Enhance router agent resilience and inject project overview context

[presstab]

Overview
This update strengthens the command‐routing logic by introducing a fallback “salvage” flow that reformats malformed LLM outputs, reduces duplicate parsing code, and integrates the project’s high-level overview directly into the system prompt. The deprecated get_project_summary tool has been removed in favor of on-the-fly context injection. Prompt guidelines have also been refined to clarify when to invoke the /code command.

What This Means for Users

• The agent will recover more gracefully from unexpected or malformed LLM responses, asking follow-up questions only when genuinely needed.
• The system automatically includes a concise project summary in its reasoning, improving overall relevance without manual file reads.
• Code generation boundaries are clearer: the router will no longer launch a full /code agent for simple questions or text-only tasks.

A Closer Look at the Changes

• .gitignore

  • Added a new entry for runs/ to prevent local run artifacts from being tracked.

• CommandInterpretationAgent (router_agent.py)

  • Inserted add_project_summary() during prompt construction to supply the agent with an “overview” file from the project state.
  • Centralized JSON parsing: moved message append and decision‐handling into a single flow after successful parsing.
  • On JSON decoding failures, triggers a salvage routine:
    • Builds a new prompt from router/salvage_response.md
    • Calls the “quick_reasoning” model to reformat the LLM output
    • Retries parsing and falls back to an error message if salvage also fails

• MessageBuilder (message_builder.py)

  • Introduced add_project_summary() method to include the overview file in outgoing system messages when available.

• agent_tools.py

  • Removed the legacy get_project_summary() function.
  • Expanded documentation for get_indexed_files_context() to emphasize its role in serving key file summaries.

• Prompt Updates

  • router/salvage_response.md: New template instructing how to reformat malformed JSON responses into the expected schema.
  • router/select_command.md: Refined rule on /code usage—clarifies that code agent calls should be reserved for actual code edits, not simple Q&A or planning tasks, and adds guidance to always confirm before invoking /code when in doubt.
    Generated by JrDev AI using o4-mini-2025-04-16