fix(droid): match tool names on word boundaries when inferring agent tools

[LukeTheoJohnson]

Problem

When converting an agent to a Factory Droid, mapAgentTools infers the agent's tool list by scanning its name, description, and body for tool names with bodyLower.includes(claudeTool). A tool name that appears as a substring of an ordinary word is treated as a reference, so a prose only agent that names no tools still gets a list:

body: "Review the request using already gathered context then do tasks
        as a checklist with details. Never open an editor and think globally
        about tradeoffs."

inferred tools: ["Edit", "Glob", "LS", "Read", "Task"]

Each is a false positive: already -> Read, details -> LS, tasks -> Task, editor -> Edit, globally -> Glob. Because the frontmatter tools list restricts the droid to those tools, a spurious match silently narrows the agent's access. An agent that should run with full tools gets locked to five it never asked for.

Fix

Match tool names on word boundaries (\bread\b) instead of as substrings. This is the same approach the kiro converter already uses for its tool-name mapping in claude-to-kiro.ts. Real references still resolve; incidental substrings no longer do.

Tests

Two cases added to tests/droid-converter.test.ts:

• the prose only agent above now infers no tools (was ["Edit", "Glob", "LS", "Read", "Task"])
• an agent that references Read and Bash still infers ["Execute", "Read"]

Full droid suite green (14/14); tsc --noEmit clean.

Note

The draft #358 also touches mapAgentTools: it adds a fast path for agents that declare an explicit tools field. Agents without one (the case above) still fall through to the substring inference.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: dfe8ee79ea

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".