Live model tracking after set_model (AgentClient.CurrentModel)

Root cause

docs/design-notes/2026-07-31-history-restore-flow-and-model-picker.md section 3 documented (and the implementation followed) “the current model” as client.InitMessage.Model — the resolved model name captured once from the system/init message at connect time. That was correct as long as AgentClient.InitMessage was the only place a model could ever be recorded.

It stopped being correct the moment AgentHub.SwitchModel started sending a live set_model control_request (same note, same section): a successful set_model genuinely changes which model answers the next turn (R02 section 5.4), but nothing ever touched InitMessage.Model afterwards — AgentClient.HandleControlResponse only logs failures and raises ControlRequestResolved(kind, success, error); it never mutates any “current model” field. Every reader of “the current model” (StatusBarView.ResolveModelName, HeaderView’s picker button label, its GenericMenu checkmark, and StatusBarView.RefreshContextMeter’s currentResolvedModel key for SelectPrimaryModelUsage) reads InitMessage.Model directly, so all four kept showing the pre-switch model until the next full reconnect re-ran system/init.

The ctx meter compounds this: AgentHub._lastModelUsage was also left untouched by SwitchModel (only StartFresh/SwitchToSession cleared it), so after the next turn completed under the new model, SelectPrimaryModelUsage’s exact-match branch keyed off the stale currentResolvedModel failed to find the new model’s modelUsage entry and silently fell back to the largest-token heuristic — which can pick the wrong entry entirely when a helper model (e.g. title generation) is also present in the same result.

Options considered

1. Have set_model’s control_response carry the resolved model name and read it off the response. Rejected: no captured fixture or research note documents the set_model success response’s payload shape (unlike can_use_tool/initialize, R02/02b never captured one against a live process). Depending on an unverified wire field would repeat exactly the mistake AgentClient.cs’s --allowedTools comment already warns about elsewhere in this file. LiveCliIntegrationTests.SetModel_LiveRoundTrip only asserts success/error, not response shape.
2. Re-run the whole initialize handshake after every set_model. Rejected: wasteful (a second full commands/agents/models round trip for a one-field change) and set_model is documented as a lightweight “keeps running” control request, not a resync point.
3. Track the requested model value locally and resolve it against the ALREADY-CACHED InitializeResponse.Response["models"] array (the same value -> resolvedModel mapping HeaderView.ParseModels already uses to render the picker), committing it as the live model only once the matching control_response reports success. Chosen.

Decision

AgentClient gains:

• A private _liveModel field, reset to null in Start() (a fresh system/init is always authoritative again).
• SetModel(model) resolves model (the alias/value the caller passed, e.g. "haiku") against InitializeResponse.Response["models"] the same way HeaderView.ParseModels does (value -> resolvedModel), falling back to the raw model string when no match is found (manual/unlisted model ids still get a best-effort label instead of silently keeping the old one). It then tracks the outbound request with a callback (the PendingRequestMap.Track callback parameter already existed and was unused here) that commits _liveModel ONLY when the resulting control_response is Success == true. A failed or timed-out set_model leaves _liveModel untouched, matching the CLI’s own behavior of leaving the previous model running.
• A new public CurrentModel property: _liveModel ?? InitMessage?.Model. This is additive — InitMessage.Model itself is untouched, so nothing that legitimately wants “the resolved model system/init reported” (none of the current call sites do) is affected.

StatusBarView.ResolveModelName, StatusBarView.RefreshContextMeter’s currentResolvedModel, and HeaderView.OnModelPickerClicked’s isCurrent comparison now all read client.CurrentModel instead of client.InitMessage.Model. This supersedes the “現在選択中モデルのハイライトは client.InitMessage.Model…と比較する” line in the 2026-07-31 history-restore-flow-and-model-picker.md note section 3 — read CurrentModel wherever that note says InitMessage.Model for “the current model” from now on.

AgentHub.SwitchModel additionally resets _lastModelUsage to a fresh empty dictionary (the same pattern StartFresh/SwitchToSession already use), so the ctx meter shows “no data” (hidden) rather than a stale model’s percentage under the new model’s label until the next turn’s modelUsage arrives.

Tradeoffs accepted

• If the CLI ever silently ignores a model value that IS in the models[] list (accepts the control_request with success:true but does not actually switch), the panel would show the new label anyway. No evidence of this happening has been captured; the alternative (never trusting success:true) would make the picker permanently stuck showing the old model, which is strictly worse for the common case.
• CurrentModel resets to InitMessage.Model only on the next Start() (full reconnect), not on e.g. Stop() alone — consistent with every other per-connection field this class already resets exactly once, in Start().

Regression-guard tests

All in AgentClientStateTests.cs (Core-layer coverage; StatusBarView.ResolveModelName and HeaderView’s picker are thin one-line read-throughs of AgentClient.CurrentModel with no InternalsVisibleTo from Tests/Editor into the UI assembly, so the behavior both actually delegate to is exercised at the AgentClient level instead):

• CurrentModel_MirrorsInitMessage_WhenNoLiveSwitchHasHappened
• SetModel_CommitsLiveModelOnSuccess_AndResolvesAliasViaModelsList — asserts the resolved name ("claude-haiku-4-5-20251001"), not the raw alias ("haiku"), only commits once the matching control_response arrives, and that InitMessage.Model itself stays untouched.
• SetModel_DoesNotCommitLiveModelOnFailureOrTimeout
• SetModel_UnmatchedValue_FallsBackToRawValueOnSuccess
• Start_ResetsLiveModel_SoAFreshConnectIsAuthoritativeAgain

AgentHub.SwitchModel’s _lastModelUsage reset was left without a dedicated test (the field is private/static and only observable through AgentHub.LastModelUsage; the pre-existing StartFresh/SwitchToSession clears of the same field were also never given a dedicated regression test) — consistent with this codebase’s existing test-coverage boundary for AgentHub’s static composition-root methods.