openclaw - 💡(How to fix) Fix Bug: Bedrock Converse Streaming silently aborts on long-context agent sessions (~6 min timeout, no retry, no fallback)

Summary

Agent sessions using the amazon-bedrock provider abort silently after ~6 minutes when handling long-context prompts that require long streaming responses. The session dies with no retry, no fallback, and no transcript written — all work is lost.

Environment

• OpenClaw version: 2026.5.4
• Provider: amazon-bedrock (Bedrock Converse Streaming API)
• Model: amazon-bedrock/global.anthropic.claude-sonnet-4-6
• Trigger: Large prompt (~36,000 chars) + expected long output (exam paper response)
• Platform: macOS, launchd-managed gateway

Reproduction

1. Configure an agent with amazon-bedrock/global.anthropic.claude-sonnet-4-6 as primary model
2. Send a session prompt of ~36,000 characters that requires a long, detailed response
3. Observe session abort at approximately 6 minutes

Observed behaviour

Session trajectory metadata after abort:

{
  "status": "error",
  "aborted": true,
  "externalAbort": false,
  "timedOut": false,
  "idleTimedOut": false,
  "timedOutDuringCompaction": false,
  "timedOutDuringToolExecution": false,
  "promptError": "This operation was aborted"
}

• Session started: 2026-05-29T02:04:32Z
• Session ended: 2026-05-29T02:10:33Z (6 min 1 sec)
• Session .jsonl transcript file never written (ENOENT) — no output saved
• externalAbort: false — not killed externally
• timedOut: false — did not hit the configured session timeout
• promptError: "This operation was aborted" — abort originated inside the Bedrock streaming call

Expected behaviour

1. On stream drop: automatic retry (at least once) before declaring failure
2. If retry exhausted: route to configured fallback model rather than failing the session
3. If session must fail: flush partial transcript to disk before dying
4. Surface a meaningful error to the operator (not just "This operation was aborted")

Root cause hypothesis

The Bedrock Converse Streaming API appears to silently drop the connection after ~6 minutes on large payload requests. This is not surfaced as a named timeout error — it manifests as an internal abort in the bedrock-converse-stream adapter. The 6-minute window is suspiciously consistent and may correspond to an undocumented AWS streaming connection limit.

Impact

• All work is lost — no transcript written, no partial output saved
• Silent failure — operator sees only "This operation was aborted" with no actionable context
• No fallback fires — configured fallback models are not tried on stream abort

Workaround

Provision ANTHROPIC_API_KEY on the instance and use anthropic/claude-sonnet-4-6 direct instead of Bedrock. No abort on the same prompt.

Related

• #87318 — params.modelId override ignored on Bedrock provider (separate bug, same provider layer)

Reported by Haderach-Ram. Production instance, not a test environment.