Bug: prompt loop ID comparison still broken in v1.3.0 when running `opencode web` remotely (with workaround)

[chnlkw]

Description

The prompt loop exit condition in SessionPrompt.loop() still uses message ID lexicographic comparison in v1.3.0:

if (lastAssistant?.finish
    && !["tool-calls", "unknown"].includes(lastAssistant.finish)
    && lastUser.id < lastAssistant.id) {
    break;
}

This breaks when opencode web runs on a remote server (e.g. Kubernetes) and the browser is on a different machine. The user message ID is generated in the browser via Date.now(), while the assistant message ID is generated on the server. Any clock skew (even a few ms where the browser is ahead) causes lastUser.id > lastAssistant.id, so the loop never exits after finish=stop.

Symptoms:

• claude-opus-4-6 (no prefill support): "This model does not support assistant message prefill" error on every turn
• Models with prefill support (e.g. claude-opus-4-5): step 1 silently overwrites step 0 → blank/stuck session in Web UI

Confirmed by extracting from the v1.3.0 binary:

$ strings /usr/local/bin/opencode | grep 'lastUser.id < lastAssistant.id'
if (lastAssistant?.finish && !["tool-calls", "unknown"].includes(lastAssistant.finish) && lastUser.id < lastAssistant.id) {

This is the same root cause as #13768 and #17982. Community PRs #17010 and #17149 fix both backend and frontend sides but remain open and unreviewed — could these be considered for merge?

Workaround

For anyone running opencode web remotely, a minimal Node.js proxy that delays only prompt requests by the exact clock skew (all other traffic zero-delay passthrough):

"use strict";
const http = require("http");
const TARGET = { host: "127.0.0.1", port: parseInt(process.env.TARGET_PORT || "3000") };
const LISTEN_PORT = parseInt(process.env.LISTEN_PORT || "8080");
const PROMPT_RE = /\/session\/[^/]+\/prompt(?:_async)?$/;
const MASK48 = (1n << 48n) - 1n;

function calcDelay(messageID) {
  if (!messageID?.startsWith("msg_") || messageID.length < 16) return 0;
  const hex = messageID.slice(4, 16);
  if (!/^[0-9a-f]{12}$/i.test(hex)) return 0;
  try {
    const idEnc = BigInt("0x" + hex);
    const nowEnc = (BigInt(Date.now()) * 0x1000n + 1n) & MASK48;
    return idEnc > nowEnc ? Math.min(Number((idEnc - nowEnc) / 0x1000n), 5000) : 0;
  } catch { return 0; }
}

function proxy(req, res, body) {
  const p = http.request({ ...TARGET, method: req.method, path: req.url,
    headers: { ...req.headers, host: `${TARGET.host}:${TARGET.port}` }
  }, (r) => { res.writeHead(r.statusCode, r.headers); r.pipe(res); });
  p.on("error", () => { if (!res.headersSent) { res.writeHead(502); res.end("Bad Gateway"); } });
  body ? p.end(body) : req.pipe(p);
}

http.createServer((req, res) => {
  if (req.method === "POST" && PROMPT_RE.test(req.url)) {
    const chunks = [];
    req.on("data", (c) => chunks.push(c));
    req.on("end", () => {
      const body = Buffer.concat(chunks);
      let delay = 0;
      try { const j = JSON.parse(body); if (j.messageID) delay = calcDelay(j.messageID); } catch {}
      if (delay > 0) setTimeout(() => proxy(req, res, body), delay);
      else proxy(req, res, body);
    });
  } else proxy(req, res);
}).on("upgrade", (req, sock, head) => {
  const p = http.request({ ...TARGET, method: req.method, path: req.url, headers: req.headers });
  p.on("upgrade", (r, up, uh) => {
    let h = "HTTP/1.1 101 Switching Protocols\r\n";
    for (let i = 0; i < r.rawHeaders.length; i += 2) h += r.rawHeaders[i] + ": " + r.rawHeaders[i+1] + "\r\n";
    sock.write(h + "\r\n"); if (uh.length) sock.write(uh); if (head.length) up.write(head);
    sock.pipe(up); up.pipe(sock);
    sock.on("error", () => up.destroy()); up.on("error", () => sock.destroy());
  });
  p.on("error", () => sock.destroy());
  p.end();
}).listen(LISTEN_PORT, "0.0.0.0", () =>
  console.log("[proxy] :" + LISTEN_PORT + " -> " + TARGET.host + ":" + TARGET.port)
);

Usage: TARGET_PORT=3000 LISTEN_PORT=8080 node proxy.js, point browser at port 8080. The proxy extracts the 48-bit timestamp from messageID, compares with server Date.now(), and delays only when the browser clock is ahead.

Plugins

none

OpenCode version

1.3.0

Steps to reproduce

1. Deploy opencode web on a remote server (e.g. Kubernetes container)
2. Open the Web UI from a browser on a different machine
3. Select any Anthropic model (e.g. claude-opus-4-6)
4. Send any message
5. Observe: step 0 completes with finish=stop, but the loop continues to step 1 which fails with prefill error (or produces a blank response for prefill-capable models)

The issue does NOT reproduce when browser and server share the same machine (local opencode web), because they share the same hardware clock.

Screenshot and/or share link

N/A — server-side bug, visible in --log-level INFO logs as step=1 firing after finish=stop on step 0

Operating System

Linux (Kubernetes, Debian bookworm container)

Terminal

N/A (Web UI)

[github-actions]

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

• Bug: OpenCode prompt loop continues after finish=stop, triggering prefill error on claude-opus-4-6 #17982: Same root cause — prompt loop continues after finish=stop, triggering the prefill error on claude-opus-4-6. That issue covers the same lastUser.id < lastAssistant.id comparison bug and links to the same fix PRs (fix(session): use parentID instead of message ID ordering in prompt loop #17010, fix(web): order session messages by creation time #17149).
• This model does not support assistant message prefill / Github Copilot with Opus 4.6 #13768: Same prefill error symptom (This model does not support assistant message prefill) from the same underlying loop bug.

Your report adds useful new context (remote deployment / clock skew as a concrete trigger for the ID ordering race, plus a practical workaround proxy), but the core defect and the proposed fixes are already tracked in #17982. It may be worth adding your clock-skew reproduction details as a comment there instead.

[chnlkw]

Moved the v1.3.0 confirmation, reproduction details, and workaround to #17982 as suggested — see #17982 (comment)

Closing this as duplicate.