## What happens

The agents bar widget attributes **every** rollout in `$CODEX_HOME/sessions` to the Codex subscription, including turns Codex CLI served from a local or third-party backend. On a machine that uses `codex` as a front-end for Ollama, the panel's "Tokens by day", "Tokens by model", session and prompt counts are all inflated by usage the subscription never paid for.

Real record from my machine, `~/.local/state/omarchy/agents/usage/codex.json`:

```json
"todayTokensByModel": {"qwen3-coder:30b": 21037, "gpt-6-astra": 31580}
```

`qwen3-coder:30b` is a local Ollama model. It is 40% of the number on that tab.

The rate-limit meters are unaffected — those come from the app-server RPC, not from session files — so this is a stats-accuracy bug, not a false quota warning.

## Why

`scan_native_codex_sessions()` in `bin/omarchy-agent-usage-codex` walks every `*.jsonl` under the session roots and counts all of it. It reads `model` from `turn_context` and never looks at the provider.

The provider *is* recorded, but on `session_meta`, which the scan skips past:

```
$ head -1 rollout-…-01a0a6c1.jsonl | jq '.payload.model_provider'
"ollama"
$ head -1 rollout-…-01a0a6c7.jsonl | jq '.payload.model_provider'
"openai"
```

`turn_context` carries `model` but no provider field at all, so the one record that knows is the one the scan ignores.

Notably the other two scanners in the same file already guard against exactly this:

- `scan_pi_sessions()` requires `provider == "openai-codex"`
- `scan_opencode_sessions()` does an exact match on `providerID == "openai"`, with a comment that reads "a custom `openai-local` gateway is not this subscription"

So the intent is established and consistent; the native scanner is the one path that never got the check.

## Reproduce

```bash
export CODEX_HOME=$(mktemp -d)
mkdir -p "$CODEX_HOME/sessions/$(date +%Y/%m/%d)"
cat > "$CODEX_HOME/sessions/$(date +%Y/%m/%d)/rollout-local.jsonl" <<'JSON'
{"type":"session_meta","payload":{"model_provider":"ollama"}}
{"type":"turn_context","payload":{"model":"qwen3-coder:30b"}}
{"type":"event_msg","payload":{"type":"token_count","info":{"last_token_usage":{"input_tokens":9000,"output_tokens":900}}}}
JSON

omarchy-agent-usage-codex --force | jq .todayTokensByModel
# {"qwen3-coder:30b": 9900}   <- a local model billed to the Codex subscription
```

Any non-OpenAI provider reproduces it. In testing, an OpenRouter-backed rollout showed up on the Codex tab as `"claude-sonnet-5"`, which makes the conflation especially visible.

## Suggested fix

Read `session_meta.model_provider` and skip the rollout when it is present and is not `openai`. Rollouts written before Codex recorded the field carry no provider and must keep counting, so the check should be "present and different", not "not equal".

## A related gap, for the record

`scan_projects()` in `bin/omarchy-agent-usage-claude` has the same structural shape — no endpoint filter on native transcripts — so pointing Claude Code at a proxy would conflate the same way.

It is **not** fixable by the same means, though: Claude Code transcripts record no provider or base URL. Dumping the fields from an assistant entry gives

```
cwd, entrypoint, gitBranch, isSidechain, parentUuid, requestId, sessionId,
timestamp, type, userType, uuid, version,
message.{content, diagnostics, id, model, role, stop_details, stop_reason, usage, ...}
```

Codex writes `session_meta.model_provider`; Claude Code writes nothing equivalent. The only available signal would be guessing from the model name, which seems too fragile to be worth it. Mentioning it so the bug class is on the record rather than looking overlooked — I would not want it bundled into the Codex fix.

## Environment

- omarchy 4.0.3-1, Arch
- codex CLI 0.154.0
- `omarchy.agents` bar plugin
- Present on `quattro`: `bin/omarchy-agent-usage-codex` is byte-identical to the installed 4.0.3 copy