feat(models): add Sonnet 1M and GPT-5.4 with usage warnings + recover…

[aletc1]

Summary

• Sonnet 4.6 1M is now selectable alongside Opus 1M. Backend strips the [1m] suffix and enables the shared context-1m-2025-08-07 beta header for either Opus or Sonnet — same code path, no per-model branching.
• GPT-5.4 + GPT-5.4 mini added to the Codex model list, with both the backend default (gpt-5.4/high) and the frontend last-selected default flipped from gpt-5.3-codex to gpt-5.4. 5.3-codex stays available; users with stored preferences are unaffected.
• Cost warning badge: any model with a [1m] id renders an amber "1M · higher cost" tag in the model selector with a tooltip, so users see the pricing tradeoff before opting in.
• Recovery action on errors: when a 1M model hits a rate-limit / quota error, the toast action becomes "Switch to " instead of "Copy Error" — one click moves the active sub-chat back to the 200K variant. Description is augmented to explain why.

Pricing for gpt-5.4 / gpt-5.4-mini was already in pricing.ts so billing flows through unchanged.

Files Changed

• src/renderer/features/agents/lib/models.ts — register sonnet[1m], gpt-5.4, gpt-5.4-mini
• src/renderer/features/agents/ui/agent-context-indicator.tsx — 1M context window for Sonnet 1M
• src/renderer/features/agents/atoms/index.ts — whitelist + MODEL_ID_MAP + flip Codex default
• src/main/lib/trpc/routers/claude.ts — generalize 1M beta-header to any [1m] suffix; thread model into error debugInfo
• src/main/lib/trpc/routers/codex.ts — DEFAULT_CODEX_MODEL → gpt-5.4/high
• src/renderer/features/agents/components/agent-model-selector.tsx — "1M · higher cost" badge
• src/renderer/features/agents/lib/ipc-chat-transport.ts — swap toast action to "Switch to " on 1M rate-limit errors

Test plan

• Open the model selector — verify Sonnet 4.6 1M appears with the amber "1M · higher cost" badge
• Select Sonnet 4.6 1M, send a message — verify the child-process log shows [claude] 1M context enabled for sonnet — ANTHROPIC_BETAS=…,context-1m-2025-08-07
• Verify the context indicator tooltip shows … / 1.0M context
• Open the Codex section — verify GPT-5.4 and GPT-5.4 mini appear above Codex 5.3
• Fresh install / cleared localStorage — verify the default Codex model is GPT-5.4
• Existing user with stored gpt-5.3-codex selection — verify it is preserved (5.3 still in the list)
• Trigger a rate-limit error on a [1m] model (or simulate by injecting RATE_LIMIT_SDK with a model: "sonnet[1m]" debugInfo) — verify the toast action reads "Switch to sonnet" and clicking it switches the sub-chat
• Same error on a non-1M model — verify the toast still shows "Copy Error"

Sources for the model decisions

• OpenAI Codex Models — confirmed gpt-5.4 (no -codex suffix)
• Anthropic Context Windows — confirmed shared context-1m-2025-08-07 beta for Sonnet 4.6
• Anthropic Effort Parameter — confirmed Sonnet 4.6 supports low/medium/high/max (no xhigh — Opus 4.7 only)

🤖 Generated with Claude Code