fix(security): enforce disableExternalLLM in team worker contracts

[blue-int]

Summary

Enforce disableExternalLLM security setting at runtime. When OMC_SECURITY=strict or disableExternalLLM: true in config, getContract() throws for codex/gemini agent types, preventing external LLM workers from being spawned in team mode.

Claude workers are always allowed regardless of this setting.

Depends on

• fix(security): harden security-config with strict override protection #2016 (security-config with isExternalLLMDisabled())

Changes

• src/team/model-contract.ts: Guard in getContract() that checks isExternalLLMDisabled() for non-claude agents
• src/team/__tests__/model-contract.test.ts: 3 new tests (codex blocked, gemini blocked, claude allowed)

Test plan

• OMC_SECURITY=strict → getContract('codex') throws "disabled by security policy"
• OMC_SECURITY=strict → getContract('gemini') throws "disabled by security policy"
• OMC_SECURITY=strict → getContract('claude') succeeds
• 46/46 model-contract tests pass

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Repo admins can enable using credits for code reviews in their settings.

[Yeachan-Heo]

Reviewed via OMX session. REQUEST_CHANGES.

Blocking issues:

1. Logic-order bug in getContract() (src/team/model-contract.ts:223)
The external-LLM check now runs before validating that agentType is a known agent. In strict mode, getContract('unknown' as any) will throw the external-LLM policy error instead of the existing Unknown agent type... error. This changes behavior for invalid inputs and hides the real problem.

Fix: Move the disableExternalLLM enforcement after the agent-type validation guard.

2. Misleading error message
The thrown error says:

Set OMC_SECURITY to a non-strict value or configure security.disableExternalLLM=false to allow.

But in this PR's own security model, strict mode prevents config-file relaxation. The remediation suggestion contradicts the security policy.

What passed review:

• Enforcement is correctly scoped for known agents (codex/gemini blocked, claude allowed in strict)
• Mocked-fs security-config tests are solid (existsSync/readFileSync simulation, strict override protection, cache reset)
• silentAutoUpdate interaction test is sensible

Please fix the two blocking items and force-push.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[blue-int]

Fixed both blocking issues in 0a09dc5d:

1. Logic order: Moved disableExternalLLM guard after CONTRACTS[agentType] validation. Unknown agent types now correctly throw "Unknown agent type" first.

2. Error message: Removed misleading remediation suggestion. Now says:

   External LLM provider "codex" is blocked by security policy (disableExternalLLM). Only Claude workers are allowed in the current security configuration.

46/46 tests pass.

[Yeachan-Heo]

Solid work — this PR addresses all the concerns from the #2016 review:

✅ disableExternalLLM enforcement in getContract() — correct guard logic
✅ fs-mocked strict override tests — properly simulates config file override attempts
✅ silentAutoUpdate + security config interaction test — covers the gap
✅ Claude allowed even when external LLM disabled — correct

Minor nit (non-blocking)

The error message in getContract() suggests security.disableExternalLLM=false as an escape hatch, but in strict mode this config override is explicitly ignored by the || protection. Consider adjusting to: "Remove OMC_SECURITY=strict or use a non-strict security mode to allow external LLM providers." — but this is non-blocking.

Since this PR is a superset of #2016, we should close #2016 once this merges.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]