Nightly E2E failure: agent reply JSON extraction broken by OpenClaw schema change

Run: https://github.com/NVIDIA/NemoClaw/actions/runs/26260886472
Date: 2026-05-22
Commit: 74c0246
Affected jobs (5):

• sandbox-operations-e2e (job 77293833899)
• bedrock-runtime-compatible-anthropic-e2e (job 77293833749)
• messaging-compatible-endpoint-e2e (job 77293833829)
• openclaw-inference-switch-e2e (job 77293833863)
• launchable-smoke-e2e (job 77293833838)

Bug group: agent-reply-json-extract
Failure class: config_error
Confidence: high

Root cause

OpenClaw ≥ 2026.5.18 changed its openclaw agent --json output from nesting payloads under {"result": {"payloads": [...]}} to returning them at the top level {"payloads": [...]}. Five E2E test scripts use inline Python extractors that hardcode doc.get("result") or {}, which returns {} under the new schema — so .get("payloads") always yields [] → empty reply → assertion failure.

Evidence

All 5 jobs show the same pattern:

• Raw agent output contains correct payload text (e.g., "42", "PONG") at top level
• Python extractor returns reply='' because it looks only in doc["result"]
• Assertion fails on empty reply

Fix

PR: #4030

The one-line fix in each script changes doc.get("result") or {} → doc.get("result") or doc, which falls back to the document root when the result wrapper is absent. Backwards-compatible with the old schema.

Files changed

• test/e2e/test-bedrock-runtime-compatible-anthropic.sh (line 713)
• test/e2e/test-launchable-smoke.sh (line 531)
• test/e2e/test-messaging-compatible-endpoint.sh (line 534)
• test/e2e/test-openclaw-inference-switch.sh (line 284)
• test/e2e/test-sandbox-operations.sh (line 393)