OpenClaw 'No Callable Tools Remain' Bug: modelRun + Allowlist Conflict Blocks All LLM-Only Inference

Three linked GitHub issues (April 28-30) reveal a systemic bug in OpenClaw 4.25-4.27 where LLM-only model runs (tools intentionally disabled) fail with 'No callable tools remain' when the runtime inherits any explicit tool allowlist from tools.profile or tools.alsoAllow. Affects CLI gateway mode, Lobster workflow llm-task calls, and any embedded model run.

Product Idea from this Signal

A CLI tool that snapshots your OpenClaw state before updates, runs the upgrade in a sandboxed dry-run against your live config, and auto-rolls back if any health check fails

1293.8k ▲

OpenClaw ships updates every 2-3 days and each one risks breaking exec permissions, crashing the gateway, saturating CPU with plugin loading bugs, or silently deleting cron jobs. Four separate regressions in April 2026 alone (v2026.4.1 sandbox, v2026.4.5 CPU, v2026.4.14 memory timeouts, update command wiping Feishu and crons) hit thousands of self-hosters with no rollback path. The built-in openclaw doctor only validates config schema after the fact. This tool wraps the entire update lifecycle: snapshot state, dry-run the upgrade against your actual agent/cron/channel config, diff the before/after, and block or auto-rollback if anything regresses.

CLIOPEN-SOURCEDEVTOOLSELF-HOSTEDRELIABILITY

CompetitiveView Opportunity →