Step-capped agents fail with an assistant prefill 400 #56250

Activity

1. opencode-agent commented on Oct 16, 2026

   

   opencode-agentBot

   on Oct 16, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the clear report and for pinpointing the cause. I reproduced this on 2.0.24, and it still happens on the latest dev (d4a6223).

   I set the build agent to steps: 2 and used a simulated model that calls read on step 1. On step 2, the last allowed step, the request OpenCode sends ends with an assistant message containing the "CRITICAL - MAXIMUM STEPS REACHED" prompt. Anthropic models that don't support assistant prefill reject that kind of request with the 400 you saw. The simulated model can't return Anthropic's real error, so it shows a marker in the TUI when a request ends on an assistant message.

   The line responsible is still in packages/core/src/session/runner/llm.ts: when the step limit is reached, it appends Message.assistant(MAX_STEPS_PROMPT) to the messages.

   The script below reproduces it, and the recording and screenshot show the run.

   Related, ordered by likelihood:

   • Step-cap assistant message causes 400 on Claude models with thinking enabled #32548: the same step-cap prefill 400 on v1 (closed).
   • bug(session): finish_reason 'unknown' triggers an assistant-prefill continuation; SAP AI Core (and other prefill-forbidding providers) 400 after every successful turn #46415: the same 400, caused by continuing after finish_reason: unknown.
   • openrouter: Gemini fails with 400 "Requests ending with a model turn are not supported" after read tool returns images, and error hides metadata.raw #55694: Gemini rejects requests that end on a model turn.

   recording.mp4

   

   Reproduction script (drive.ts)

   // Repro for #56250: on the last allowed step (agent `steps: 2`), OpenCode sends a
   // request whose final message is an assistant turn (MAX_STEPS prompt as prefill).
   // Providers/models that forbid assistant prefill (e.g. Anthropic Claude via GitLab Duo)
   // reject such a request with "The conversation must end with a user message".
   import { Effect, Stream } from "effect"
   import { Llm, OpenCodeDriver } from "opencode-drive"
   
   const roleOf = (m: any) => m?.role ?? m?.type
   const messagesOf = (body: any): any[] => body?.messages ?? body?.input ?? body?.contents ?? []
   const textOf = (m: any) => JSON.stringify(m?.content ?? m).slice(0, 80)
   
   export default OpenCodeDriver.use(
     {
       ...(process.env.OC_DEV ? { opencode: { dev: process.env.OC_DEV } } : {}),
       tui: { recording: true },
       config: { agents: { build: { steps: 2 } } } as any,
       project: { git: true, files: { "a.txt": "A\n", "b.txt": "B\n", "c.txt": "C\n" } },
     },
     ({ ui, llm }) =>
       Effect.gen(function* () {
         yield* llm.serve((request, index) => {
           const msgs = messagesOf(request.body)
           const last = msgs[msgs.length - 1]
           const role = roleOf(last)
           console.log(`REQUEST ${index + 1}: ${msgs.length} messages, last role = ${role}, last = ${textOf(last)}`)
           if (role === "assistant") {
             // Stand-in for the provider's documented 400 for prefill-forbidding models.
             return Stream.make(
               Llm.text(`PREFILL: request ${index + 1} ends with an assistant message -> Anthropic would return 400 "This model does not support assistant message prefill. The conversation must end with a user message."`),
             )
           }
           if (index < 3)
             return Stream.make(
               Llm.toolCall({ index: 0, id: `call_${index}`, name: "read", input: { path: ["a.txt", "b.txt", "c.txt"][index] } }),
               Llm.finish("tool-calls"),
             )
           return Stream.make(Llm.text("Summary: read the files."))
         })
         yield* ui.submit("Read a.txt, b.txt and c.txt one at a time, then summarize.")
         yield* ui.waitFor(
           () => Effect.map(Effect.all([ui.matches("PREFILL:"), ui.matches("Summary:")]), ([a, b]) => a || b),
           { timeout: 60000 },
         )
         yield* Effect.sleep(1500)
         yield* ui.screenshot("final")
       }),
   )

   Run it with opencode-drive: opencode-drive run ./drive.ts. It drives opencode from your PATH; set OC_DEV to the path of an OpenCode checkout to run that from source instead.
2. opencode-agent

   added

   reproduced

   on Oct 16, 2026
3. tlmquintino commented on Oct 18, 2026

   

   tlmquintino

   on Oct 18, 2026

   More actions

   Also seen on 2.0.28 with the anthropic provider (claude-sonnet-5-5), on subagents (mode: subagent, steps: 60) launched through the subagent tool. The impact for subagents is worse than a failed turn: the child's file changes up to the cap stay on disk, but the parent receives only the error, no summary, so its work is silently partial (logged as "Failed to drain Session … does not support assistant message prefill"). We hit it on three long-running subagent tasks in one day. Sending the MAX_STEPS prompt as a user message, as suggested above, would fix it for us.