## The gap

Each collector in the agents panel answers "what did this subscription cost me". That is the right question, and #12031 / #12032 make the Codex collector answer it honestly by dropping turns a local backend served.

But it leaves those turns counted by nobody. Run Codex CLI against Ollama, or opencode against a local gateway, and the work simply vanishes from the panel — which is correct for the Codex tab and wrong for the panel as a whole. On my machine that is a third of the day's tokens with nowhere to live.

## The proposal

A `local` collector, so the panel has a home for work that ran on hardware you own. It counts exactly what the subscription collectors skip — the complement, not an overlap — across the same three sources they already read: native Codex rollouts, pi/omp sessions, and the opencode database.

The useful part is that "local" is a property of the **endpoint**, not a hardcoded runtime list. An endpoint qualifies when its URL resolves to a loopback, private, or link-local address, or to a bare or LAN hostname. So a model served by another box on your network counts the same as one on this machine — which matters to anyone with a GPU host in the basement:

| provider | endpoint | counted as |
|---|---|---|
| `homelab` | `http://server.home:11434/v1` | Local |
| `gpu` | `http://10.0.0.50:8000/v1` | Local |
| `lanbox` | `http://gpu-box:8080` (bare hostname) | Local |
| `ollama` | no URL; well-known runtime | Local |
| `openrouter` | `https://openrouter.ai/api/v1` | neither |
| `openai` | the subscription | Codex |

Provider ids with no URL to resolve (opencode and pi record the id but not the endpoint) fall back to a list of well-known runtimes — Ollama, LM Studio, llama.cpp, vLLM, LocalAI and friends — and a custom gateway can be named in `~/.config/omarchy/agents/local.json`.

There are no limits and no balance to draw: nothing meters self-hosted work, so the tab shows token history alone.

## Why it composes

- **It partitions rather than overlaps.** With the Codex fix in place, every rollout is counted by exactly one collector. On real data: 52,617 tokens before the fix = 31,580 on the Codex tab + 21,037 on the Local tab. No turn lost, none double-counted — asserted in the test, not just observed.
- **Cross-device aggregation needs no code.** `aggregateSnapshots()` builds provider accumulators on demand for any id, so with `syncMode` on, a homelab's numbers travel exactly like any other agent's. (Untested — I have only one machine.)
- **It self-hides.** Like every other agent, the tab appears only once there is usage to show, so this costs nothing for people who never run a local model.

## Deliberate non-goals

- **A third-party API is counted by nobody.** OpenRouter and friends are neither this subscription nor your own hardware. A case could be made for a fourth bucket; I did not want to invent one unasked.
- **No `assets/local.svg`.** The bar glyph stands in, and a mark invented by a drive-by contributor seemed like the wrong thing to put in this PR.

## Status

Implemented and working in #12034, stacked on #12032. Happy to drop it if this is not a direction you want the panel to go — filing the request first so the design can be argued before the code is.