refactor(agent): extract text tool-call parser

[alectimison-maker]

Summary

• extract the raw-text tool-call parser from both browser Agent classes into byte-identical, browser-free modules
• keep the existing Agent method as a compatibility wrapper with the current default and mode-specific allowlists
• exercise every currently supported text format directly against production parser code in both builds

Motivation

Local/chat-template models can return tool calls as raw text rather than through a provider's structured tool_calls field. That fallback parser had grown inside both large Agent classes, while most formats were not directly regression-tested. Isolating it makes the security boundary and future parser changes easier to review without changing current behavior.

Design

• preserve the existing 10,000-character cap, allowlist checks, parsing order, output shape, and fallback IDs
• keep browser-specific tool-name sets at the Agent call sites
• mirror the small module in each independently packaged extension tree and enforce byte parity in tests
• leave parser behavior improvements and broader context-trimming work for separate changes

Testing

• node test/run.js: 1329 passed, 1 inherited failure (package.json is 25.9.7 while the newest CHANGELOG.md entry is 25.9.0)
• npm run test:security: 60/60
• npm run test:fixtures: 125/125
• npm run test:webmcp: passed with Chrome 150
• npm run test:ci: passed
• node --check on both parser modules and both Agent modules
• git diff --check

Compatibility

No public API, tool schema, accepted format, or browser packaging behavior changes.

[vercel]

@alectimison-maker is attempting to deploy a commit to the esokullu's projects Team on Vercel.

A member of the Team first needs to authorize it.