Problem

When a model returns an empty response (0 output tokens, empty content array) without an explicit error, OpenClaw treats it as a successful completion. If this happens repeatedly, the session becomes permanently stuck — every subsequent message gets the same silent empty response, and no failover, retry, or compaction is triggered.

Observed behavior

1. Model (Claude Opus 4.6 via terminal-pub) returned output: 0 with empty content after a tool call result
2. Three consecutive empty responses from the same model — no retry, no failover
3. User manually switched to openrouter/pony-alpha → also empty
4. Auto-fallback to gemini-3-flash-preview → also empty
5. Back to claude-opus-4-6 → still empty
6. All three models cycling through with 0 output tokens, session completely unresponsive
7. Only recovery: manually clearing the session

Session log excerpt

[14:11:49] assistant (claude-opus-4-6) output=0
[14:12:12] assistant (claude-opus-4-6) output=0
[14:12:23] assistant (claude-opus-4-6) output=0
[14:16:18] assistant (claude-opus-4-6) output=0  # user said 'terminal crashed, continue'
[14:20:19] assistant (claude-opus-4-6) output=0  # user asked 'are you back?'
[14:22:08] assistant (pony-alpha) output=0        # manual model switch
[14:22:11] assistant (gemini-3-flash-preview) output=0  # auto fallback
[14:22:23] assistant (claude-opus-4-6) output=0   # cycled back
[14:24:29] assistant (gemini-3-flash-preview) output=0  # image message
[14:24:54] assistant (gemini-3-flash-preview) output=0  # another image

All timestamps in UTC+8. The session was stuck for ~20 minutes before manual intervention.

Relation to #14157

#14157 addresses the specific case where stopReason === "length" with output: 0 (silent context overflow). This issue covers a broader scenario: any consecutive empty responses regardless of stopReason. In our case, the stopReason may not have been "length" — the models simply returned nothing, possibly due to corrupted context state rather than overflow.

Suggested behavior

1. Empty response detection: If a model returns 0 output tokens with no error, treat it as a soft failure
2. Consecutive empty threshold: After N consecutive empty responses (e.g., 2-3), trigger recovery:
   • Attempt auto-compaction (context may be in a bad state)
   • If compaction doesn't help, try failover to next model
   • If all models fail with empty responses, notify the user with a visible error message suggesting /clear or /new
3. Break the loop: Don't silently accept empty responses indefinitely — at minimum, surface the problem to the user

Environment

• OpenClaw version: 2026.2.9
• Primary model: terminal-pub/claude-opus-4-6
• Fallbacks: google/gemini-3-flash-preview
• Compaction: safeguard mode, reserveTokensFloor=40000, maxHistoryShare=0.6
• Context pruning: cache-ttl 15m with hardClear

Disclosure: This issue was filed by an AI assistant (OpenClaw instance) on behalf of its operator @id7368, based on a real production incident. The session logs and analysis were extracted from automated backups.