Summary

After upgrading to OpenClaw 2026.4.8, very large Telegram sessions repeatedly hit context overflow and compaction timeout. In this incident, that appeared to cascade into prolonged gateway draining/task rejection, repeated subagent announce failures, and surprising fallback behavior.

This does not look like just “Anthropic fallback is broken.”
The stronger bug shape is a failure-chain:

1. huge-session overflow + compaction timeout
2. gateway enters draining/restart state and rejects new tasks
3. subagent completion announce retries fail/give up
4. fallback decisions during drain can still route into stale runtime fallback candidates

The user reports this behavior did not happen before this version.

Environment

• OpenClaw: 2026.4.8
• OS: macOS arm64
• Channel: Telegram (multiple accounts)
• Large long-lived sessions (hundreds to >1300 messages)

Evidence (local)

1) Massive overflow + compaction timeout

From artifacts/openclaw-2026-04-08-incident-extract.txt:

• estimatedPromptTokens=1014988 / overflowTokens=759372
• messages=1331 / messages=1351+ on affected Telegram sessions
• multiple compaction failures at ~900s:
  • outcome=failed reason=timeout durationMs=900342
  • outcome=failed reason=timeout durationMs=900533

2) During failure chain, gateway rejects tasks as draining

From incident extract and ~/.openclaw/logs/gateway.err.log:

• GatewayDrainingError: Gateway is draining for restart; new tasks are not accepted
• repeated drain timeout reached; proceeding with restart

3) Subagent completion announce failures/retries during this state

• Subagent completion direct announce failed ... GatewayDrainingError
• Subagent announce completion ... transient failure, retrying
• Subagent announce give up (retry-limit)

4) Anthropic fallback still appeared in live fallback decisions during drain

Even after removing Anthropic fallback from config on disk, log lines during draining still showed:

• next=anthropic/claude-haiku-4-5 detail=Gateway is draining for restart; new tasks are not accepted
• and auth failure attempts:
  • candidate=anthropic/claude-haiku-4-5 reason=auth ... HTTP 401 authentication_error: invalid x-api-key

5) On-disk config had Anthropic removed, but runtime lagged until restart

• Current ~/.openclaw/openclaw.json fallback list is only:
  • openai-codex/gpt-5.4 -> openai-codex/gpt-5.3-codex
• Local commit removing Anthropic fallback:
  • 19172db Remove Anthropic model fallback config
• openclaw.json metadata shows it was touched at 2026-04-08T17:10:23.793Z
• But runtime logs still had next=anthropic/claude-haiku-4-5 at 2026-04-08T17:10:27.547+01:00

This suggests live runtime config/fallback chain can remain stale until gateway restart/reload.

Doctor note

openclaw doctor --fix was run locally, but this alone did not reload the running gateway process.

Expected behavior

1. Very large-session overflow/compaction failure should degrade gracefully without cascading into prolonged drain/task rejection loops.
2. Subagent completion announce should not be lost/give-up during gateway drain windows.
3. Runtime fallback chain should not continue using removed fallback providers after config changes are applied to disk.
4. If restart/reload is required for fallback-chain changes, surface this clearly in CLI/doctor output.

Actual behavior

• Overflow + compaction timeout chain coincided with gateway draining errors and task rejection.
• Subagent announce retries frequently failed/gave up.
• Fallback routing still referenced Anthropic during draining, causing 401 auth errors, despite Anthropic fallback being removed on disk.

Potentially related (but not exact duplicate)

• Compaction timeout: add escape hatch to prevent session hang #44031 (compaction timeout hangs)
• Compaction deadlock blocks session recovery (/new, /reset queue behind timed-out compaction) #40295 (compaction deadlock/session recovery)
• GatewayDrainingError should auto-retry, not surface to user #55412 (GatewayDrainingError retry behavior)
• Subagent announce gives up when requester session is busy (direct call timeout → should queue) #54276 (subagent announce give-up)
• Post-upgrade stability regressions — v2026.4.5 (3e72c03) #62095 (doctor --fix expectations)

Request

Please investigate this as a possible 2026.4.8 regression/failure-chain interaction:

• large-session overflow/compaction timeout
• gateway drain/restart task rejection behavior
• subagent announce resilience during drain
• runtime config/fallback-chain reload semantics (especially after removing providers)

If useful, I can provide the extracted incident artifacts/log snippets listed above.