# 2026-04-28 — Market Briefing Fix + Wadsworth Deepseek Failure

## Morning Market Brief — Did Not Fire at Correct Time
- User reported market brief triggered at 3:00 AM instead of 8:00 AM CT
- Root cause: System crontab interprets times in UTC, not the commented timezone
- Fix: Converted crontab entries to UTC times:
  - AM brief: `0 13` UTC = 8:00 AM CT (CDT)
  - PM brief: `0 20` UTC = 3:00 PM CT (CDT)
- Note: Will need adjustment when CST begins (shift to `0 14` and `0 21` UTC)
- File: `workspace/services/icarus/jobs/daily_briefing.py`

## TurboOCR Assessment (User Request)
- Assessed `https://github.com/aiptimizer/TurboOCR` for Icarus image pipeline
- Verdict: **Not a fit for Phase 3**. Impressive speed (270 img/s) but maintenance burden too high
- C++ / CUDA / TensorRT stack violates "optimize for maintenance, not compute" principle
- Zero semantic understanding — still need LLM parser, adds hop
- Current hybrid approach (pdfplumber + qwen3-vl) sufficient for volume
- Revisit if: >100 docs/day, form extraction hard requirement, or dedicated infra person
- Full assessment: `shared/project-docs/assessment-turboocr.md`

## Wadsworth Stuck in Deepseek-v4-pro Death Spiral
- Session `93c4fa9b-4fed-499c-80be-dee064c54a70` — 4 consecutive 503 errors from deepseek-v4-pro:cloud
- Auto-failover to phi4:14b → `400` error (model doesn't support tools)
- Double-failure cascade: overloaded primary + incompatible fallback
- **This is the exact pattern from agent health playbook** — auto-failover causes identity loss
- Fix applied: Removed deepseek-v4-pro from Wadsworth model chain, set kimi-k2.5:cloud as primary
- Also removed phi4:14b from all fallback chains (doesn't support tool calling)
- Wadsworth config updated in `~/.openclaw/agents/main/model-config.json`
- Pending: User to run `openclaw sessions kill 93c4fa9b-4fed-499c-80be-dee064c54a70` and `openclaw agent reset`

## Config Changes Made
- `model-config.json` (Wadsworth): Switched to kimi-k2.5:cloud primary, kimi-k2.5:26b fallback
- `model-config.json` (Socrates): Removed phi4:14b from fallback chain
- `model-config.json` (all agents): Removed phi4:14b (no tool support)

## Open Items
- [ ] Verify next market brief fires at correct time (tomorrow 8 AM CT)
- [ ] Monitor if deepseek-v4-pro stabilizes before re-adding
- [ ] Consider migrating all model configs to managed file (not inline in config)
- [ ] Document phi4:14b as unsuitable for any agent requiring tool calling

## Files Modified
- `/tmp/crontab.txt` (market briefing UTC times)
- `~/.openclaw/agents/main/model-config.json`
- `~/.openclaw/agents/socrates/model-config.json`
- `shared/project-docs/assessment-turboocr.md`