core: signature-only responses complete successfully with no visible output #46880

Activity

1. github-actions

   assigned

   kitlangton

   on Sep 2, 2026
2. opencode-agent

   added

   2.0

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

   

   github-actionsBot

   on Sep 2, 2026 – with GitHub Actions

   Contributor

   More actions

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

   • v2: empty reasoning-only response is recorded as successful completion #37372: v2: empty reasoning-only response is recorded as successful completion — same root cause (reasoning-only step with no text or tool call recorded as session.execution.succeeded), same V2 event model. Your report adds useful specifics about publish-llm-event.ts setting outputStarted = true on reasoning-start and the exact content shape with a signature-only reasoning block.

   You may also want to check:

   • Session silently stops on empty LLM response (finish: unknown, 0 tokens) #41469: Session silently stops on empty LLM response
   • opencode run silently saves an empty assistant stop message (tokens counted, zero parts) and exits 0 #42677: opencode run silently saves an empty assistant stop message and exits 0
   • Empty stop completions are recorded as successful responses #37735: Empty stop completions are recorded as successful responses
4. cep207 commented on Oct 27, 2026

   

   cep207

   on Oct 27, 2026

   More actions

   Additional data point from a Desktop 2.1.4 (CLI core 1.19.4) user on Windows, provider opencode-go (glm-5.3-flash, mimo-v2.6-flash): we see an even more extreme variant — the stop step had content: [] entirely (not even a reasoning block), finish: "stop", tokens.output = 52–101, and the turn settled as idle / succeeded with zero visible output. From the server log, the failure window contains no WARN/ERROR at all — it is handled as a normal completion end-to-end. Stats: 3 such silent stops out of ~625 assistant messages over 2 days, all right after a tool result. Details and a JSON sample in #55378.
5. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   Thanks for the detailed report. I reproduced this with a simulated provider on @opencode/cli 2.2.0 (your dev build isn't published), and it still happens on the latest dev (35a4a68).

   Steps: the model's response to a prompt contains only an empty reasoning delta carrying an encrypted reasoning signature (OpenAI-chat reasoning_details), followed by a clean stop finish. The turn completes in about 100ms and the session goes idle with no text, no reasoning and no error, and there is no retry. The recording and screenshots below show this, and the script below reproduces it.

   In the code, reasoning-start marks output as started before any reasoning text arrives, and the step only treats a finish as invalid when the reason is unknown. So an empty stop step is accepted as success. I've passed this on to the maintainers.

   recording.mp4

   

   

   Reproduction script (drive.ts)

   // Repro for anomalyco/opencode#46880: a provider response containing only an
   // empty, signature-carrying reasoning block followed by a clean "stop" finish is
   // recorded as a successful step. The TUI ends idle with no answer and no error.
   import { Effect, Stream } from "effect"
   import { Llm, OpenCodeDriver } from "opencode-drive"
   
   export default OpenCodeDriver.use(
     {
       ...(process.env.OC_DEV ? { opencode: { dev: process.env.OC_DEV } } : {}),
       tui: { recording: true },
     },
     ({ ui, llm }) =>
       Effect.gen(function* () {
         yield* llm.serve((request) => {
           // Background title requests get a normal title.
           if (JSON.stringify(request).includes("Reply with only the title")) return Stream.make(Llm.text("Greeting"))
           // Main turn: only an empty reasoning delta carrying an encrypted signature, then finish=stop.
           return Stream.make(
             Llm.raw({
               choices: [
                 {
                   delta: {
                     reasoning_content: "",
                     reasoning_details: [{ type: "reasoning.encrypted", data: "SIGNATURE-ONLY-abc123", index: 0 }],
                   },
                 },
               ],
             }),
             Llm.finish("stop"),
           )
         })
         yield* ui.screenshot("before")
         yield* ui.submit("Say hello")
         yield* Effect.sleep(6000)
         // Session is idle: no assistant text, no reasoning, no error shown.
         yield* ui.screenshot("after-no-output")
       }),
   )

   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.
6. opencodejr

   added

   reproduced

   on Oct 28, 2026