Summary

openclaw infer model run --local --model <id> --prompt <p> can abort before attempting any model call when ~/.openclaw/openclaw.json has a non-empty tools.alsoAllow.

The error is:

No callable tools remain after resolving explicit tool allowlist (tools.allow: *, lobster, llm-task)

This happens even though infer model run sets modelRun:true / disables tools intentionally. The target model and auth are valid. The same model succeeds when tools.alsoAllow is removed in a temp config.

Reproduce

1. Configure ~/.openclaw/openclaw.json with:

   "tools": {
     "profile": "coding",
     "alsoAllow": ["lobster", "llm-task"]
   }

2. Run:

   openclaw infer model run --local --model openai-codex/gpt-5.5 --prompt 'reply exactly: ok' --json

Actual

The command aborts with:

No callable tools remain after resolving explicit tool allowlist (tools.allow: *, lobster, llm-task)

No provider request is made. The gateway never reaches the model call.

Expected

Because modelRun:true disables tools for this call shape, tool allowlist validation should not fail the model smoke. The command should attempt the model call and return the model response or a provider/model auth error.

Root cause

Observed in OpenClaw 2026.4.25 bundled output (file hashes will churn between builds; locations cited as observed):

• runModelRun() sets modelRun:true and promptMode:"none" (capability CLI module).
• The execution attempt path propagates disableTools: params.opts.modelRun === true.
• The selection module zeroes toolsRaw when disableTools || modelRun is true, but then still calls buildEmptyExplicitToolAllowlistError(...) because the global tools.alsoAllow has been expanded into an explicit allowlist of the form ["*", ...alsoAllow] by the sandbox tool-policy module.

So the modelRun path intentionally disables tools, but the empty-explicit-allowlist guard still evaluates the global tool policy and fails closed.

Observed source locations (bundled)

• dist/capability-cli-*.js — runModelRun()
• dist/attempt-execution-*.js — propagates disableTools
• dist/selection-*.js — toolsRaw zeroing + buildEmptyExplicitToolAllowlistError call
• dist/sandbox-tool-policy-*.js — tools.alsoAllow → explicit allowlist expansion

Suggested fix

Skip the empty-explicit-allowlist guard when tools are intentionally disabled:

if (!params.modelRun && !params.disableTools && /* existing condition */) {
  throw buildEmptyExplicitToolAllowlistError(...);
}

Alternatively, do not pass the global tool-policy block into that guard for modelRun/disableTools executions.

Impact

This breaks model/routing smoke tests and can produce false negatives during migrations.

We hit it during a gpt-5.5 routing cleanup on 2026-04-27 and initially rolled back a correct config patch because infer model run failed before making a provider call.

Workarounds

These production-path checks are unaffected:

• openclaw cron run <id>
• openclaw agent --agent <id> ...

A temp config that removes tools.alsoAllow also allows the same model/auth combo to succeed.

Environment

• OpenClaw: 2026.4.25
• Install: ~/homebrew/lib/node_modules/openclaw
• macOS: Darwin 25.3.0