openclaw - 💡(How to fix) Fix Session reset: /new prompt drops sender envelope, blocking personalized greetings [1 participants]

Summary

On a bare /new (or /reset) in a channel like Telegram, the agent's first turn receives the BARE_SESSION_RESET_PROMPT_BASE string without the Sender (untrusted metadata) envelope block. This means the LLM has no way to see who just started the session — even though the channel adapter already populated SenderName / SenderId / SenderUsername in sessionCtx.

Net effect: agents cannot personalize the /new greeting (e.g. "Hey David" vs. "Hey, what's your name?") on the first turn. They are forced into a two-turn workaround where Turn 1 asks the name and Turn 2 uses the envelope that finally arrives with the first real user message.

Repro

1. Configure any Telegram agent (e.g. the Divergent Systems bot) with a persona that should greet known senders by name.
2. From a Telegram account that is in the agent's contacts.json, send /new.
3. Observe the run jsonl for that turn: the user message contains only the bare reset prompt — no Sender (untrusted metadata): {...} block.
4. The LLM therefore greets generically ("What's your name?") even though Telegram already knows the sender's id and display name.

Root cause

src/auto-reply/reply/get-reply-run.ts around lines 281–300:

```ts const isBareNewOrReset = rawBodyTrimmed === "/new" || rawBodyTrimmed === "/reset"; const isBareSessionReset = isNewSession && ((baseBodyTrimmedRaw.length === 0 && rawBodyTrimmed.length > 0) || isBareNewOrReset); const baseBodyFinal = isBareSessionReset ? buildBareSessionResetPrompt(cfg) : baseBody; const envelopeOptions = resolveEnvelopeFormatOptions(cfg); const inboundUserContext = buildInboundUserContextPrefix( isNewSession ? { ...sessionCtx, /* history stripped */ } : { ...sessionCtx, ThreadStarterBody: undefined }, envelopeOptions, ); const baseBodyForPrompt = isBareSessionReset ? baseBodyFinal // <-- envelope dropped : [inboundUserContext, baseBodyFinal].filter(Boolean).join("\n\n"); ```

The isBareSessionReset branch deliberately throws away inboundUserContext, which is the exact block that contains the sender metadata. buildInboundUserContextPrefix in src/auto-reply/reply/inbound-meta.ts already builds that block correctly from sessionCtx.SenderName / SenderId / SenderUsername, and the Telegram channel adapter populates those fields for /new in extensions/telegram/src/bot-native-commands.ts around lines 870–900.

So everything upstream works — the metadata just gets stripped off at the last step.

Proposed fix

Always prepend inboundUserContext, even on bare session reset:

```ts const baseBodyForPrompt = [inboundUserContext, baseBodyFinal].filter(Boolean).join("\n\n"); ```

This is a one-liner for the channel reply path. The agent's system prompt already tells the LLM the envelope is untrusted metadata, so no trust boundary changes.

Secondary call site

src/gateway/server-methods/agent.ts around line 525 also calls buildBareSessionResetPrompt(cfg) in the JSON-RPC agent handler, but there is no rich sessionCtx in scope there. Fixing that path would require looking up session origin from the session entry — probably a follow-up.

Impact

Unblocks personalized /new greetings for every agent in every channel that populates sender metadata (Telegram confirmed; Discord/iMessage adapters likely already populate the same fields). Today's workaround is a forced two-turn flow in each agent's IDENTITY.md, which is brittle and wastes a turn.

Environment

• openclaw b5c597cc66 (local build), behavior also present in upstream/main 9e4f478f86
• Channel: Telegram (bot-native commands)
• Observed via DS bot @DivergentSystemsBot