Add claude_tool_bridge plugin

[spcdatapro]

Add claude_tool_bridge

Claude Tool Bridge makes Claude 5 models (for example claude-sonnet-5-5) work in Agent Zero's chat mode.
Agent Zero talks to Anthropic in chat mode, without native tools, so a model must write each call as
{"tool_name": ..., "tool_args": {...}}. Some Claude models keep writing their native syntax instead
(<invoke name="..."><parameter name="...">...</parameter></invoke>); Agent Zero treats those replies as unusable and
stops the agent after repeated ones. In a replay of 20 real prompts, claude-sonnet-5-5 produced an unusable reply in
45-90 % of the turns depending on its thinking settings, almost always a correct call in the wrong syntax.

• Translates the first complete <invoke> of a reply into a JSON tool request before _context_doctor sees it; prose
  written before the call becomes thoughts.

• Types arguments from the tool schema, then from Agent Zero's own usage examples; strings that look like JSON stay strings.

• Understands tool:method names, MCP tools, antml: prefixes, <function_calls> wrappers and thoughts / headline /
  tool_args written as parameters.

• Optionally stops the model stream at the first translatable </invoke> (early_stop, on by default).

• Leaves a reply untouched when it already holds a JSON call, when the <invoke> is quoted (code fence or mid-line),
  when the tool is unknown or a required argument is missing, or when the chat model is not configured
  (providers: [anthropic], model_patterns: ["claude-*"] by default).

• Plugin repository: https://github.com/spcdatapro/claude-tool-bridge

• License: MIT (LICENSE at the repository root)

• plugin.yaml name (claude_tool_bridge) matches the folder name in this PR

Checks

• Ran scripts/validate_plugin_submission.py locally on this branch, with the current generated-index release
  index.json in place for the duplicate checks: passed.
• The PR adds exactly one file: plugins/claude_tool_bridge/index.yaml.
• Tested on Agent Zero v2.13: 47 unit tests; an offline replay of 120 recorded model outputs (29 of 36 unusable
  Sonnet 5.5 replies rescued, no change to the 84 usable ones); a live replay of the same 20 prompts with the plugin
  (claude-sonnet-5-5: 18/20 to 1/20 unusable replies in one configuration and 9/20 to 1/20 in another;
  claude-opus-5-5 control unchanged at 0/20); and a multi-step task in a real instance.
• Overlap: the community context_doctor plugin repairs malformed JSON tool calls; this plugin handles replies that
  use Claude's XML call syntax instead of JSON, and defers to JSON whenever the reply holds a usable call.

Safety notes for reviewers

• No network calls, no subprocesses, no eval/exec, no third-party dependencies (standard library and the framework only).
• It only rewrites the text of the model reply in memory (message_loop_result) and wraps the response callback of the
  current model call (call_chat_model_turn). Nothing is written to disk and no API handlers are added.
• A translated call goes through the framework's normal path afterwards (history, _context_doctor, tool policy and
  tool execution). If anything in the plugin fails, the reply is left untouched.

🤖 Generated with Claude Code