Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

We observed two failures in an OpenClaw Telegram direct session using the OpenAI Codex runtime. At first sight they looked like aggressive/unsafe compaction causing context loss, but the logs point to two different failure modes:

1. A Codex app-server failure before any assistant reply.
2. A spurious visible "Something went wrong while processing your request. Please try again." message injected after a successful Codex run.

This happened on OpenClaw `2026.5.22 (a374c3a)`.

Steps to reproduce

Environment

• OpenClaw: `2026.5.22 (a374c3a)`
• Surface: Telegram direct
• Runtime/provider: OpenAI Codex
• Model: `openai-codex/gpt-5.5`
• Context configuration:
  • `contextTokens`: `230000`
  • `contextPruning.mode`: `cache-ttl`
  • `compaction.mode`: `default`
  • `reserveTokensFloor`: `32000`
  • `midTurnPrecheck.enabled`: `true`
  • `memoryFlush.enabled`: `true`

Incident 1 - Codex app-server closed before reply

Local time: `2026-05-26 22:44:27 Europe/Paris`

Session:

• `3f9fde1f-8027-48e1-9b51-f827518d827c`

User prompt before the failure:
```
C'est laborieux la synchro iCloud
```

Trajectory shows:
```
session.ended status="error"
promptError="codex app-server client is closed"
```

Gateway log around the same timestamp shows:
```
Codex agent harness failed; not falling back to embedded PI backend
error="codex app-server client is closed"
lane task error: lane=main ... error="Error: codex app-server client is closed"
lane task error: lane=session: ... error="Error: codex app-server client is closed"
Embedded agent failed before reply: codex app-server client is closed
```

There was no assistant reply for that turn. The user then used `/new` at about `22:50:35`, and OpenClaw saved the session context.

Incident 2 - visible error after successful Codex run

Local time: `2026-05-26 23:53:09 Europe/Paris`

Session:

• `c72748c1-2cc5-42b0-8760-473a03e85f49`

The Codex rollout completed successfully at about `23:52:15` local time:
```
task_complete
last_token_usage.total_tokens=209473
model_context_window=258400
```

The assistant reply was successfully produced.

Then, without a new user message and without a transcript compaction event at that moment, OpenClaw appended a visible assistant message at `23:53:09`:
```
Something went wrong while processing your request. Please try again.
```

That message appears as:
```
provider/model: openclaw / delivery-mirror
usage: 0
```

The user then used `/new` at about `23:54:59`.

Expected behavior

• If the Codex app-server closes, OpenClaw should either restart/retry cleanly or surface a precise runtime error without making the user suspect context loss.
• A generic visible "Something went wrong" message should not be injected after a successful assistant turn.
• Large Telegram sessions should not require manual `/new` recovery when the model run itself succeeded.

Actual behavior

• At `22:44`, the Codex app-server closed and the Telegram session failed before reply.
• At `23:53`, Codex succeeded, but OpenClaw still injected a generic visible error from `delivery-mirror`.
• From the user side, both failures looked like unstable context/compaction behavior.

OpenClaw version

2026.5.22

Operating system

macOS26.4

Install method

npm global

Model

gpt5.5

Provider / routing chain

openclaw - OpenAI Codex

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response