# SESSION STATE - Live Memory

Last update: 2026-03-10 (builder glass preset + ignore cleanup)
Owner: Codex + user

## Current objective
Stabilize the UB24 builder workflow with reproducible progress and no context loss between sessions.

## Current context snapshot
- Canonical workspace: `C:\word`
- Active branch target: `ai/ub24-builder-v1`
- Runtime command: `python -m demo.app`
- Canonical URL: `http://127.0.0.1:5001/elementor/1`

## Done recently
1. Defined memory/versioning strategy based on stable rules + live state + history.
2. Added startup/close hooks to enforce session continuity.
3. Updated startup protocol to always read `agent.md` in addition to core memory files.
4. Disabled legacy routes `/customizer2/<site_id>` and `/customizer3/<site_id>` in `demo/routes/customizer_ascii.py` (now return `404`).
5. Stabilized `elementor_builder` for restaurante in free-drag mode while preserving manual placement.
6. Reset action now restores base template per rubro (instead of wiping all blocks).
7. Menu mapping improved to resolve links by semantic intent + block type (contact/map/social/review/cards).
8. Added draft autosave and preview-final device propagation (`desktop/tablet/phone`).
9. Added universal LLM router toolkit for multi-project reuse at `tools/llm_universal` with provider fallback, cooldown, and SQLite usage tracking.
10. Stored cross-project LLM execution plan in root file `llm`.
11. Configured and verified `GITHUB_TOKEN` and `CLOUDFLARE_API_TOKEN`/`CLOUDFLARE_ACCOUNT_ID` in user env.
12. Improved universal router behavior: missing-key fallback handling + round-robin balancing by route.
13. Validated unified multi-provider flow with real calls (fallback to GitHub Models when other candidates fail).
14. Added unified chat interface `tools/llm_universal/chat_cli.py` with persistent context via `chat_history.json`.
15. Documented full LLM implementation log in `codex/LLM_UNIVERSAL_LOG.md`.
16. Enabled OpenRouter in user env from local Cline secrets and validated live calls.
17. Added unified HTTP API server for LLM (`tools/llm_universal/api_server.py`) with persistent sessions and context continuity.
18. Added run/test scripts for immediate usage (`run_server.ps1`, `test_api.ps1`) and validated E2E flow.
19. Finalized universal assistant as standalone service in `tools/llm_universal` (detached from `demo.app`), with browser chat on `:5055`.
20. Installed Task Scheduler autostart (`GKACHELE-LLM-Chat-Autostart`) to launch at user logon.
21. Locked assistant web to fixed agent mode and route (`agent`) to avoid manual model switching.
22. Added startup memory preload in agent sessions (`AGENTS.md`, `agent.md`, `codex/SESSION_STATE.md`, `codex/VERSIONADO_IA.md`).
23. Verified LLM GUI server health on `127.0.0.1:5055` and clarified `304` responses are normal cache behavior for `GET /assistant`.
24. Verified Groq API key is valid (direct `GET /models` and `POST /chat/completions` return `200`).
25. Started and validated both local GUIs:
    - LLM assistant: `http://127.0.0.1:5055/assistant` (health: `GET /api/llm/health`)
    - UB24 builder: `http://127.0.0.1:5001/elementor/1`
26. Added xAI (Grok) provider routing to `tools/llm_universal/provider_config.example.json` and set `XAI_API_KEY` in user env.
27. xAI endpoints returned `403` for both `GET /v1/models` and `POST /v1/chat/completions` with the provided key (base URL `https://api.x.ai/v1`).
28. Added OpenAI provider using Responses API with `codex-mini-latest` in routes; implemented Responses payload parsing in router and updated key setup/checklist and autostart env propagation.
29. Added menu glass/dark/default presets and hover/blur styling in `elementor/templates/elementor_builder.html`; logging of incoming presets to `logs/elementor_save.log`.
30. Expanded `.gitignore` to drop local artifacts (db/logs/snapshots/__pycache__/gitea_data/free-llm resources, etc.).

## In progress
1. Final QA of restaurante flow end-to-end (order, menu links, responsive, publish persistence).
2. Enable additional provider keys (beyond OpenRouter) for universal LLM fallback.
3. Stabilize free-tier provider mix (monitor Groq/auth errors if they reappear and keep fallback stable).
4. Prepare first project integration using `/api/llm/chat`.
5. Standalone service smoke validated (`/assistant` + `/api/llm/health` + chat POST).

## Blockers
1. xAI endpoints returned `403` with the provided key; need a valid xAI API key and/or the correct base URL from the xAI console.
2. OpenAI provider added but not validated yet (requires `OPENAI_API_KEY`).
3. Agent UI occasionally shows raw tool-call JSON (e.g. `{"type":"tool_call",...}`) instead of executing tool and returning final natural-language response.

## Next 3 steps
1. If `:5055` goes down again, inspect `tools/llm_universal/_api_err.log` for crash details and adjust autostart to restart on failure.
2. Persist and auto-restore last `session_id` in web UI so context resumes automatically after PC restart.
3. Fix agent tool-call parser/loop so raw JSON is never rendered to user and every valid tool call is executed before final answer.

## Quick handoff template (copy and fill at close)
### What changed today
- 

### What was validated
- 

### What failed (if any)
- 

### Next immediate action
-