Summary

When /reasoning on is explicitly enabled in an iMessage-connected session, reasoning/thinking blocks are delivered as Reasoning:-prefixed iMessages. These outbound messages are re-ingested by the iMessage provider as new inbound messages, triggering new agent runs and creating an infinite feedback loop that produces hundreds of runaway agent runs until force-killed.

This is a composite bug formed by two known issues firing simultaneously:

1. Reasoning block delivery to iMessage — explicit /reasoning on still leaks thinking blocks as visible messages to iMessage. Reasoning/thinking blocks leaking into iMessage channel delivery #18667 reported this and was closed as superseded by PR fix(agents): suppress reasoning blocks from channel delivery #18935, but leakage persists on 2026.2.22-2 and 2026.2.23 when reasoning is explicitly enabled via /reasoning on.
2. 2. iMessage echo re-ingestion — outbound iMessages are picked up by the imsg polling bridge as new incoming messages (iMessage self-chat creates echo loop despite v2026.1.22 fixes #1649, still open).
      Why this is distinct from Reasoning/thinking blocks leaking into iMessage channel delivery #18667: That issue addresses only the reasoning leakage side. Even if fix(agents): suppress reasoning blocks from channel delivery #18935 fully resolves leakage, the underlying iMessage echo loop (iMessage self-chat creates echo loop despite v2026.1.22 fixes #1649) remains unfixed — meaning any future regression in reasoning suppression would immediately produce another runaway loop. This issue tracks the composite interaction and the need for defense-in-depth (dedup + suppression + rate limiting).

Steps to reproduce

1. Connect an iMessage session via the imsg CLI bridge
2. Enable /reasoning on in the session
3. Send any message that triggers a model response
4. Observe: the reasoning block is delivered as an iMessage, re-ingested as a new inbound message, triggering another agent run — looping indefinitely

The loop path:

/reasoning on → model produces thinking block
  → delivery pipeline formats as "Reasoning: ..." text
    → iMessage provider sends via imsg CLI
      → macOS stores message in chat DB
        → imsg polling picks it up as "new incoming message"
          → OpenClaw dispatches new agent run with "Reasoning: ..." as user input
            → model produces new thinking block → LOOP

Expected behavior

Reasoning/thinking blocks should not be delivered as visible messages to iMessage. Alternatively, outbound messages from the bot should be deduped and not re-ingested as inbound. At minimum, a rate limiter or circuit breaker should prevent hundreds of runaway agent runs.

Actual behavior

Hundreds of agent runs fire in rapid succession until the process is force-killed. Each iteration generates a new thinking block (the model sees "Reasoning: ..." as user input), produces a new response with reasoning, which gets delivered and re-ingested again. No rate limiting or circuit breaker exists to stop the loop.

OpenClaw version

2026.2.22-2, 2026.2.23

Operating system

macOS (iMessage requires macOS)

Install method

No response

Logs, screenshots, and evidence

Impact and severity

Affected: iMessage users with /reasoning on on 2026.2.22-2 and 2026.2.23
Severity: Critical (runaway agent loop, hundreds of runs before force-kill)
Frequency: 100% repro when both conditions met
Composite of #1649 (iMessage echo loop, still open) + reasoning leakage (#18667, closed by #18935 but not fully effective for explicit /reasoning on)

Additional information

Workaround: Do not enable /reasoning on in iMessage-connected sessions.

Suggested fixes:

• Outbound message dedup filter: After sending via imsg, record message text + timestamp in a short-lived cache (~30s TTL). Skip processing on poll if match. Client-side equivalent of what Slack/Discord handle at platform level.
• Suppress reasoning delivery for iMessage: Strip Reasoning: blocks when target is iMessage, matching what fix(telegram): filter reasoning segments when reasoningLevel is off #24518 does for Telegram.
• Rate limiting / circuit breaker: Detect rapid-fire runs from same channel (>10 runs/minute) and pause. Would limit runaway loops regardless of cause.
• Self-message detection: Broader fix for iMessage self-chat creates echo loop despite v2026.1.22 fixes #1649 — if sender matches bot's configured number or allowFrom entry, apply stricter dedup.

Related issues & PRs: #18667, #18935, #24290, #24335, #24518, #24626, #1649, #655, #659, #1031, #1080