[Model] deepseek-v4-flash repeatedly emits the meaningless token "curd" in long tool-calling sessions (Claude Code CLI + Go plan)

Description

Setup / Environment

• Frontend: Claude Code CLI v2.1.237 (NOT the OpenCode TUI) — so OpenCode /share is unavailable and no share link can be provided from here.
• Model access: deepseek-v4-flash via the OpenCode Go plan (subscription-based model access). Observed on 2026-08-20.
• OS: Windows 11; Terminal: Windows Terminal (PowerShell + Git Bash).

What happened

In long, high-context turns (documentation edits, test runs, git stage/commit/push), the model repeatedly emits the token curd after its normal reasoning text and before the intended tool call. Runs range from a single curd to thousands of consecutive repeats. It requires a manual interrupt to stop.

Measured evidence (extracted from the session transcript, all from assistant output of deepseek-v4-flash):

Metric	Value
Assistant messages affected	26
Total curd tokens emitted	8,298
Single-message peak	3,886 (second peak: 3,299)
Messages with ≥100 curd	6
Messages with ≥1,000 curd	2
First → last occurrence	2026-08-20 04:18 → 13:32 UTC (~9 h)
Per-message distribution	median 5 / mean 319

Transcript excerpts (verbatim, from the raw session log; all from assistant output, model deepseek-v4-flash; [before]→[curd]→[after] shows where the burst lands):

① small burst (1 token) — docs/test-edit turn
   [before] Let me update the ToolAccessGuardTest for the refined non-blocking dept behavior:
   [curd]    curd
   [after]   (end of message)

② medium burst (6 tokens) — test-run turn
   [before] Let me run the tool-service test explicitly and capture the tail output.
   [curd]    curdcurdcurdcurdcurdcurd

③ repeating bursts mid-message (27 tokens) — build-diagnose turn
   [before] Let me run without the grep filter to see what happened.
   [curd]    curdcurdcurdcurdcurdcurdcurdcurdcurd...
   [after]   Let me run the build capturing full output. curdcurd curdcurdcurdcurd Let me run...

④ peak burst (3,886 tokens) — docs-update turn
   [before] Now update the architecture README to note P3c.
   [curd]    curd\n\nLet me update the architecture REA[curd][curd]...(3,886 in a row, then the real text resumes)

Pattern observed

• Always: normal reasoning → burst of curd → actual tool call.
• Correlates with long context + chained multi-step tool calls (docs update, test runs, git commit/push). Absent on short single-turn replies.
• Same single token repeated with no variation — characteristic of a generation-side repetition / decoding loop, not a frontend rendering issue.
• Repeats even mid-message: the model emits curd, then a real sentence, then curd again (see excerpt ③).

Request
Please route this to the upstream model provider (DeepSeek). Happy to attach the raw transcript excerpt if useful.

Plugins

No response

OpenCode version

No response

Steps to reproduce

1. Use Claude Code CLI (v2.1.237) pointed at the OpenCode Go plan (deepseek-v4-flash)
2. Start a long, high-context task with chained tool calls (multi-file doc edits, test runs, git stage/commit/push)
3. After normal reasoning text and before a tool call, the model streams "curd" (single to thousands of repeats)
4. Short single-turn tasks do not trigger it — needs long context

Screenshot and/or share link

No response

Operating System

Windows 11

Terminal

Windows Terminal (PowerShell 5.1 + Git Bash)

No repository file, entry point, or test is identified; start by reviewing the raw session transcript and confirming the behavior with the stated Claude Code CLI, OpenCode Go plan, and deepseek-v4-flash setup. Done means the repeated token emission is reproduced and resolved or confirmed as an upstream DeepSeek issue, with the result routed to the model provider.