Summary

When a subagent completes and triggers an announce back to the requester session, if the main session file is concurrently being written (e.g., main agent is sending a Feishu message at the same moment), the announce throws EmbeddedPromptLockFileChangedError and retries — even though the message was already successfully delivered. This causes duplicate messages in the channel.

Steps to Reproduce

1. Spawn a subagent via sessions_spawn
2. Subagent completes and triggers announce to main session
3. At the same moment, main agent is writing to the session file (e.g., delivering a reply to the user)
4. Announce throws session file changed while embedded prompt lock was released
5. Announce retries 3 times (5s / 10s / 20s intervals)
6. Each retry successfully sends the message → user receives 3 identical duplicate messages

Log Evidence

feishu: sendText target=user:ou_xxx, textLength=25          ← message sent successfully
[warn] Subagent completion direct announce failed: text completion direct delivery failed: session file changed while embedded prompt lock was released: /path/to/session.jsonl
feishu: sendText target=user:ou_xxx, textLength=25          ← retry #1, duplicate sent
[warn] Subagent completion direct announce failed: ... session file changed ...
feishu: sendText target=user:ou_xxx, textLength=25          ← retry #2, duplicate sent
[warn] Subagent announce give up (retry-limit) run=xxx

Root Cause

In subagent-announce-delivery.js, the error session file changed while embedded prompt lock was released (thrown as EmbeddedPromptLockFileChangedError) is not classified in either:

• TRANSIENT_ANNOUNCE_DELIVERY_ERROR_PATTERNS
• PERMANENT_ANNOUNCE_DELIVERY_ERROR_PATTERNS

Despite this, the announce still retries. Critically, no idempotency check is performed before retrying — the code does not verify whether the message was already delivered before attempting to send again.

Suggested Fix

Either:

1. Add /session file changed while embedded prompt lock was released/i to PERMANENT_ANNOUNCE_DELIVERY_ERROR_PATTERNS to prevent retry after this error
2. Record a delivery idempotency token when sendText succeeds, and check it before each retry to avoid duplicate delivery

Environment

• OpenClaw version: 2026.6.1-beta.3 (3b53e2c)
• Channel: Feishu (via @larksuite/openclaw-lark plugin v2026.5.20)
• OS: Darwin 25.3.0 (arm64)
• Node: v24.14.1
• Reproducible: consistently, every subagent announce when main session has concurrent write activity