v1.17.18: thinking leaks as literal text + TypeValidationError on usage (undefined) in message_delta (minimax-coding-plan/MiniMax-M3, Fedora/Kitty) #36387

Summary

Two distinct bugs that both reproduce in the same session on opencode 1.17.18, provider minimax-coding-plan/MiniMax-M3, Fedora + Kitty terminal. Both surface during streaming; the second one aborts the turn.

Bug 1 — thinking blocks leak as literal text between content blocks

During streaming, the assistant's reasoning / chain-of-thought appears as visible plain-text content bracketed by <think>…</think> (or equivalent), instead of being rendered in a collapsible reasoning panel like Anthropic / OpenAI o-series / DeepSeek outputs.

Expected: thinking content streams into a hidden/collapsible reasoning section (same as Claude, GPT-o, DeepSeek, etc.), invisible by default.

Actual: raw <think>…</think> text appears inline in the assistant message, between visible content blocks.

Bug 2 — TypeValidationError: usage is undefined on message_delta

When the stream reaches stop_reason: "", opencode aborts with:

Type validation failed: Value: {"type":"message_delta","delta":{"stop_reason":""}}.
Error message: [
  {
    "expected": "object",
    "code": "invalid_type",
    "path": ["usage"],
    "message": "Invalid input: expected object, received undefined"
  }
]

The provider emits a message_delta SSE event whose top-level usage field is undefined (Anthropic-compatible shape), but opencode's Zod schema requires usage to be an object on every message_delta. The turn dies before the final message_stop.

Repro

1. Install opencode 1.17.18.
2. Configure provider minimax-coding-plan/MiniMax-M3 (this is the Minimax coding-plan provider, see https://opencode.ai).
3. Send any prompt that triggers reasoning + a multi-token response.
4. Observe: thinking renders as literal text in the output, then the turn aborts with the validation error above.

Environment

• opencode: 1.17.18
• OS: Fedora (latest)
• Terminal: Kitty
• Provider: minimax-coding-plan/MiniMax-M3 (model: MiniMax-M3)
• Mode: TUI (default CLI)

Related issues

• [Bug] AI 推理链（thinking）泄露到可见输出，导致"自说自话"循环 #28659 — thinking leaks into visible output (similar symptom, different provider)
• OpenAI Chat parser treats standalone </think> before tool_calls as assistant text #34126 — OpenAI Chat parser treats standalone </think> before tool_calls as assistant text
• Too many thought messages with MiniMax-M3 #31999 — Too many thought messages with MiniMax-M3 (same model family)
• @ai-sdk/anthropic 3.0.71 rejects usage.iterations[].type: "fallback_message" — turn fails with AI_TypeValidationError #31579 — @ai-sdk/anthropic 3.0.71 rejects usage.iterations[].type: "fallback_message" (closed; AI SDK usage validation has been fragile here for months)

Suggested fix (skim)

1. Bug 2 (one-liner): make usage optional in the message_delta Zod schema for the Anthropic-shaped provider. Anthropic's SSE spec explicitly allows usage to be absent on message_delta (it is sent on message_start and optional on later deltas); current strict schema rejects valid streams.
2. Bug 1: route reasoning content through the same hidden/collapsible rendering pipeline used for Claude/GPT-o, regardless of provider — the parser is likely treating the provider's reasoning field as plain text instead of as a structured reasoning part.