Summary

When the bundled claude-cli backend hits AbortError / no-output / wedged-session conditions, Claude Code CLI injects a synthetic assistant message with model: "<synthetic>" and content "No response requested.". OpenClaw's stream-json parser does not extract text from assistant message blocks (only tool_use / tool_result), and the trailing type: "result" event commonly carries an empty result string in this case. The turn is then classified as empty and dropped silently. The user sees no message, no status, no retry indicator.

Why it matters

End users can't distinguish "OpenClaw is stuck" from "model decided not to reply." In a long-running multi-step session (workflows, deep-research, PDF generation), this consistently happens after large tool-heavy turns. The user has to manually probe with "are you there?" to trigger a fresh turn.

Reproduction shape

• backend: claude-cli/Claude-Opus-4.7 (bundled @anthropic-ai/claude-code)
• channel: telegram (group with topics, but also reproduces in DM)
• trigger: long context turn after a heavy tool sequence (e.g. background workflow + HTML/PDF generation) followed by a new user message
• transcript line shape:

  {
    "type": "assistant",
    "message": {
      "model": "<synthetic>",
      "role": "assistant",
      "content": [{"type": "text", "text": "No response requested."}],
      "stop_reason": "stop_sequence",
      "usage": {"input_tokens": 0, "output_tokens": 0}
    }
  }

• Observed 4 times in a single 4h conversation on 2026-06-05

Root cause (from code inspection)

1. src/agents/cli-output.ts:573-602 — assistant message handler only iterates tool_use / tool_result blocks, no text extraction. Synthetic-message text is silently dropped at parse time.
2. src/agents/cli-output.ts:374-388 — type: "result" extractor returns {text: ""} when result string is empty (a real codepath for tool-only turns).
3. src/agents/embedded-agent-runner/result-fallback-classifier.ts:11-13 — hasDeliberateSilentTerminalReply calls isSilentReplyPayloadText, which only matches NO_REPLY (case-insensitive). No response requested. is not in the silent-token set.
4. result-fallback-classifier.ts:73 — if (!isGpt5ModelId(params.model)) return null — non-GPT-5 backends never trigger the auto-fallback/retry branch for empty/synthetic results.
5. Net effect: empty turn → no outbound delivery, no retry, no user-visible signal.

Background on the upstream placeholder

No response requested. is a Claude Code CLI synthetic assistant message injected by ensureToolResultPairing and similar repair paths, exported as NO_RESPONSE_REQUESTED in src/utils/messages.ts of Claude Code, marked with model: "<synthetic>" (sources: Yuyz0112/claude-code-reverse, claude-code-best/claude-code). It's a well-known noise string filtered out by third-party Claude Code GUIs (e.g. sting8k/pi-vcc, edmundmiller/dotfiles).

Proposed fix

Minimum (correctness)

1. In cli-output.ts parseClaudeStreamJsonResult, detect parsed.type === "assistant" with message.model === "<synthetic>" and content text matching the known synthetic-placeholder set ("No response requested.", "Continue from where you left off."). Tag the result with a new classification, e.g. synthesized-placeholder.
2. In result-fallback-classifier.ts, treat synthesized-placeholder as hasDeliberateSilentTerminalReply === true AND additionally surface a structured status signal back to the channel adapter (so Telegram can render a status reaction or one-shot debug breadcrumb).

Nice-to-have (observability)

3. Lift the GPT-5-only restriction on auto-retry in classifyEmbeddedAgentRunResultForModelFallback, or at least extend it to classify synthesized-placeholder as retryable for claude-cli backends (Claude CLI's own AbortError path already produces these placeholders; the retry could be one-shot to avoid loops).
4. Add a silentErrorReplies-style status surface for synthetic-placeholder turns so users can see "🟡 backend produced no output, no retry" instead of nothing.

Related

• [Bug]: Telegram DM fabricates silent-reply chatter for no-visible-response turns #70628 — opposite direction of the same defect surface (Telegram DM fabricates No added response from me.)
• [Bug]: Telegram streaming=partial silently drops response when model fallback occurs mid-stream #44340 — Telegram streaming=partial silent drop on model fallback (closed 2026-04-24)
• [claude-cli backend] Auto-invalidate stored CLI session id after AbortError / FailoverError #78785 — Auto-invalidate claude-cli session id after AbortError (closed 2026-05-17) — directly relevant: the wedged-session scenario this fix addresses is the most common trigger of the synthetic placeholder
• pi-embedded-runner: reasoning-only retry short-circuited in group chats by silentReplyPolicy default #85684 — silentReplyPolicy short-circuits reasoning-only retry (related observability gap)

Environment

• OpenClaw 2026.6.1
• claude-cli/Claude-Opus-4.7
• macOS Darwin 25.4.0 (arm64)
• node v26.0.0
• channel: telegram (group + topic)