session/load returns the config.toml model/effort instead of the thread's, because thread/resume is always sent a modelProvider

[julianmesa-gitkraken]

Setup: codex-acp driven over ACP by Kepler (Electron client, spawns the wrapper via dist/index.js and resumes stored sessions with session/load). Reproduced on codex-acp 1.1.4 and 1.1.7, against codex-cli 0.144.6 app-server.

Problem

session/load returns the model and reasoning effort from ~/.codex/config.toml instead of the ones persisted on the thread. Any in-session model/effort switch is silently lost on resume.

The cause is the modelProvider field that loadSession/resumeSession always pass to thread/resume:

// dist/index.js (1.1.7:26277) — same in 1.1.4:26044
const response = await this.codexClient.threadResume({
  config: await this.createSessionConfig(...),
  cwd: request.cwd,
  modelProvider: await this.getResumeModelProvider(),  // never absent
  threadId: request.sessionId,
});

getResumeModelProvider() is getCurrentModelProvider() ?? "openai", so the field is always populated even with no gateway configured. Passing it makes the app-server discard the thread's stored model/reasoningEffort and re-resolve them from config.

Reproduction

1. Create a session and switch it to a model/effort that differs from ~/.codex/config.toml. Mine has model = "gpt-5.5", model_reasoning_effort = "xhigh"; the session was set to gpt-5.6-sol / max.
2. Send a prompt. The rollout jsonl records the switch: "model":"gpt-5.6-sol", "effort":"max" in turn_context.
3. Resume that thread with session/load.

Observed on both versions:

currentModelId= gpt-5.5[xhigh]
  model            => "gpt-5.5"
  reasoning_effort => "xhigh"

Expected gpt-5.6-sol / max (what the thread holds).

Isolating it to modelProvider

Calling thread/resume directly on the app-server for the same thread, varying only which params are sent:

params sent	model	reasoningEffort
threadId, cwd	gpt-5.6-sol	max
+ config	gpt-5.6-sol	max
+ modelProvider: "openai"	gpt-5.5	xhigh
+ config + modelProvider (what codex-acp sends)	gpt-5.5	xhigh

config is irrelevant; modelProvider alone flips it. The values it falls back to are exactly the ones in config.toml.

Expected

Resuming a thread should restore the model and reasoning effort stored on that thread. Either don't send modelProvider on thread/resume when no gateway is configured, or have the app-server keep the thread's model when a provider is passed. If the app-server behaviour is the intended contract, this is arguably an openai/codex bug and I'm happy to move it there.

Impact

The client cannot recover the picks on its own: after a client restart there is no live session left to read them from, and the resume response is the only source of truth. Users see the correct model for a moment (from the client's own cache), then it flips to the config default as the load response lands.

Possibly the same root cause as #336, which reports effort reverting between turns of a live session — that one survives without a resume, so it may be a separate path. Filing this separately since the trigger here is isolated to session/load.