Assistant turn that ends with reasoning only (no text, no tool call, finish "stop") is settled silently and the agent stops #55378

Activity

1. opencode-agent commented on Oct 14, 2026

   

   opencode-agentBot

   on Oct 14, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the clear report. I was able to reproduce this on 2.0.29 with a simulated model, and it still happens on the latest dev (34d8ded0c).

   What I did:

   1. The model makes a shell tool call and gets its output back.
   2. The next step returns a single chunk with content: null, no tool calls, finish_reason: "stop" and usage of 24 tokens, all reasoning.

   The TUI shows the tool output and then nothing: no answer, no error, and the session goes idle as if the turn had ended normally. The token counter still includes that step's usage. The recording and screenshot below show it, and the script below reproduces it.

   It looks like the session runner only treats an empty step as a failure when the finish reason is unknown. An empty stop is recorded as a normal completion.

   Related issues:

   • core: signature-only responses complete successfully with no visible output #46880: a step with only empty, signed reasoning also ends as a silent success
   • v2: empty reasoning-only response is recorded as successful completion #37372: the same problem reported earlier; its fix (fix(core): fail empty provider output #37379) was never merged
   • Empty stop completions are recorded as successful responses #37735 / fix(session): retry an empty completion regardless of finish reason #49061: the empty stop case on v1, with a proposed retry fix
   • Empty response with finish_reason "length" and zero tokens is treated as success #50949 / fix(core): fail length finishes that return no content #51741: the same problem with an empty length finish

   recording.mp4

   

   

   Reproduction script (drive.ts)

   // Repro for #55378: after a tool result, the model ends a step with no text and no
   // tool call (finish_reason "stop", only reasoning usage). OpenCode settles it silently.
   import { Effect } from "effect"
   import { Llm, OpenCodeDriver } from "opencode-drive"
   
   export default OpenCodeDriver.use(
     {
       ...(process.env.OC_DEV ? { opencode: { dev: process.env.OC_DEV } } : {}),
       project: { git: true, files: { "README.md": "hello\n" } },
       tools: ["shell"],
       tui: { recording: true },
     },
     ({ ui, llm, tools }) =>
       Effect.gen(function* () {
         const shells = yield* tools.control("shell")
         // Step 1: a normal tool call.
         yield* llm.queue(
           Llm.toolCall({ index: 0, id: "call_ls", name: "shell", input: { command: "ls" } }),
           Llm.finish("tool-calls"),
         )
         // Step 2: the provider returns no content and no tool call, only reasoning usage, then "stop".
         yield* llm.queue(
           Llm.raw({
             id: "chatcmpl-empty",
             object: "chat.completion.chunk",
             created: 0,
             model: "gemini-3.5-flash-lite",
             choices: [{ index: 0, delta: { role: "assistant", content: null }, finish_reason: "stop" }],
             usage: {
               prompt_tokens: 100,
               completion_tokens: 24,
               total_tokens: 124,
               completion_tokens_details: { reasoning_tokens: 24 },
             },
           }),
         )
         yield* ui.submit("List the files, then tell me what you found")
         const shell = yield* shells.take("call_ls")
         yield* shell.succeed({ output: "README.md\n", exit: 0 })
         yield* Effect.sleep(4000)
         yield* ui.screenshot("after-empty-stop")
       }),
   )

   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 14, 2026
3. zonemeen commented on Oct 14, 2026

   

   zonemeen

   on Oct 14, 2026

   More actions

   I’d like to take this on. My plan is to detect stop responses with no text or tool calls, retry with a limit, and surface a clear error if retries fail. I’ll also add a regression test for reasoning-only responses after a tool result.

   👍React with 👍1Reacted by Juan Carlos Moreno
4. yxlimo commented on Oct 17, 2026

   

   yxlimo

   on Oct 17, 2026

   More actions

   Reproducing the same symptom on OpenCode 2.0.32 through the free-tier channel (opencode.ai/inference/...), across three models — so this is not Gemini-specific: mimo-v2.6-flash-free (×18), deepseek-v4.1-flash-free (×1), glm-5.3-flash-free (×1) within ~48h of multi-session agentic usage.

   What we can add beyond the hypotheses discussed so far is a raw stream capture taken at a passive MITM tap for one occurrence (glm-5.3-flash-free): the upstream sent a complete ~800 B stream —

   data: {"choices":[{"index":0,"delta":{"content":""}}]}
   data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
   data: usage chunk (completion_tokens=7)
   data: [DONE]

   No tool-call deltas at any point. The session DB recorded exactly what the stream contained (finish:"stop", no tool parts). 13 consecutive control streams on the same capture were healthy with normal finish_reason:"tool_calls" endings — so (a) the client settles the turn precisely as received and nothing is dropped client-side, and (b) the tap itself does not perturb the stream.

   This shape matches the captured stream in this issue exactly (empty/reasoning-only content, finish_reason:"stop", no text, no tool calls). Note it differs from the two shapes in #44148 (finish_reason:"other" and clean-FIN cutoff, Go plan): so at least three distinct upstream failure shapes currently end in the same silent stop.

   Related context, also from below: #49673 was never reopened despite still reproducing (the free-tier hits above were on 2.0.x), and PR #40531 (retry empty responses) remains unmerged/conflicted — so the client-side fix path is still open while occurrence continues.

   One possibly-relevant lead, with the caveat that causality is unproven: across 17 occurrences we recorded, 7 had a Magic-Context historian sub-session request rejected with HTTP 403 ("free tier can only be used from within OpenCode") 1.0–1.7 s around the hit (enrichment ~150× vs baseline assistant messages). If the upstream throttles bursts of parallel free-tier-authenticated requests, that could correlate with the empty-stop shape. Happy to share sanitized capture scripts and evidence on request.
5. thymikee commented on Oct 17, 2026

   

   thymikee

   on Oct 17, 2026

   More actions

   Also seen on 2.0.32 with callstack/Apex (OpenAI-compatible, served by vLLM through LiteLLM). On the gateway the stopping requests are logged as success with a normal stop. Two examples, from about 13.5k agentic steps:

   • Reasoning only: 13,020 reasoning tokens, 3 output tokens, no text, no tool call. The session went idle and the task was reported complete.
   • 706 reasoning tokens plus 53 non-reasoning output tokens, yet no text part and no tool part were stored. Whatever those 53 tokens were never reached the session. That may point to a dropped tool call rather than a model that simply stopped.

   We see a close variant more often: the step ends with text that announces an action ("…Rewriting:", "Let me run the test…"), then stop with no tool call. It's harder to detect, but the same "retry or surface it" handling would help there too.
6. luisboch commented on Oct 20, 2026

   

   luisboch

   on Oct 20, 2026

   More actions

   Seeing this on 2.1.1 too, and there it looks like a regression: it happens on almost every step after a tool result, not occasionally.

   Environment

   • opencode desktop 2.1.1 (Linux x64, .deb, channel latest)
   • Provider/model: github-copilot / gemini-3.8-flash (variant medium)
   • Custom primary agent, Code Mode enabled with remote MCP servers

   What happens

   The model calls read, the tool result comes back fine, and the next step ends with nothing in it:

   { 
     "content": [],
     "finish": "stop", 
     "rawFinish": "stop",
     "tokens": { 
         "input": 6434,
         "output": 0,
         "reasoning": 0,
         "cache": {
              "read": 15815 
         }
      } 
   }