v3.26.3

Fixed

• The Claude model picker listed models that no longer exist, and switching between them did nothing. Two independent defects that presented as one broken control:
  • The list was a hardcoded manifest constant. backends/claude-code/manifest.ts pinned claude-opus-4-7 / claude-sonnet-4-6, so the picker still offered Opus 4.7 long after Opus 5 and Fable shipped. Codex fills available_models from its model/list RPC and Kimi from ACP; claude-code never filled it at all, so the switcher always fell through to that static list. The CLI does expose its own list — control_request { subtype: "initialize" } answers with models[], the same source the Agent SDK's query.supportedModels() reads — and the bridge now fires that probe when the CLI transport attaches, publishing the result through the very same available_models field the other two backends use. system.init does not carry the list (only the single active model), so this is the only route.
  • set_model was a no-op. A top-level {"type":"set_model"} frame is accepted by the stream-json reader, silently discarded, and the next turn runs on the old model anyway; only the control-request form takes effect. Measured on Claude Code 2.1.220 with one prompt: the top-level frame billed to claude-opus-5, the control request to claude-haiku-4-5-20251001. The bridge's optimistic UI update is what hid this for so long — the label changed, the billed model didn't. A rejected switch now snaps the picker back instead of leaving it naming a model the session isn't running.

Changed

• available_models entries carry an optional resolvedId — Claude Code lists aliases (default, opus[1m]) that are many-to-one with concrete models, so "which row is active?" has to resolve to a single winner rather than run a predicate per row. The manifest's defaultModels survives only as a fallback for CLIs too old to answer the probe, and is now expressed as aliases (opus / sonnet / haiku) rather than pinned ids — an alias tracks whatever the installed CLI treats as current, so it cannot rot the way the pinned trio did.

Notes

• backends/claude-code/README.md had documented the wrong wire shape for set_model and asserted that Claude Code has no model-discovery endpoint; both are corrected, and .claude/rules/backends.md records the two traps.