openclaw - 💡(How to fix) Fix Auto-compaction produces empty-content assistant messages that hard-kill sessions (Venice HTTP 400) [1 comments, 2 participants]

Summary

When auto-compaction fails or is interrupted mid-run, it can write structurally invalid assistant messages with content: [] (empty array) into the session JSONL file. Every subsequent turn then returns HTTP 400 (no body) from Venice — zero input tokens, zero output tokens — because the API rejects a conversation history containing empty assistant turns. The fallback provider (OpenAI) also fails for the same reason: same broken payload, same rejection.

The session becomes permanently dead. OpenClaw's own repair pass (session file repaired: rewrote 2 assistant message(s)) does not fully resolve it — the 400s continue after repair.

Reproduction

1. Run a tool-heavy session until context overflow is detected during a tool loop
2. Auto-compaction triggers (compactionAttempts=0, source=assistantError)
3. Compaction fails or is interrupted (e.g. WhatsApp watchdog timeout firing simultaneously)
4. Session JSONL now contains trailing assistant messages with content: [], stopReason: error, usage.input: 0, usage.output: 0
5. Every subsequent turn — including a simple "ping" — returns 400 status code (no body) from Venice immediately
6. session file repaired log entry appears but does not resolve the 400s
7. Session is unrecoverable without manual file quarantine/rename

Log Evidence

2026-04-28T18:28:52.442+01:00 [agent/embedded] [context-overflow-diag] sessionKey=*** provider=venice/claude-sonnet-4-6 source=assistantError messages=407 compactionAttempts=0 error=Context overflow: estimated context size exceeds safe threshold during tool loop.
2026-04-28T18:28:52.442+01:00 [agent/embedded] context overflow detected (attempt 1/3); attempting auto-compaction for venice/claude-sonnet-4-6

2026-04-28T21:12:46.854+01:00 [agent/embedded] session file repaired: rewrote 2 assistant message(s) (b21b827c-...)
2026-04-28T21:12:47.662+01:00 [agent/embedded] embedded run agent end: runId=... isError=true model=claude-sonnet-4-6 provider=venice error=400 status code (no body)
2026-04-28T21:12:48.289+01:00 [agent/embedded] embedded run agent end: runId=... isError=true model=claude-sonnet-4-6 provider=venice error=400 status code (no body)

2026-04-28T21:19:43.895+01:00 [agent/embedded] session file repaired: rewrote 2 assistant message(s) (b21b827c-...)
2026-04-28T21:19:44.755+01:00 [agent/embedded] embedded run agent end: runId=... isError=true model=claude-sonnet-4-6 provider=venice error=400 status code (no body)

Poisoned Session File Pattern

Trailing entries in the dead session JSONL:

{
  "type": "message",
  "message": {
    "role": "assistant",
    "content": [],
    "provider": "venice",
    "model": "claude-sonnet-4-6",
    "usage": { "input": 0, "output": 0 },
    "stopReason": "error",
    "errorMessage": "400 status code (no body)"
  }
}

Multiple such entries exist consecutively at the tail of the file. content: [] violates the Anthropic/OpenAI API contract — empty assistant turns are structurally invalid.

Impact

• Session becomes permanently unrecoverable
• The user sees [assistant turn failed before producing content] on every turn including trivial inputs like "ping"
• The fallback provider also fails (same payload sent to both)
• session file repaired log entry is misleading — it does not resolve the 400s
• Without manual intervention, every new message appended to the session continues to fail

Expected Behaviour

Compaction failure should either:

1. Not write empty-content assistant entries into the session file, or
2. Detect that a session tail contains consecutive zero-token error-stop assistant messages and quarantine/reset automatically before attempting the next turn

Workaround

Rename the poisoned .jsonl to .jsonl.poisoned so it is not resumed. OpenClaw will create a fresh session on the next turn.

Environment

• OpenClaw: 2026.4.25 (aa36ee6)
• Provider: venice/claude-sonnet-4-6 via openai-completions API
• Compaction mode: safeguard
• Compaction model: openai/gpt-4o-mini
• Context at overflow: 407 messages, overflow during tool loop
• OS: macOS Darwin 25.4.0 arm64