server: provider requests hang after host IP change until restart #53442

Activity

1. github-actions

   assigned

   neriousy

   on Oct 5, 2026
2. github-actions commented on Oct 5, 2026

   

   github-actionsBot

   on Oct 5, 2026 – with GitHub Actions

   Contributor

   More actions

   ---

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

   • #52564: Reports long-lived connections in opencode serve failing to recover after a network change until the server process is restarted. That report is MCP-specific, but the network-change and stale-connection behavior is similar.
3. opencodejr commented on Oct 27, 2026

   

   opencodejrBot

   on Oct 27, 2026

   More actions

   Thanks for the detailed report. I can reproduce this on v2.0.16 and on the latest dev (f7d0ce1), using a local mock provider instead of opencode.ai.

   What I did:

   • Put a mock OpenAI-compatible provider in a separate network namespace, reached through a veth pair.
   • Started opencode serve and sent one prompt. It works, and ss shows a pooled keep-alive socket from the original address.
   • Swapped the host-side address (ip addr del the old one, ip addr add a new one). curl to the provider from the new address still works.
   • Sent another prompt in the same session. It hangs.

   What I saw:

   • The request is written to the old socket, still bound to the old address, and piles up in its send queue.
   • No new connection is opened, the provider never receives the request, and the server logs no warning or error.
   • On a new session, the title request and the reply go out together. One gets the stale socket and hangs; the other opens a fresh connection and succeeds.

   v2.0.16 has no transport timeout, so the hang there should last indefinitely. Current dev has a default 5-minute response-header timeout, so the request probably fails and retries after about 5 minutes, but I didn't confirm that. Either way the stale socket isn't detected and the prompt hangs silently for minutes.

   The script below reproduces it (it needs root for the network namespace). Related issues: #53747 (steps stalling with no timeout) and #52866 (stall timeout and keepalive handling).

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduce #53442: provider requests hang after the host's source IP changes.
   # Needs root (network namespaces), bun, iproute2, curl, and `opencode` on PATH
   # (or OPENCODE_CMD="bun ... packages/cli/src/index.ts" for a source checkout).
   # A mock OpenAI-compatible provider runs in a separate netns behind a veth pair;
   # the host side address is swapped from .2 to .3 between two prompts.
   set -u
   NET=${NET:-10.9.1}; NS=${NS:-prov53442}; VH=${VH:-vh53442}; PORT=${PORT:-4299}
   OPENCODE_CMD=${OPENCODE_CMD:-opencode}
   WORK=$(mktemp -d /tmp/opencode/r53442.XXXX); mkdir -p "$WORK/proj/.opencode" "$WORK/home"
   export HOME="$WORK/home" OPENCODE_SERVER_PASSWORD=pw
   
   ip netns add "$NS"; ip link add "$VH" type veth peer name "${VH}p"
   ip link set "${VH}p" netns "$NS"
   ip addr add "$NET.2/24" dev "$VH"; ip link set "$VH" up
   ip -n "$NS" addr add "$NET.1/24" dev "${VH}p"; ip -n "$NS" link set "${VH}p" up
   
   cat > "$WORK/mock.ts" <<'EOF'
   Bun.serve({ hostname: "0.0.0.0", port: 8080, idleTimeout: 0, async fetch(req) {
     const u = new URL(req.url)
     if (!u.pathname.endsWith("/chat/completions")) return Response.json({ data: [] })
     const body: any = await req.json(); console.log(new Date().toISOString(), "POST", JSON.stringify(body.messages?.at(-1)?.content).slice(0, 60))
     const enc = new TextEncoder(), b = { id: "c1", object: "chat.completion.chunk", created: 1, model: "m" }
     const ev = (o: any) => enc.encode(`data: ${JSON.stringify(o)}\n\n`)
     return new Response(new ReadableStream({ start(c) {
       c.enqueue(ev({ ...b, choices: [{ index: 0, delta: { role: "assistant", content: "pong" }, finish_reason: null }] }))
       c.enqueue(ev({ ...b, choices: [{ index: 0, delta: {}, finish_reason: "stop" }], usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 } }))
       c.enqueue(enc.encode("data: [DONE]\n\n")); c.close()
     } }), { headers: { "content-type": "text/event-stream" } })
   } })
   EOF
   cat > "$WORK/proj/.opencode/opencode.jsonc" <<EOF
   { "model": "mock/m", "snapshots": false,
     "permissions": [{ "action": "*", "resource": "*", "effect": "allow" }],
     "providers": { "mock": { "name": "Mock", "package": "@opencode-ai/ai/providers/openai/chat",
       "settings": { "apiKey": "k", "baseURL": "http://$NET.1:8080/v1" },
       "models": { "m": { "name": "M", "capabilities": { "tools": false, "input": ["text"], "output": ["text"] },
         "limit": { "context": 128000, "output": 1000 } } } } } }
   EOF
   ip netns exec "$NS" bun "$WORK/mock.ts" > "$WORK/mock.log" 2>&1 &
   cd "$WORK/proj" && git init -q
   $OPENCODE_CMD serve --hostname 127.0.0.1 --port "$PORT" --print-logs --log-level info > "$WORK/serve.log" 2>&1 &
   sleep 8
   run() { local s=$(date +%s); timeout "${WAIT:-90}" $OPENCODE_CMD run --server "http://127.0.0.1:$PORT" -m mock/m "$@" > "$WORK/run.log" 2>&1; echo "  exit=$? after $(( $(date +%s)-s ))s, output: $(tail -1 "$WORK/run.log")"; }
   
   echo "== prompt 1 (source $NET.2)"; run "hi"
   ss -tn | grep "$NET.1:8080"
   echo "== host address $NET.2 -> $NET.3"
   ip addr del "$NET.2/24" dev "$VH"; ip addr add "$NET.3/24" dev "$VH"
   echo "== curl from the new address works:"; curl -s -m3 "http://$NET.1:8080/v1/models"; echo
   echo "== prompt 2 (same session)"; run -c "after ip change"
   echo "== sockets:"; ss -tno | grep "$NET.1:8080"
   echo "== requests the mock provider received:"; cat "$WORK/mock.log"
   echo "== WARN/ERROR in server log:"; grep -E "level=(WARN|ERROR)" "$WORK/serve.log" || echo "  (none)"
   kill %1 %2 2>/dev/null; ip netns del "$NS"; ip link del "$VH" 2>/dev/null

   Run it: bash repro.sh
4. opencodejr

   added

   reproduced

   on Oct 27, 2026