Console Go: DeepSeek V4.1 Flash #max fails on long multi-step runs — reasoning_content not passed back (HTTP 400) #48364

Summary

Multi-step agentic runs using DeepSeek V4.1 Flash with max reasoning effort
via OpenCode Go ("Console Go") consistently fail with HTTP 400
The reasoning_content in the thinking mode must be passed back to the API.
The same long task succeeds with the same model without the #max variant,
and with glm-5.3-flash#max, so reasoning content is dropped on some replay
path specific to this model + max effort.

Environment

• opencode version: v0.0.0-next-17403 (opencode2)
• OS: Linux 7.0.0-28-generic x86_64 (Ubuntu 24.04 based)
• Terminal: TERM=xterm-256color
• Shell: /bin/bash
• Install/channel: next
• Active plugins: none found in config

Reproduction

1. Register a custom model under the opencode-go provider mapping to the
   new upstream id (released 2026-09-10):

   "providers": { "opencode-go": { "models": {
     "deepseek-v4.1-flash": {
       "modelID": "deepseek-flash",
       "capabilities": { "tools": true, "input": ["text", "image"], "output": ["text"] },
       "limit": { "context": 1000000, "output": 384000 },
       "variants": [
         { "id": "low", "settings": { "reasoningEffort": "low" } },
         { "id": "high", "settings": { "reasoningEffort": "high" } },
         { "id": "max", "settings": { "reasoningEffort": "max" } }
       ]
     }
   } } }

2. Create a read-only subagent with model: opencode-go/deepseek-v4.1-flash#max
   (deny edit/shell, steps: 8).
3. Give it a task requiring ~7+ sequential tool calls (e.g. read 7 files +
   a grep) so the run needs many thinking+tool-call rounds.
4. After several successful reasoning + tool-call steps, a later model call
   fails and the session drain errors out.

Expected Behavior

reasoning_content from thinking blocks is passed back automatically on
every continuation (as happens with glm-5.3-flash#max and with this same
model without a variant), and the run completes.

Actual Behavior

The run fails with:

Provider request failed with HTTP 400: Error from provider (Console Go):
Upstream request failed: [invalid_request_error] The `reasoning_content`
in the thinking mode must be passed back to the API.

Failing assistant message has finish: error, empty content, zero tokens.
Stack: SessionRunner.callModel (core/src/session/runner/llm.ts) via
RequestExecutor.execute (ai/src/route/executor.ts).

Additional Context

• Frequency: 3/3 failed sessions on long tasks with #max; short tasks
  with #max succeed; the identical long task with the base model
  (no variant) succeeds end-to-end. Possibly the final
  summarize-without-tools call after steps exhaustion replays history
  without reasoning_content, but not confirmed.
• Workaround: drop the #max variant (use the base model).
• Related: model released 2026-09-10 (deepseek-flash canonical API id;
  OpenCode listed as day-one partner); opencode-go/deepseek-flash
  catalog entry exists but the v4.1 id does not yet, hence the custom
  mapping.