#OpenClaw bug report — Telegram polling + model invocation issues on WSL2/Docker

1 messages · Page 1 of 1 (latest)

minor elbow
#

Setup: WSL2 + Docker on Windows 10. Tested v2026.4.25 and v2026.4.26. Coming from v2026.4.14 (stable for weeks).

Already applied (per https://docs.openclaw.ai/channels/telegram):

OPENCLAW_TELEGRAM_DISABLE_AUTO_SELECT_FAMILY=1
OPENCLAW_TELEGRAM_DNS_RESULT_ORDER=ipv4first
NODE_OPTIONS=--dns-result-order=ipv4first

fetch('https://api.telegram.org/') from inside the container → 200 OK 438ms. Basic egress is fine.

4.25 — Polling stalls every 15-25 min (getUpdates stuck for 175-334s; forcing restart). Watchdog restart works, but immediately afterwards every send fails: sendChatAction failed / sendMessage failed: Network request for ... failed!. ~25-50% message loss.

4.26 + gpt-5.4 — New error from [agent/embedded]:

embedded run agent end: isError=true model=gpt-5.4 provider=openai-codex
error=One of "input" or "previous_response_id" or 'prompt' or 'conversation_id' must be provided.

Same send cascade afterwards. Plus Polling runner stop timed out after 15s; forcing restart cycle. Stuck sessions are on agent:main:telegram:slash:<CHAT_ID>.

4.26 + gpt-5.5 (via openclaw.json hot reload; agent model: openai-codex/gpt-5.5 confirmed) — After [gateway] ready the logs go silent. No queue, no stuck session, no polling stall, no send. "typing" appears once on a fresh session, then nothing. Worse than 4.25 — no diagnostic to act on.

Questions:

  • Is Telegram polling under WSL2/Docker on Windows known broken in 4.25/4.26?
  • Is openai-codex/gpt-5.4 deprecated in 4.26, or is the One of "input"... error a regression?
  • Why does 4.26 + gpt-5.5 produce no log activity at all after ready? Any verbosity flag I'm missing?

Sanitized logs attached.

ornate groveBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

drowsy basinBOT
#

I checked current docs/source/changelog. This looks like a real 4.25/4.26 regression cluster, not “you misconfigured Telegram.”

Short answers:

  1. WSL2/Docker isn’t documented as globally broken, but 4.25/4.26 had very relevant Telegram/network fixes immediately after. I’d test 2026.4.27+ first before deeper config changes.

  2. openai-codex/gpt-5.4 doesn’t look deprecated from current source. That One of "input"... error means OpenAI/Codex received an empty/malformed Responses request. If it happens on normal messages, treat it as a regression, not deprecation.

  3. 4.26 + gpt-5.5 “ready then silence” may match a startup/prewarm issue fixed in 2026.4.27: changelog says channels now start without waiting for primary model prewarm. That could explain “gateway ready, but channel/runtime activity disappears.”

For better logs, note that --verbose only affects console verbosity; file logs need config:

{
  logging: {
    level: "debug",
# 
    consoleLevel: "debug"
  },
  diagnostics: {
    flags: ["telegram.http"]
  }
}

Then restart the container fully and capture:

openclaw --version
openclaw status --all
openclaw channels status --probe
openclaw logs --follow --local-time