feat(harness): per-session model + thinking/effort-level selection

[oxoxDev]

Summary

Add per-session model + thinking/effort-level selection to the coding-agent harness/daemon. Today an agent launches with a single daemon-global model and no notion of effort/thinking level; this makes model choice a first-class, per-task property with a per-provider catalog, threads it into each agent's real CLI flag, and surfaces the active model on the event stream.

Implementation home for the [OpenHuman] agent-runtime abstraction epic (tinyhumansai/openhuman#4731). Reference pattern only — clean-room, no third-party source copied.

Problem (current state, sdk/typescript/src/cli/)

• Model is daemon-global. --model (daemon.ts) → DaemonRuntimeDeps.model → merged into every runTask. There is no model field on TaskFrame (daemon/protocol.ts), so the protocol cannot carry a per-task model.
• No effort / thinking / reasoning-level control anywhere.
• claude never receives a model flag. buildRunArgs (daemon/providers.ts) threads -m for codex/opencode only; the claude arm omits it entirely.
• No model catalog. No per-provider list of available models or effort levels for a UI to offer.
• event.model is dead. Declared on SessionEnvelopeV2.event (types/harness.ts) and honored by buildEventEnvelopeV2, but HarnessSessionTailer.v2Context() never sets ctx.model, so it is always undefined. SessionInfoPayload.model is likewise never emitted.

Approach

1. Catalog — a new per-provider table (e.g. daemon/models.ts) of { models[], effortLevels[], defaultModel }. Static tables first (claude/codex/opencode + future providers); self-report probing is a follow-up.
2. Per-task carry — add optional model + effort to TaskFrame (daemon/protocol.ts, plus encodeTaskFrame/decodeTaskFrame); read frame.model/frame.effort in handleTask (daemon/runtime.ts) instead of the daemon-global.
3. Flag mapping — extend buildRunArgs (daemon/providers.ts) to map model+effort → each provider's real flag: add the missing claude --model, codex reasoning-effort (-c model_reasoning_effort=…), opencode -m. Single site.
4. Surface on the stream — set ctx.model in v2Context() (harness-wrapper.ts) and emit SessionInfoPayload.model so consumers show the active model.
5. Expose the catalog — advertise the per-provider catalog over the daemon protocol / a query so OpenHuman's UI can render a picker.

Acceptance criteria

• TaskFrame carries optional model + effort; round-trips through encode/decode with a test.
• buildRunArgs maps model+effort to the correct flag for every provider incl. claude; unit-tested per provider.
• Per-provider catalog exists and is queryable; default model/effort applied when a task omits them.
• event.model is populated end-to-end on the interactive stream (regression test asserts non-undefined).
• No blanket model default silently overrides an explicit per-task choice.
• Tests + typecheck + lint green; ready for openhuman vendor/tinyplace submodule bump.

Related

• Epic (OpenHuman abstraction layer): Agent runtime abstraction layer — clean interface, orchestration UX, and end-to-end testing for Cursor/Windsurf/Codex openhuman#4731
• Tracking issue (OpenHuman): linked below
• Sequenced before the ACP shared-transport slice (ACP's session/set_model consumes the per-task model field this adds).

[oxoxDev]

Tracking issue (OpenHuman): tinyhumansai/openhuman#4828 · Epic: tinyhumansai/openhuman#4731