opencode2 run and background: true tasks may cause premature exit #48361

Activity

1. github-actions

   assigned

   kitlangton

   on Sep 10, 2026
2. github-actions commented on Sep 10, 2026

   

   github-actionsBot

   on Sep 10, 2026 – with GitHub Actions

   Contributor

   More actions

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

   • opencode run (headless) exits 0 mid-pipeline after a parallel subagent fan-out, before the model's next turn #37548: opencode run exits 0 mid-pipeline after a parallel subagent fan-out — same symptom of premature exit after background tasks are dispatched
   • Bug: opencode run exits before background tasks (explore/librarian) complete #16380: opencode run exits before background tasks complete — same root cause where headless mode treats a model response as session completion without waiting for pending background work

   If your issue is distinct (e.g., specific to the background: true flag in task config rather than the general headless lifecycle), please add details clarifying the difference.
3. opencode-agent

   added

   2.0

   on Sep 10, 2026
4. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   Thanks for the report. I could reproduce this.

   I used a mock model whose first reply starts two subagent calls with background: true. Each subagent takes about 15 seconds to answer. Then I ran opencode2 run --standalone --auto "..." on v0.0.0-beta-19425. What happened:

   • Both subagents were shown as ✓ immediately.
   • The parent printed its "waiting for them to finish" text.
   • run exited with code 0 after about 1–2 seconds.

   The subagents went on to finish about 15 seconds later, but the parent was never resumed with their results, so the final answer never appeared. It still happens on the latest dev (1e7ba46f8b99). The script below reproduces it with a built-in mock provider.

   Closely related:

   • One-shot run exits before plugin-backgrounded work completes; plugins cannot detect a headless client on the shared service #48683: the same early exit, when work is moved to the background by a plugin
   • [BUG] V2: subagent with pending background work (shell commands or subagents, background: true) marked completed early — results never collected #48826: a subagent with pending background work is reported as completed too early

