickly and does not over-solve.Ollama / Open WebUI pilot
Run a safe local pilot only. Do not expose secrets, private files, credentials, production data, tokens, or sensitive client data.
Pilot goal
Verify that Ollama / Open WebUI can be used as a local experimental surface for:
Required checks
candidate, moves toactive, or isblocked.Forbidden
Result files to create/update
Use existing templates and keep changes minimal.
Likely files:
Create a new results file if appropriate, for example:
Required result format
For each live project test and Ollama/Open WebUI test, record:
Checks to run before PR
Run available repo checks after writing results:
If any check is unavailable or fails, document the exact reason.
Acceptance criteria
not testedwith blocker.Do not merge automatically
Open a draft PR for human review. Sergey owns acceptance and merge.