opencode run (v2) exits 1 after a transient provider error that the runner retried and recovered from #50062

Activity

1. rajmehta53 commented on Oct 16, 2026

   

   rajmehta53

   on Oct 16, 2026

   More actions

   Still reproduces on v2.0.30, macOS arm64 (latest channel). Our variant is
   provider.invalid-output (OpenAI Chat stream ended without finish_reason, HTTP 200)
   rather than a transport error, but the outcome is identical.

   One automation run of:

   opencode run --dangerously-skip-permissions --agent build --model openrouter/z-ai/glm-5.3-flash "<prompt>"
   

   Timeline (local time):

   • 02:54:17 provider stream dies mid-run
   • 02:54:56 runner auto-retries, agent resumes and completes all work
   • 02:57:00 commit lands, tests pass, session finalizes outcome: "succeeded"
   • 02:57:02 CLI exits 1, stderr empty after the retry

   Checked the v2.0.30 tag of packages/cli/src/run/noninteractive.ts:
   session.step.failed still sets emittedError = true; process.exitCode = 1,
   session.execution.succeeded is still just return, and nothing in the file resets
   the exit code after a success.

   Downstream: our runner gates on exit code, so the successful attempt was recorded as a
   failure, and the next attempt found the work already committed, so retries could never
   produce a new commit. DB rows and server log available if useful.
   Watching #50068 and #52885.
2. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   Contributor

   More actions

   Thanks for the detailed report. I was able to reproduce this on v2.0.10, and it still happens on the latest dev (62ae125).

   Setup: a small OpenAI-compatible mock provider. On the agent's first step request it sends 200, the SSE headers and a couple of text deltas, then drops the socket. Every later request gets a normal answer ending with finish_reason: "stop". Running opencode run --format json -m mock/mock "Say hello" against it:

   • the stream shows the partial text, then an error event with provider.transport (ECONNRESET: The socket connection was closed unexpectedly, status 200);
   • the step is retried and returns the full answer; the session ends with finish stop and outcome succeeded;
   • stderr is empty and the exit code is 1.

   The cause matches what you describe. Core publishes session.step.failed for a step it then retries, and the run command's handler for that event sets process.exitCode = 1 without checking for a retry. A later success never resets it. The script below reproduces it, either against the installed opencode or against a source checkout with OC_DEV=<checkout>.

   Related: #39573 (the same symptom on the v1 runner, through a different code path).

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Repro for #50062: `opencode run` exits 1 after a step that failed mid-stream was retried and the run finished with "stop".
   # Usage: ./repro.sh            (uses `opencode` from PATH)
   #        OC_DEV=/path/to/opencode-checkout ./repro.sh   (run the CLI from source)
   set -u
   OC=${OC:-opencode}
   WORK=$(mktemp -d /tmp/opencode/repro50062.XXXX)
   PORT=${PORT:-18765}
   cat > "$WORK/mock.mjs" <<'JS'
   // OpenAI-compatible mock (node:http). Request #1: 200 + headers + two SSE deltas, then the socket is destroyed
   // (client sees ECONNRESET / "socket connection was closed unexpectedly"). Later requests: a normal answer, finish "stop".
   import http from "node:http"
   let n = 0 // agent step requests seen
   const chunk = (delta, finish = null) =>
     `data: ${JSON.stringify({ id: "c1", object: "chat.completion.chunk", created: 0, model: "mock", choices: [{ index: 0, delta, finish_reason: finish }] })}\n\n`
   http.createServer((req, res) => {
     console.error(`[mock] ${req.method} ${req.url}`)
     if (req.url.endsWith("/models")) { res.writeHead(200, { "content-type": "application/json" }); return res.end(JSON.stringify({ object: "list", data: [{ id: "mock" }] })) }
     let raw = ""
     req.on("data", (d) => (raw += d))
     req.on("end", () => {
       const body = JSON.parse(raw || "{}")
       // Only the agent's own step requests carry tools; title generation etc. do not.
       const agentStep = Array.isArray(body.tools) && body.tools.length > 0
       if (agentStep) n++
       console.error(`[mock] request agentStep=${agentStep} #${n}`)
       res.writeHead(200, { "content-type": "text/event-stream" })
       if (agentStep && n === 1) {
         res.write(chunk({ role: "assistant", content: "Partial " }))
         res.write(chunk({ content: "answ" }))
         setTimeout(() => res.socket.destroy(), 300)
         return
       }
       res.write(chunk({ role: "assistant", content: "Hello, " }))
       res.write(chunk({ content: "world." }))
       res.write(chunk({}, "stop"))
       res.write(`data: ${JSON.stringify({ id: "c1", object: "chat.completion.chunk", created: 0, model: "mock", choices: [], usage: { prompt_tokens: 10, completion_tokens: 3, total_tokens: 13 } })}\n\n`)
       res.end("data: [DONE]\n\n")
     })
   }).listen(Number(process.env.PORT), () => console.error("[mock] listening"))
   JS
   PORT=$PORT node "$WORK/mock.mjs" 2> "$WORK/mock.log" &
   MOCK=$!
   sleep 1
   mkdir -p "$WORK/project" && cd "$WORK/project" && git init -q
   cat > opencode.json <<JSON
   {
     "\$schema": "https://opencode.ai/config.json",
     "autoupdate": false,
     "provider": {
       "mock": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Mock",
         "options": { "baseURL": "http://127.0.0.1:$PORT/v1", "apiKey": "x" },
         "models": { "mock": { "name": "Mock" } }
       }
     }
   }
   JSON
   # From source (OC_DEV=/path/to/checkout): the spawned server child needs the Solid preload too.
   if [ -n "${OC_DEV:-}" ]; then
     OC="bun --conditions=browser --preload=$OC_DEV/packages/tui/node_modules/@opentui/solid/scripts/preload.js $OC_DEV/packages/cli/src/index.ts"
     printf 'preload = ["%s"]\n' "$OC_DEV/packages/tui/node_modules/@opentui/solid/scripts/preload.js" > bunfig.toml
   fi
   export XDG_DATA_HOME=$WORK/data XDG_CONFIG_HOME=$WORK/config XDG_STATE_HOME=$WORK/state XDG_CACHE_HOME=$WORK/cache
   $OC --version
   timeout 40 $OC run --format json -m mock/mock "Say hello" > "$WORK/out.jsonl" 2> "$WORK/err.txt"
   CODE=$?
   echo "--- JSON events (type / error / reason):"
   cat "$WORK/out.jsonl"
   echo "--- stderr:"; cat "$WORK/err.txt"
   echo "--- mock:"; cat "$WORK/mock.log"
   echo "EXIT CODE: $CODE"
   kill $MOCK 2>/dev/null
   # opencode run leaves a background service process; stop it so the next run starts fresh.
   pkill -f "[s]erve --service" 2>/dev/null || true

   Run it: bash repro.sh
3. opencodejr

   added

   reproduced

   on Oct 28, 2026