Bug Summary

When a Codex-backed agent (e.g. GPT-5.5) calls the OpenClaw message(action=send) tool to deliver a progress update mid-turn, the Codex app-server treats the tool result as sufficient to release the turn. The agent announces intent, sends a progress message, and then goes completely silent — no further tool calls, no final response, no artifact written.

Reproduction

1. Agent receives a task from user via Telegram
2. Agent runs some bash commands (works fine)
3. Agent calls message(action=send) to send a progress update to Telegram: {"ok": true, "messageId": "9805"}
4. Turn ends. No further tool calls. No final response.
5. User asks "Did you complete?" → same cycle repeats

Observed in transcript (stopReason: "toolUse" on the assistant message — the model wanted to continue, but the app-server ended the turn).

Environment

• OpenClaw: 2026.7.1-beta.1 (4eb1d33)
• Agent: Stark (gpt-5.5, Codex app-server, yolo mode)
• Channel: Telegram
• Config: turnCompletionIdleTimeoutMs: 60000 (default), postToolRawAssistantCompletionIdleTimeoutMs: 300000

Root Cause Analysis

In run-attempt (dist), the message tool is handled as a dynamic tool with a 120s timeout (CODEX_DYNAMIC_MESSAGE_TOOL_TIMEOUT_MS = 12e4). After the tool returns successfully:

1. shouldBlockTerminalReleaseForNonTerminalDynamicToolResult(response) checks response.asyncStarted !== true
2. For the message tool, the response is synchronous (ok: true, messageId: ...), so asyncStarted is not set
3. This means shouldBlockTerminalReleaseForNonTerminalDynamicToolResult returns true → sets currentTurnHadNonTerminalDynamicToolResult = true
4. But then shouldReleaseTurnAfterTerminalDynamicTool checks !state.currentTurnHadNonTerminalDynamicToolResult — which is now false
5. The turn is released

The logic appears to be: synchronous dynamic tool results block the terminal-release shortcut, but the combination of currentTurnHadNonTerminalDynamicToolResult = true + no pending terminal dynamic tool + no active app-server turn requests results in the turn being released anyway via the idle watch.

The message tool should not be turn-terminal. It's a delivery mechanism, not a completion signal.

Proposed Fixes

1. Runtime fix (structural)

message(action=send) must never be turn-terminal unless the assistant has emitted a final response or an explicit terminal flag is set. The message tool is a delivery mechanism, not a completion signal. Consider either:

• Treating message results as async for turn-completion purposes (so asyncStarted = true), or
• Adding a turnTerminal: false flag to the message tool's response, or
• Excluding message from the terminal-release check entirely

2. Integration test

Add a test: agent sends progress via message, then runs a bash command, writes an artifact, and sends a final message. The test should fail if the bash command or artifact never executes.

3. Stalled-turn watchdog

If a turn sends only a message tool call and then has no further tool activity or final response within 30-60s, mark it as stalled. Either auto-resume the turn or alert the user.

4. Delivery metadata distinction

Distinguish progress_message from final_delivery in OpenClaw tool metadata so Telegram delivery does not collapse the work loop. The message tool could accept an optional kind: "progress" | "final" parameter. Progress messages would be explicitly non-terminal; final messages could optionally be terminal.

Workaround Applied

• Agent-side: Instructed agent via AGENTS.md to never send progress messages mid-turn — do all work first, then send only a final summary. This prevents the bug from triggering but limits agent communication.
• Config-side: Bumped turnCompletionIdleTimeoutMs from 60s → 300s. This is a shot in the dark — if the app-server is explicitly treating message as terminal, the timeout bump doesn't help.

Impact

This affects all Codex-backed agents that use the message tool mid-turn for progress updates. Stark (gpt-5.5) has been consistently failing on multi-step tasks because of this. Other agents (Donna, Hermione, Jack) may be affected but haven't exhibited the pattern as clearly.