Problem
The 20260701 scored run's debug bundle (openclaw_failure_cases_debug_bundle_20260703, machine bot1) reports two OpenClaw CLI compat failures that we cannot reproduce locally and could not attribute because the CLI version was never recorded:
openclaw agent --session-key agent:main:bench-<id> rejected with a parameter error (the adapter had to switch to --agent main --session-id bench-<id>);
- provider config
timeoutSeconds rejected by the config schema (had to be removed from the generated openclaw.json).
The bundle's run occurred on bot1 around 2026-07-01, suspected on a newer CLI than ours (possibly 2026.6.11). On the locally installed OpenClaw 2026.6.10 (aa69b12) BOTH forms work:
So the current adapter form (--agent --session-id, no timeoutSeconds) is a cross-version conservative choice — the only form with green evidence on both known environments — not a verified fix. Whether/when the old form actually breaks, and on which version, is unknown.
What's now in place (PR #81)
openclaw --version is executed once per process and logged in run.log (OpenClaw CLI version: ...);
- every task result carries
agent_metadata.openclaw_cli_version.
Next time a CLI compat failure appears
- Read the failing task's
agent_metadata.openclaw_cli_version (or the run.log line) — the version is no longer a mystery.
- Diff against the last-known-good version;
npm install -g openclaw@<version> locally and reproduce with a single bubench run --agent openclaw --data LexBench-Browser --mode single.
- Record the exact rejected flag/config field and version pair here, so the adapter's compat choices stop being guesses.
Optional proactive close-out: obtain openclaw --version from the bot1 machine (bundle author) and re-test both CLI forms on that exact version — that would confirm or refute the bundle report's claims directly.
🤖 Generated with Claude Code
Problem
The 20260701 scored run's debug bundle (openclaw_failure_cases_debug_bundle_20260703, machine bot1) reports two OpenClaw CLI compat failures that we cannot reproduce locally and could not attribute because the CLI version was never recorded:
openclaw agent --session-key agent:main:bench-<id>rejected with a parameter error (the adapter had to switch to--agent main --session-id bench-<id>);timeoutSecondsrejected by the config schema (had to be removed from the generated openclaw.json).The bundle's run occurred on bot1 around 2026-07-01, suspected on a newer CLI than ours (possibly 2026.6.11). On the locally installed OpenClaw 2026.6.10 (aa69b12) BOTH forms work:
openclaw agent --helpdocuments--session-keyand--session-idside by side;--session-key+ providertimeoutSeconds) green on this machine (run 20260704_002844);So the current adapter form (
--agent --session-id, notimeoutSeconds) is a cross-version conservative choice — the only form with green evidence on both known environments — not a verified fix. Whether/when the old form actually breaks, and on which version, is unknown.What's now in place (PR #81)
openclaw --versionis executed once per process and logged in run.log (OpenClaw CLI version: ...);agent_metadata.openclaw_cli_version.Next time a CLI compat failure appears
agent_metadata.openclaw_cli_version(or the run.log line) — the version is no longer a mystery.npm install -g openclaw@<version>locally and reproduce with a singlebubench run --agent openclaw --data LexBench-Browser --mode single.Optional proactive close-out: obtain
openclaw --versionfrom the bot1 machine (bundle author) and re-test both CLI forms on that exact version — that would confirm or refute the bundle report's claims directly.🤖 Generated with Claude Code