🤖 fix: agent tool inheritance from base agents

[ammar-agent]

Summary

Fixed a bug where agents using base inheritance (like ask which extends exec) had no tools available despite claiming tools in the UI.

Root Cause

resolveToolPolicyForAgent only looked at the current agent's frontmatter.tools.add patterns, ignoring the base inheritance chain.

For example, the ask agent has:

base: exec
tools:
  remove:
    - file_edit_.*

It has no add patterns - it relies on inheriting add: [.*] from exec. The old code only saw an empty add list, so it created a policy that denied all tools.

Fix

Modified resolveToolPolicyForAgent to:

1. Accept the full list of agents (including built-ins)
2. Walk the inheritance chain using collectToolConfigs() (base first, then child)
3. Apply each layer's add/remove patterns in order

Now for ask (base: exec):

• Start with deny-all baseline
• Process exec: enable .*, disable propose_plan, disable ask_user_question
• Process ask: disable file_edit_.*

Result: All tools enabled except propose_plan, ask_user_question, and file_edit_.* - exactly as intended.

Testing

• Updated existing tests to use new interface
• Added tests for single-level and multi-level inheritance
• All tests pass

---

Generated with mux • Model: anthropic:claude-opus-4-5 • Thinking: high

[chatgpt-codex-connector]

💡 Codex Review

mux/src/node/services/aiService.ts

Lines 1149 to 1153 in f0156ae

	{
	id: effectiveAgentId,
	base: agentDefinition.frontmatter.base,
	tools: agentDefinition.frontmatter.tools,
	},

Include full agent list when resolving tool inheritance

resolveToolPolicyForAgent now walks the provided agents list to apply inherited add/remove rules, but allAgents is built from the built-ins and just the requested agent. Any agent that inherits from another workspace-defined agent (non built-in) will still resolve with a deny-all policy because its base is absent from this list, so the base tools.add patterns are never applied. Agents that rely on custom base definitions therefore continue to lose their tools.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".