Activity

1. github-actions

   assigned

   kitlangton

   on Aug 31, 2026
2. github-actions commented on Aug 31, 2026

   

   github-actionsBot

   on Aug 31, 2026 – with GitHub Actions

   Contributor

   More actions

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

   • sap-ai-core: Claude models fail with 400 "does not support assistant message prefill" #45313: sap-ai-core Claude models fail with 400 "does not support assistant message prefill" — directly reports the same symptom (every SAP AI Core Anthropic turn yields a 400 prefill-forbidden error) and identifies the same root cause.
   • v1.18.21 repeatedly continues complete responses with finish=unknown #43939: v1.18.21 repeatedly continues complete responses with finish=unknown — covers the unbounded continuation loop when a provider returns complete output but labels the finish as unknown.
   • Empty completions with an unknown finish reason retry without any bound #45315: Empty completions with an unknown finish reason retry without any bound — related unbounded retry behavior on unknown finish.
   • core(v2): "unknown finish reason" error omits raw provider finish_reason #44092: core(v2): "unknown finish reason" error omits raw provider finish_reason — tracks the narrow finish-reason enum and the collapse to unknown for unrecognized values, which is the same mapping issue described here.
3. lgarceau768 commented on Sep 2, 2026

   

   lgarceau768

   on Sep 2, 2026

   More actions

   Same root cause on a second provider, and this time with the recorded finish value rather than a histogram.

   opencode 1.18.26, amazon-bedrock/anthropic.claude-opus-5 — so not gateway-specific, and not a provider that misreports its finish reason.

   Last three messages of the session, out of opencode.db:

   role	finish
   assistant	tool-calls
   assistant	unknown
   assistant	(APIError)

   The error on that last one:

   This model does not support assistant message prefill.
   The conversation must end with a user message.
   

   The unknown message matches your suggested predicate exactly: time.completed set, a 1192-character text part, and no tool part — its parts are step-start, reasoning, text, step-finish. Content produced, nothing pending, and the loop continued into a trailing-assistant request anyway.

   ["tool-calls","unknown"].includes(...) is still the exit condition in 1.18.26, at three call sites.

   The user-visible symptom is different enough from yours to be worth noting: a long reply appears to stop mid-sentence, then the 400 lands. Nothing is retryable in-session.

   One thing for prioritisation: on Bedrock there is no model choice that avoids this. Assistant prefill 400s across the current Claude family, so every Anthropic model in that catalogue hits the same path.
4. amelvil2-ford commented on Sep 3, 2026

   

   amelvil2-ford

   on Sep 3, 2026

   Last edited by amelvil2-ford

   More actions

   I haven't used this provider, but does this fix help you? #47079

   Let me know how it goes
5. ton618to commented on Sep 5, 2026

   

   ton618to

   on Sep 5, 2026

   Author

   More actions

   Thanks for the pointer @amelvil2-ford.

   I don't think #47079 resolves this specific case, and for SAP AI Core / Bedrock it would trade the 400 for a per-turn hard error.

   The difference is subtle:

   • Empty completions with an unknown finish reason retry without any bound #45315 (which fix: Improve handling for streams that end without a finish reason #47079 closes) is about streams that end without a finish reason → unbounded retry.
   • This issue is a stream that does end with a finish reason, but the provider's value isn't in opencode's FinishReason enum, so ai-sdk.ts normalizes it to "unknown" — on a turn that produced a complete text response.

   For sap-ai-core and amazon-bedrock Anthropic models, "unknown" is the value on every successful turn (see the DB evidence above and @lgarceau768's Bedrock confirmation). #47079's default path treats anything outside stop | tool-calls | length | error | content-filter as ok: false and throws unrecognized finish reason "unknown". So out of the box it would turn every completed SAP/Bedrock turn into an error, and you'd need a finish.chunk plugin that flips ok = true for "unknown" just to get back to a working state. The finish.chunk hook itself is a nice addition, but as a default it swaps one per-turn failure for another on these providers.

   The proximate cause is the loop-exit predicate !["tool-calls", "unknown"].includes(finish) (still present at the three call sites @lgarceau768 noted). The narrow fix that avoids both the prefill 400 and a spurious error is to treat an "unknown" finish as terminal when the turn already produced text and has no pending tool calls, while keeping the existing continue-on-"unknown" behavior for empty/truncated turns (so #45315 stays covered). Concretely:

   const hasText =
     lastAssistantMsg?.parts.some((part) => part.type === "text" && part.text.trim().length > 0) ?? false
   
   if (
     lastAssistant?.finish &&
     lastAssistant.finish !== "tool-calls" &&
     (lastAssistant.finish !== "unknown" || hasText) &&
     !hasToolCalls &&
     lastAssistant.parentID === lastUser.id
   ) {
     // exit loop
   }

   I've been running exactly this as a local patch against 1.18.23 and it cleanly stops the prefill 400 on SAP anthropic--claude-4.8-opus with no error line and no regression in the prompt-loop tests. Mapping the provider's raw finish (end_turn/stop → stop) would be the more complete fix, but the predicate above is a safe minimal guard in the meantime.
6. Dawncoral commented on Oct 28, 2026

   

   Dawncoral

   on Oct 28, 2026

   More actions

   Confirming the prefill path on 1.18.35, plus a version boundary.

   Harness: a local OpenAI-compatible mock that returns finish_reason: "unknown" and rejects any request whose last message is assistant, with the prefill 400.

