Files
claudetools/.claude/memory/feedback_ollama_tier0_routing.md
Howard Enos 7e2e3a5882 sync: auto-sync from HOWARD-HOME at 2026-04-23 06:21:23
Author: Howard Enos
Machine: HOWARD-HOME
Timestamp: 2026-04-23 06:21:23
2026-04-23 06:21:24 -07:00

47 lines
2.6 KiB
Markdown

---
name: Route Tier-0 tasks through Ollama (Mike's ClaudeTools design intent)
description: Drafts, summaries, classifications, extractions MUST go through Ollama per Mike's tiered-model architecture. Don't default to Claude inference for low-stakes text generation.
type: feedback
---
Route Tier-0 tasks (summaries, classifications, drafts, extractions) through Ollama. Not optional — this is how Mike designed ClaudeTools to work.
**Why:** Mike built the tiered-model architecture (`CLAUDE.md` Model Routing section + `.claude/OLLAMA.md`) deliberately. Tier 0 is free + fast + private. Defaulting to Claude for every drafting task burns context window and Anthropic tokens on work that qwen3:14b does fine.
**How to apply:**
- Drafting emails, session-log paragraphs, status-update sentences, commit-message first-drafts → qwen3:14b
- Summarizing long output (Graph JSON, PowerShell transcripts, log tails) → qwen3:14b
- Extracting structured data from text → qwen3:14b
- Suggesting refactors / generating docstrings → codestral:22b (then review)
- NEVER for: auth decisions, credential handling, production migrations, security review, citation work, production-change scripts
**Endpoint resolution (updated 2026-04-22 in `.claude/OLLAMA.md`):**
```bash
if curl -s -m 2 http://localhost:11434/api/tags >/dev/null 2>&1; then
OLLAMA="http://localhost:11434"
else
OLLAMA="http://100.92.127.64:11434"
fi
```
HOWARD-HOME has the canonical models loaded locally (qwen3:14b, codestral:22b, nomic-embed-text, plus bonus qwen3-coder:30b) — so HOWARD-HOME uses local Ollama, not Mike's. Zero Tailscale hop.
**Call pattern for qwen3 — use `/api/chat` with `think:false`**, NOT `/api/generate`. qwen3 on generate endpoint dumps reasoning into internal thinking tokens and returns empty `response` field. Chat endpoint with `think:false` returns clean content in `message.content`:
```python
body = json.dumps({
'model':'qwen3:14b',
'messages':[{'role':'user','content': prompt}],
'stream':False,
'think':False
}).encode()
# POST to OLLAMA + '/api/chat'
# Read res['message']['content']
```
Codestral doesn't need `think:false` — just use it on `/api/chat` normally.
Cold-start ~30-50s on first call per model per session; warm calls 1-5s.
**Incident 2026-04-22:** Spent an entire Cascades rollout session (G1 hygiene, orphan cleanup, risk register, synology discovery, etc.) without routing a single task through Ollama despite many drafting opportunities (report drafts, summary text, email drafts). Howard called this out: "just make sure ollama is being used as mike has designed claudetools to work."