Summary

When an agent sends media (image + caption) via the message tool mid-turn, the runtime records a mirrored assistant message (provider: "openclaw", model: "delivery-mirror") into the session transcript as a side-branch (leaf record, appendMode: "side") attached between an assistant tool_use turn and its toolResult entries. On every subsequent turn, the compiled Anthropic request is rejected with HTTP 400 (invalid request, ~0.5s, zero tokens). The session never recovers: every turn ends non_deliverable_terminal_turn, the user-facing channel receives nothing, and there is no operator alert. /status still works, so the bot appears healthy while being permanently mute.

Environment

• OpenClaw 2026.6.11 (e085fa1), macOS arm64, node v25.6.0
• Provider: anthropic (anthropic-messages API), model claude-opus-4-8
• Channel: Telegram group session

Reproduction (observed, 2x in one week)

1. Telegram group session; agent works normally.
2. Agent calls message tool with an image attachment + caption targeting the same group, mid-turn (before its final reply).
3. Transcript gains, in order: assistant msg with tool_use blocks → mirrored assistant msg {role:"assistant", content:[{type:"text", text:"<media filename>"}], provider:"openclaw", model:"delivery-mirror"} → {type:"leaf", targetId:<assistant msg id>, appendParentId:<mirror id>, appendMode:"side"} → the toolResult entries.
4. Next inbound message → provider POST returns 400 in <1s, lastCallUsage all zeros, transcript records assistant placeholder [assistant turn failed before producing content], turn ends non_deliverable_terminal_turn.
5. Every subsequent turn repeats step 4 forever. No delivery, no alert.

Update — second occurrence, text-only trigger

Three minutes after the first repaired session came back online, the agent re-poisoned it with a plain text message send issued in parallel with other tool calls (edit + sessions_spawn + message in one assistant turn). The mirror entry landed between the tool_use block and its tool_result entries again. So the trigger is not media-specific: it is the delivery mirror being appended as a side-branch while sibling tool calls from the same assistant message are still pending.

Impact

Two production group sessions killed in one week; one stayed silently dead for 4 days. Humans in the group assume the agent is ignoring them.

Expected

1. Media delivery mirroring must not create a transcript branch that compiles into an invalid provider message sequence (assistant text between tool_use and tool_result).
2. Self-healing: after N consecutive instant 400s on an unchanged transcript, the runtime should repair/quarantine the offending branch (or at minimum surface a loud error to the channel/operator) instead of failing silently forever.

Evidence available

Sanitized transcript excerpts (entry ids, parent links, leaf record) and gateway log excerpts showing the 400 streak — happy to attach.