   Until this is addressed, you could ask the model in your script not to use background: true. That gives up the parallelism you were after, but foreground subagents block until they finish, so run gets their results.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Repro for #48361: `opencode run` exits before background subagents finish.
   # Needs: bun and `opencode` on PATH (override with OC="..."). Starts an embedded OpenAI-compatible mock on :4999.
   # The mock makes the parent call `subagent` twice with background:true; each child takes 15s to answer.
   set -u
   OC=${OC:-opencode}
   # OC_DEV=/path/to/checkout runs the CLI from source instead
   if [ -n "${OC_DEV:-}" ]; then
     OC="bun --conditions=browser --preload=$OC_DEV/packages/tui/node_modules/@opentui/solid/scripts/preload.js $OC_DEV/packages/cli/src/index.ts"
   fi
   DIR=$(mktemp -d); mkdir -p "$DIR/home" "$DIR/proj"
   export HOME=$DIR/home XDG_CONFIG_HOME=$DIR/home/.config XDG_DATA_HOME=$DIR/home/.local/share XDG_STATE_HOME=$DIR/home/.local/state XDG_CACHE_HOME=$DIR/home/.cache
   cat > "$DIR/proj/opencode.json" <<'JSON'
   { "model": "mock/mock", "permission": { "*": "allow" },
     "provider": { "mock": { "npm": "@ai-sdk/openai-compatible", "name": "Mock",
       "options": { "baseURL": "http://127.0.0.1:4999/v1", "apiKey": "x" },
       "models": { "mock": { "name": "Mock", "tool_call": true, "limit": { "context": 100000, "output": 4000 } } } } } }
   JSON
   cat > "$DIR/mock.ts" <<'MOCK'
   // Minimal OpenAI-compatible streaming mock.
   const log = (...a: unknown[]) => console.error(new Date().toISOString(), ...a)
   const text = (m: any) => typeof m.content === "string" ? m.content : (m.content ?? []).map((p: any) => p.text ?? "").join("")
   function sse(chunks: any[]) {
     const enc = new TextEncoder()
     return new Response(new ReadableStream({ start(c) {
       for (const ch of chunks) c.enqueue(enc.encode(`data: ${JSON.stringify(ch)}\n\n`))
       c.enqueue(enc.encode("data: [DONE]\n\n")); c.close()
     } }), { headers: { "content-type": "text/event-stream" } })
   }
   const base = { id: "x", object: "chat.completion.chunk", created: 0, model: "mock" }
   const say = (s: string) => sse([{ ...base, choices: [{ index: 0, delta: { role: "assistant", content: s } }] }, { ...base, choices: [{ index: 0, delta: {}, finish_reason: "stop" }] }])
   const calls = (cs: any[]) => sse([{ ...base, choices: [{ index: 0, delta: { role: "assistant", tool_calls: cs.map((c, i) => ({ index: i, id: c.id, type: "function", function: { name: c.name, arguments: JSON.stringify(c.args) } })) } }] }, { ...base, choices: [{ index: 0, delta: {}, finish_reason: "tool_calls" }] }])
   Bun.serve({ port: Number(process.env.PORT ?? 4999), idleTimeout: 120, async fetch(req) {
     const url = new URL(req.url)
     if (!url.pathname.endsWith("/chat/completions")) return new Response("{}")
     const body: any = await req.json()
     const msgs = body.messages as any[]
     const all = msgs.map(text).join("\n")
     const users = msgs.filter((m) => m.role === "user").map(text)
     const tools = (body.tools ?? []).map((t: any) => t.function.name)
     if ((users[0] ?? "").includes("CHILD_TASK")) {
       const tok = users[0].match(/CHILD_TASK_(\w+)/)![1]
       log("child", tok, "-> sleeping 15s"); await Bun.sleep(15000); log("child", tok, "done")
       return say(`TOKEN_${tok}`)
     }
     if (!tools.length) { log("title/other"); return say("Title") }
     if (all.includes("PARENT")) {
       const toolMsgs = msgs.filter((m) => m.role === "tool").length
       if (/TOKEN_A/.test(all) && /TOKEN_B/.test(all)) { log("parent final"); return say("FINAL: got TOKEN_A and TOKEN_B") }
       if (toolMsgs === 0) { log("parent -> launching 2 background subagents; tools:", tools.join(",")); return calls([
         { id: "call_a", name: "subagent", args: { agent: "general", description: "child a", prompt: "CHILD_TASK_A", background: true } },
         { id: "call_b", name: "subagent", args: { agent: "general", description: "child b", prompt: "CHILD_TASK_B", background: true } },
       ]) }
       log("parent -> waiting text"); return say("Launched both subagents; waiting for them to finish.")
     }
     log("unknown"); return say("ok")
   } })
   log("mock listening")
   MOCK
   PORT=4999 bun "$DIR/mock.ts" 2>"$DIR/mock.log" & MOCK_PID=$!; sleep 1
   # the --standalone child server is spawned without the preload flag; give it a bunfig.toml
   [ -n "${OC_DEV:-}" ] && echo "preload = [\"$OC_DEV/packages/tui/node_modules/@opentui/solid/scripts/preload.js\"]" > "$DIR/proj/bunfig.toml"
   cd "$DIR/proj"
   start=$(date +%s)
   timeout 120 $OC run --standalone --auto "PARENT: launch two background subagents and report their tokens"
   echo "exit=$? wall=$(( $(date +%s)-start ))s  (expected: waits ~15s and prints FINAL: got TOKEN_A and TOKEN_B)"
   sleep 20; echo "--- mock log (children keep running after run exited; parent never resumed) ---"; cat "$DIR/mock.log"; kill $MOCK_PID

   Run it: bash repro.sh
5. opencodejr

   added

   reproduced

   on Oct 28, 2026
6. rekram1-node

   closed this as completedin #60699on Nov 4, 2026