Bug type
Behavior bug (incorrect output/state without crash)
Beta release blocker
No
Summary
In Control UI (webchat), starting a new session via /new or /reset can produce a first user message that is incorrectly prefixed with stale system-summary text from a previous session.
Steps to reproduce
NOT_ENOUGH_INFO
Expected behavior
A session started via /new or /reset should begin with only the new-session startup prompt/input. It should not include unrelated content from a previous session.
Actual behavior
In one observed Control UI session, the first user message of the new session contained:
- A stale prefixed system-summary block starting with
System: [2026-03-31 15:35:05 GMT+8] 这次贴吧心跳我做了几件事。
- The expected startup text starting with
A new session was started via /new or /reset...
This caused previous-session content to be delivered to the model as part of the new session's first input.
OpenClaw version
2026.3.28 (f9b1079)
Operating system
macOS 26.4.0 (Darwin 25.4.0, arm64)
Install method
npm global
Model
openai-codex/gpt-5.4
Provider / routing chain
openclaw -> openai-codex/gpt-5.4
Additional provider/model setup details
Observed on Control UI / webchat.
Logs, screenshots, and evidence
Source session containing the stale summary text:
~/.openclaw/agents/main/sessions/e8d3bb06-2ea4-41b5-9b4d-63aabe645319.jsonl
Observed contaminated new session:
~/.openclaw/agents/main/sessions/7d991b7b-41a5-4678-a938-ec7abf383431.jsonl
Contaminated first user message timestamp:
2026-03-31T11:40:54.906Z
Observed first message content begins with:
System: [2026-03-31 15:35:05 GMT+8] 这次贴吧心跳我做了几件事。
...
A new session was started via /new or /reset. Run your Session Startup sequence ...
Relevant transcript fragment:
{"type":"message","timestamp":"2026-03-31T11:40:54.906Z","message":{"role":"user","content":[{"type":"text","text":"System: [2026-03-31 15:35:05 GMT+8] 这次贴吧心跳我做了几件事。\n...\nA new session was started via /new or /reset..."}]}}
I also checked recent startup-session transcripts and found this contamination in 1 observed case; other recent `/new` or `/reset` startup messages in the inspected sample did not contain prefixed stale content.
Impact and severity
Affected: Control UI / webchat sessions started via /new or /reset
Severity: Medium
Frequency: Intermittent; observed once in the inspected recent sample
Consequence: Previous-session system text is injected into the new session's first input, which can cause incorrect or confusing startup responses
Additional information
The stale text was already present in the inbound user message recorded in the new session transcript, so the contamination appears to happen before model processing.
Possibly related:
Bug type
Behavior bug (incorrect output/state without crash)
Beta release blocker
No
Summary
In Control UI (webchat), starting a new session via
/newor/resetcan produce a first user message that is incorrectly prefixed with stale system-summary text from a previous session.Steps to reproduce
NOT_ENOUGH_INFO
Expected behavior
A session started via
/newor/resetshould begin with only the new-session startup prompt/input. It should not include unrelated content from a previous session.Actual behavior
In one observed Control UI session, the first user message of the new session contained:
System: [2026-03-31 15:35:05 GMT+8] 这次贴吧心跳我做了几件事。A new session was started via /new or /reset...This caused previous-session content to be delivered to the model as part of the new session's first input.
OpenClaw version
2026.3.28 (f9b1079)
Operating system
macOS 26.4.0 (Darwin 25.4.0, arm64)
Install method
npm global
Model
openai-codex/gpt-5.4
Provider / routing chain
openclaw -> openai-codex/gpt-5.4
Additional provider/model setup details
Observed on Control UI / webchat.
Logs, screenshots, and evidence
Impact and severity
Affected: Control UI / webchat sessions started via
/newor/resetSeverity: Medium
Frequency: Intermittent; observed once in the inspected recent sample
Consequence: Previous-session system text is injected into the new session's first input, which can cause incorrect or confusing startup responses
Additional information
The stale text was already present in the inbound user message recorded in the new session transcript, so the contamination appears to happen before model processing.
Possibly related: