Title: opencode-go / deepseek-v4-flash: continuing a long session returns 400 with opaque body {"object":"error","model":"deepseek-v4-flash"} #48093

Activity

1. github-actions

   assigned

   rekram1-node

   on Sep 9, 2026
2. github-actions commented on Sep 9, 2026

   

   github-actionsbot

   on Sep 9, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • [Bug] reasoning_content "invalid_request" on follow-up messages with OpenCodeZen DeepSeek-V4-Flash in thinking mode #41061: Nearly identical issue — reasoning_content invalid_request on follow-up messages with OpenCodeZen DeepSeek-V4-Flash in thinking mode
   • DeepSeek silently stops executing (interleaved reasoning_content dropped in tool call messages) #35689: Same root cause — reasoning_content dropped in assistant tool-call messages causing 400 errors with DeepSeek thinking mode across turns
   • Non-DeepSeek custom providers fail when reasoning_content is sent in tool calls or subsequent messages #37475: Related — non-DeepSeek providers also fail when reasoning_content is sent back in subsequent messages
3. hpcx1234 commented on Sep 22, 2026

   

   hpcx1234

   on Sep 22, 2026

   More actions

   Another data point: deepseek-v4.1-flash#max on Desktop 2.0.12 — same opaque-body 400, session-history dependent

   • Provider/model: opencode-go/deepseek-v4.1-flash#max (OpenCode Go, https://opencode.ai/zen/go/v1); OpenCode Desktop 2.0.12, Windows.
   • Session: long-lived (2 days, ~22k events, heavy tool use; ~2 MB of active context, 82 assistant messages since the last compaction).
   • 5 failures on 2026-09-22 between 07:12 and 07:27 UTC. Each failed step produced an empty assistant message (finish=error, content: []) and the client showed only Provider request failed with HTTP 400 — i.e. the body carried no error.message / message, consistent with the 46-byte {object:error,model:...} shape reported above.
   • After the first failure, retrying inside the same session kept failing (3 consecutive failures, including plain continue prompts). Other sessions on the same model + same variant + same account worked fine — so it is request-history dependent.
   • INFO logs only show Failed to drain Session plus the JS stack; request/response bodies are not logged, so the upstream reason is not recoverable client-side.
   • Workaround that matches Console Go: DeepSeek V4.1 Flash #max fails on long multi-step runs — reasoning_content not passed back (HTTP 400) #48364: switching the session's model (drop #max or switch model) resolves it.

   Looks like the same class as #48180 (reasoning_content echo-back × reasoning_effort), but on deepseek-v4.1-flash with an opaque body. Happy to provide exact timestamps / message IDs if useful.
4. SamSpiri commented on Sep 29, 2026

   

   SamSpiri

   on Sep 29, 2026

   More actions

   Same class as the newer reports, on the older deepseek-v4-flash id. A client-side retry for the opaque body, looking for testers:

   https://github.com/SamSpiri/opencode-go-stub-400-retry

   It covers deepseek-v4-flash and deepseek-v4.1-flash, and matches only the body with no error information ({"object":"error","model":"..."}, the Provider request failed with HTTP 400 fallback, or a bare Bad Request).

   Worth separating first, though, because this issue's thread mixes two causes: if your failures are the thinking-mode reasoning_content contract that the earlier comments hypothesise, retrying will not help — those are deterministic and switching model or dropping the #max variant is the lever (see the sibling report). The plugin only addresses the transient stub. My own install has not fired yet at a ~1 % failure rate, so I am explicitly asking for testers rather than claiming it works.
5. erflores-tech commented on Oct 9, 2026

   

   erflores-tech

   on Oct 9, 2026

   More actions

   Cross-reference from another client (Cline) that reproduces this on the same opencode-go / deepseek-v4.1-flash Console Go route.

   • Cline issue: Desktop: reasoning_content in the thinking mode must be passed back to the API - Cline replays tool_use-only assistant turns with no reasoning block (opencode-go / deepseek-v4.1-flash) cline/cline#14965

   Finding (from on-disk session evidence): the failing session persisted 41 of 209 assistant turns as tool_use-only with no reasoning block. The request that failed replayed one of those (tool_use -> tool_result -> next request) and returned 400; every retry failed identically.

   This matches the all-or-nothing rule captured in #25000: once any assistant message carries reasoning_content, every subsequent assistant message must too, or the relay rejects the whole request. Adding this as corroborating data from another client, not as a Cline-only bug.
6. thdxr commented on Oct 11, 2026

   

   thdxr

   on Oct 11, 2026

   Member

   More actions

   Thanks for the report! OpenCode 2.0 is now out and development is focused on V2, so we're closing issues filed against OpenCode 1. Please upgrade (opencode upgrade --major, details: https://anoma.ly/notes/opencode-2-0), and if this is still relevant on 2.0, reopen this issue or open a new one with your 2.0 version.

   — from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊
7. thdxr

   closed this as not plannedon Oct 11, 2026