openclaw - 💡(How to fix) Fix [Bug]: Telegram final assistant replies after tool calls are sometimes delivered but not persisted to session JSONL [3 comments, 3 participants]

Bug type

Regression / intermittent runtime bug

Beta release blocker

No

Summary

Telegram-delivered final assistant replies after tool-calling turns are sometimes absent from the active session JSONL, so the next model turn cannot see the assistant's previous answer.

Steps to reproduce

1. Use a Telegram direct-message agent session.
2. Send a request that causes the agent to run one or more tools before answering.
3. Wait for the final assistant answer to be delivered in Telegram.
4. Inspect the active session transcript at ~/.openclaw/agents/<agentId>/sessions/<sessionId>.jsonl.
5. In affected sessions, observe a toolResult -> user sequence where the final assistant text that was visible in Telegram is missing from the JSONL transcript.

Expected behavior

If a final assistant reply is delivered to Telegram, that same assistant reply should be present in the session transcript used for future context.

Expected transcript shape:

user
assistant toolCall
toolResult
assistant final text reply
user follow-up

Actual behavior

In affected sessions, the user receives the assistant reply in Telegram, but the session JSONL has this shape:

user
assistant toolCall
toolResult
user follow-up

The assistant final text is missing from the transcript, so the next model turn cannot see its own previous answer. This breaks natural follow-ups like "add these", "do it", "use the previous answer", or "apply option 2".

OpenClaw version

NOT_ENOUGH_INFO

Operating system

Docker/Coolify deployment on Linux VPS. Exact host OS version not captured.

Install method

docker / Coolify

Model

Observed with more than one model path, including GPT-5.5 via OpenAI Codex Responses and Claude Sonnet 4.6 via OpenRouter.

Provider / routing chain

Telegram channel -> OpenClaw gateway -> agent runtime -> model provider. Observed on both openai-codex-responses and openai-completions style model paths.

Additional provider/model setup details

No provider-specific API keys, tokens, or account identifiers included. The issue appears tied to Telegram session transcript persistence after tool-call turns, not one specific model.

Logs, screenshots, and evidence

Observed by comparing session JSONL files for multiple agents/sessions.

Affected transcript examples had gaps like:

L160 assistant toolCall
L161 toolResult
L162 assistant toolCall
L163 toolResult
# assistant final text was delivered to Telegram here, but no assistant text entry appears in JSONL
L164 user follow-up

Another affected session showed multiple toolResult -> user gaps:

toolResult -> user
toolResult -> user
toolResult -> user

Control sessions did persist assistant text normally, which suggests the issue is intermittent rather than universal.

Observed sample:

Affected session A: 6 toolResult -> user gaps
Affected session B: 5 toolResult -> user gaps
Control session C: 0 gaps
Control session D: 0 gaps

The issue is not explained only by large tool output size. Gaps were observed after both small and large tool results, including very small tool results.

Impact and severity

Affected systems/channels: Telegram agent sessions.

Severity: Medium to High. The assistant reply is visible to the user, but absent from model history on the next turn. This causes context loss and failed follow-up instructions when the user refers to the previous assistant answer.

Frequency: Intermittent. It occurred in some sampled Telegram sessions and not in others.

Consequence: The model may ask the user to restate information that it had just provided, or may be unable to act on shorthand follow-ups such as "add these" because its own previous answer is missing from the transcript.

Additional information

No personal identifiers or private message content are included here. The report is based on transcript structure only: observed session JSONL event sequences and whether assistant text entries were present between tool results and subsequent user turns.