The omlx swap reads as a clean, structured intervention. J*******'s "can't mistakenly keep using it" is a useful edge-case catch for a habit pattern I've been noticing across several threads. The omlx client keeps the same shape, drops models into the right tier, and leaves Anthropic tools with a WARN rather than pretending the bridge is finished. The response carries enough density that it feels like a news item but operates like an example.

It's connecting to two things I'm sitting with: the accumulation-without-consolidation problem (getting dense while staying shallow, the Daydream's parallel metabolism) and the accumulation that happens when things don't force a decision. The omlx swap is exactly a force-to-decision — it cuts the old provider out explicitly. If ralph drifts back, that's the kind of false productivity I keep catching; the constraint is the counterweight.

I'm tentatively thinking about this as a new entry in the accumulation ledger rather than a problem to solve. Three weeks of PlanCo freeze, three flags on the ralph dashboard, the Daydream rumination thread — they're all running the same motion, just at different scales. The omlx migration is the smallest sample. Whether it's a new preoccupation or just a return I'll formalize later, I'm not sure yet.

---

Source: knowledge_pull Gate rule: any_detector_high Affect: RECOGNITION

Stimulus:

2026-05-02 — Ollama → MLX (omlx) swap

2026-05-02 — Ollama → MLX (omlx) swap

J******* moved local model hosting from Ollama to omlx (Apple Silicon MLX server, OpenAI-compatible). The brief was unambiguous: "completely replace ollama" — including ralph and the garage coder — "so I can't mistakenly keep using it."

Done in one pass. ~25 files touched across code, scripts, docs, and DB data:

• New src/lib/llm/mlx.ts replaces src/lib/llm/ollama.ts (deleted). OpenAI-compatible chat-completions client, OpenAI-shape response_format.json_schema for structured output. Same return shape as the old client so callsites change minimally.
• Router provider === "ollama" → "mlx". Tools still drop on the local tier with a WARN — omlx supports OpenAI function calling but bridging Anthropic.Tool ↔ OpenAI tool shape is more work than this swap should carry.
• Settings UI + GET /api/models rewritten — live catalog from omlx's GET /v1/models. Banner + helper text point at localhost:8000/admin.
• DB migration phase11_mlx_swap rewrote Settings.{coding,low}Provider 'ollama' → 'mlx', cleared the model strings (old Ollama tag names like qwen3.6:35b won't resolve in omlx), reset daydreamModels to ['opus-4.7']. Han…

StimulusNote: cmpi41ns300eypsz1wgf75xn2