openclaw - 💡(How to fix) Fix Codex app-server: long agent replies silently truncated at ~1000-1100 chars (stop=null, aborted=false)

Summary

Configured Codex/OAuth agent (gpt-5.5) replies are silently truncated at ~1000–1100 chars when invoked headlessly via openclaw message. The model is not aborted: data.aborted: false, stopReason/finishReason/errorMessage all null, yet assistantTexts[0] ends mid-sentence and the CLI wrapper reports status=timeout.

This is not the webchat DEFAULT_CHUNK_LIMIT = 4000 truncation — the number (1100 vs 4000) and the codepath (codex app-server stdio vs webchat outbound) are different.

Environment

• OpenClaw: 2026.5.18 stable
• Node v22.22.2, Linux 6.8.0 x86_64
• Agent: Codex on OpenAI OAuth (gpt-5.5)
• Channel: direct CLI (openclaw message), not Telegram/Discord

Symptom — last 10 invocations

Replies above ~1000 chars cut mid-sentence; replies below complete cleanly.

Reproduction

1. Configure a Codex/OAuth gpt-5.5 agent.
2. Ask for a structured 1700-char reply (five labeled sections, 3–5 sentences each, "do not stop short").
3. Inspect the resulting *.trajectory.jsonl:

import json
for line in open("<session>.trajectory.jsonl"):
    e = json.loads(line)
    if e.get("type") == "model.completed":
        d = e["data"]
        print(d["aborted"], d.get("stopReason"), len(d["assistantTexts"][0]))

Length lands 800–1100 chars; sentence is incomplete; aborted=false, stopReason=None.

Tested on 2026.5.19-beta.2 — worse, not better

Same prompt, same install, gateway restarted:

• 5.18 stable: truncated at 810 chars
• 5.19-beta.2: truncated at 455 chars

Rolled back to 5.18.

What this is NOT

• Not a real timeout (no errorMessage, no codex 5xx).
• Not DEFAULT_CHUNK_LIMIT (different number, different codepath).
• Not auth (PR #73588 is in 5.18; the agent responds, just truncated).
• Not quota (short prompts complete cleanly back-to-back).

Asks

1. Identify the buffer/window capping the codex app-server stdio path at ~1100 chars.
2. Make it configurable, or unbounded.
3. Surface a real stopReason ("buffer_full" / "stream_cutoff") when this happens — currently null makes the internally-inconsistent state hard to detect from the outside.

Full trajectory .jsonl files and additional repros available on request.