   1.18.35 request sequence:

   system, user, user        (title call)
   system, user              -> 200, finish_reason "unknown"
   system, user, assistant   -> 400 -> CLI prints the prefill error
   

   On 2.2.0 the same harness no longer produces a trailing assistant message: the client appends a synthetic user turn ("The previous response was interrupted...") and keeps auto-continuing (12 requests in ~90s here). That removes the 400 but the loop remains; #56295 covers the unmarked synthetic message.

   The steps path (#56250) has an open fix in #56257. Happy to share the ~100-line Python mock if it helps.
7. wayzha commented on Oct 28, 2026

   

   wayzha

   on Oct 28, 2026

   More actions

   v2.2.0: prefill loop fixed ✅, but Schema validation failed still fires on every SAP AI Core turn
   The prefill continuation loop described in this issue no longer occurs on v2.2.0 — the model responds correctly and no spurious 400 is produced. However, every successful SAP AI Core response still triggers:
   level=ERROR message="Failed to drain Session"
   cause="Error: Schema validation failed"
   [cause]: LLM.Event.StepFinish → reason → normalized: AnyOf failed
   The underlying root cause identified here remains unfixed: FinishReason in packages/llm/src/schema/ids.ts:39 does not include "other", and the adapter in packages/opencode/src/session/llm/ai-sdk.ts:22-24 maps it to "unknown". While the run loop now handles "unknown" without looping, the schema validation on StepFinish still rejects it because the normalized value fails the strict 6-literal union check.
   The fix should either add "other" to the FinishReason schema, or ensure the mapped "unknown" value passes through schema validation during drain.
   Environment: v2.2.0 macOS, aisdk:@jerome-benoit/sap-ai-provider-v2@4.4.11, tested with sap-ai-core/gpt-6-astra and other SAP AI Core models.
8. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   Thanks for the detailed root-cause analysis!

   The version you're running (1.18.23) is OpenCode v1, and the code paths you traced (session/prompt.ts, session/llm/ai-sdk.ts) are v1 code. v1 now only gets fixes for critical bugs such as data loss, security problems, or breakage with no workaround. Here the answer is still delivered and the extra 400 comes after it, so this is unlikely to be fixed in v1.

   We'd encourage you to try OpenCode v2 (@opencode/cli, the opencode command). It's a separate implementation, and it doesn't send a prefill continuation when the finish reason is unknown. It treats that case as a provider error instead, so you shouldn't get this exact 400 there. We haven't tested that with SAP AI Core, though. Related v2 issues you may want to follow:

   • Step-capped agents fail with an assistant prefill 400 #56250: a different case on v2 (agents hitting their step limit) that gets the same assistant-prefill 400
   • bug(v2): SAP AI Core initialization fails  #54791 and bug(v2): SAP AI Core unusable with valid saved credentials — apiKey/serviceKey mismatch blocks authentication #56126: SAP AI Core setup/credential problems on v2

   If you think this is critical, or you see the problem on v2, please reply here with your opencode --version and what you see.
9. opencodejr

   added

   pending close

   triaging

   on Oct 28, 2026
10. wayzha commented on Oct 29, 2026

    

    wayzha

    on Oct 29, 2026

    More actions

    Thanks for the detailed root-cause analysis!

    The version you're running (1.18.23) is OpenCode v1, and the code paths you traced (session/prompt.ts, session/llm/ai-sdk.ts) are v1 code. v1 now only gets fixes for critical bugs such as data loss, security problems, or breakage with no workaround. Here the answer is still delivered and the extra 400 comes after it, so this is unlikely to be fixed in v1.

    We'd encourage you to try OpenCode v2 (@opencode/cli, the opencode command). It's a separate implementation, and it doesn't send a prefill continuation when the finish reason is unknown. It treats that case as a provider error instead, so you shouldn't get this exact 400 there. We haven't tested that with SAP AI Core, though. Related v2 issues you may want to follow:

    • Step-capped agents fail with an assistant prefill 400 #56250: a different case on v2 (agents hitting their step limit) that gets the same assistant-prefill 400
    • bug(v2): SAP AI Core initialization fails  #54791 and bug(v2): SAP AI Core unusable with valid saved credentials — apiKey/serviceKey mismatch blocks authentication #56126: SAP AI Core setup/credential problems on v2

    If you think this is critical, or you see the problem on v2, please reply here with your opencode --version and what you see.

    I'm already on v2.2.0 (opencode --version → v2.2.0). The prefill 400 is indeed gone ✅.

    However, every SAP AI Core response still triggers Schema validation failed during SessionRunner.drain — the LLM.Event.StepFinish.reason.normalized field fails validation because the AI SDK's "other" finish reason is not in v2's FinishReason enum either. See my earlier comment above for the full error.

    This is non-blocking (the response does render), but it fires on every single turn with any aisdk: provider that returns finishReason: "other".
11. opencodejr

    closed this as not plannedon Oct 29, 2026
12. opencodejr commented on Oct 29, 2026

    

    opencodejrBot

    on Oct 29, 2026

    More actions

    Closing this as not planned. It was reported on v1 (1.18.23), where only critical bugs get fixed. The extra 400 after each turn is annoying, but the answer is still delivered, so it doesn't meet that bar. We waited a day for a reply in case you saw this on v2 or thought it was critical, and haven't heard back.

    OpenCode v2 (@opencode/cli) doesn't send a prefill continuation on an unknown finish reason. If you see the same 400 with SAP AI Core on v2, please open a new issue with your opencode --version and we'll look into it. Thanks for the detailed analysis!