feat: overnight R&D automation — Deep Dive + tightening + DPO export (#503)
Some checks failed
Architecture Lint / Linter Tests (push) Successful in 8s
Smoke Test / smoke (push) Failing after 7s
Validate Config / YAML Lint (push) Failing after 6s
Validate Config / JSON Validate (push) Successful in 6s
Validate Config / Python Syntax & Import Check (push) Failing after 7s
Validate Config / Python Test Suite (push) Has been skipped
Validate Config / Shell Script Lint (push) Failing after 13s
Validate Config / Cron Syntax Check (push) Successful in 5s
Validate Config / Deploy Script Dry Run (push) Successful in 5s
Validate Config / Playbook Schema Validation (push) Successful in 7s
Architecture Lint / Lint Repository (push) Failing after 7s

This commit was merged in pull request #503.
This commit is contained in:
2026-04-13 02:10:16 +00:00
parent aeefe5027d
commit e53fdd0f49
3 changed files with 410 additions and 1 deletions

68
docs/overnight-rd.md Normal file
View File

@@ -0,0 +1,68 @@
# Overnight R&D Automation
**Schedule**: Nightly at 10 PM EDT (02:00 UTC)
**Duration**: ~2-4 hours (self-limiting, finishes before 6 AM morning report)
**Cost**: $0 — all local Ollama inference
## Phases
### Phase 1: Deep Dive Intelligence
Runs the `intelligence/deepdive/pipeline.py` from the-nexus:
- Aggregates arXiv CS.AI, CS.CL, CS.LG RSS feeds (last 24h)
- Fetches OpenAI, Anthropic, DeepMind blog updates
- Filters for relevance using sentence-transformers embeddings
- Synthesizes a briefing using local Gemma 4 12B
- Saves briefing to `~/briefings/`
### Phase 2: Tightening Loop
Exercises Timmy's local tool-use capability:
- 10 tasks × 3 cycles = 30 task attempts per night
- File reading, writing, searching against real workspace files
- Each result logged as JSONL for training data analysis
- Tests sovereignty compliance (SOUL.md alignment, banned provider detection)
### Phase 3: DPO Export
Sweeps overnight Hermes sessions for training pair extraction:
- Converts good conversation pairs into DPO training format
- Saves to `~/.timmy/training-data/dpo-pairs/`
### Phase 4: Morning Prep
Compiles overnight findings into `~/.timmy/overnight-rd/latest_summary.md`
for consumption by the 6 AM `good_morning_report` task.
## Approved Providers
| Slot | Provider | Model |
|------|----------|-------|
| Synthesis | Ollama | gemma4:12b |
| Tool tasks | Ollama | hermes4:14b |
| Fallback | Ollama | gemma4:12b |
Anthropic is permanently banned (BANNED_PROVIDERS.yml, 2026-04-09).
## Outputs
| Path | Content |
|------|---------|
| `~/.timmy/overnight-rd/{run_id}/rd_log.jsonl` | Full task log |
| `~/.timmy/overnight-rd/{run_id}/rd_summary.md` | Run summary |
| `~/.timmy/overnight-rd/latest_summary.md` | Latest summary (for morning report) |
| `~/briefings/briefing_*.json` | Deep Dive briefings |
## Monitoring
Check the Huey consumer log:
```bash
tail -f ~/.timmy/timmy-config/logs/huey.log | grep overnight
```
Check the latest run summary:
```bash
cat ~/.timmy/overnight-rd/latest_summary.md
```
## Dependencies
- Deep Dive pipeline installed: `cd the-nexus/intelligence/deepdive && make install`
- Ollama running with gemma4:12b and hermes4:14b models
- Huey consumer running: `huey_consumer.py tasks.huey -w 2 -k thread`