Smarter, lower-context repository search. Coding agents can batch exact searches and receive ranked source snippets while Pica locally filters large noisy result sets.
Read-only Git analysis for coding agents. Agents can inspect project history, status, diffs, and commits while Pica keeps large outputs compact.
Fixed
More precise answers from large source files. Pica selects relevant excerpts from across a file and returns exact line-numbered source instead of unverified model prose.
More reliable code discovery. Improved indexing and retrieval make important definitions and implementations less likely to be missed in large projects.
A cleaner MCP tool list. Pica exposes only analysis tools supported by the active model and no longer advertises unavailable web tools.
Long-running analysis and indexing recover more reliably. Operations now recover more cleanly after cancellation, failure, or interruption.
Clearer guidance for coding agents using Pica. Project instructions now explain that Pica should be used to save agent context, start with broad orientation, and follow up with focused questions when needed.
Fixed
Code analysis answers are now less likely to be generic or incomplete. Pica no longer replaces failed AI synthesis with automatic project summaries or raw search results.
Better answers for broad project questions. Pica now reads more of the relevant files before finalizing, especially when search already found the right files but they were not inspected directly.
Cut-off answers recover more reliably. When a response is truncated after collecting evidence, Pica moves into the final synthesis path instead of continuing with a partial answer.
Evidence snippets render correctly for file paths with spaces. Answers now preserve full paths like folders with spaces and display the matching code block correctly.
MCP client initialization is more compatible. Clients that send initialized notifications without request IDs are now accepted without an invalid JSON-RPC response body.
Project instruction setup is less invasive. Pica now manages only the top-level agent instruction files and no longer creates hidden adapter folders or edits .gitignore.
Citations for files in folders with spaces in the name are no longer cut short. This could cause a correct answer to be wrongly rejected and re-tried, or in rare cases fall back to a lower-quality answer.
Added
Faster answers for broad questions. When the search results already thoroughly cover the relevant files, Pica can finalize the answer immediately instead of forcing extra reads.
Faster local generation on Apple Silicon. New optional speculative decoding toggle in Engine settings for the MLX engine — can significantly speed up generation depending on your model and hardware.
Visible reasoning on MLX, matching what you already see on llama.cpp.
Pica MCP instruction installer. Adding a project now writes Pica MCP setup instructions into it so other coding agents (Claude, Cursor, etc.) know to use Pica instead of their own raw search. New Fix Project action in the Projects list repairs this if it ever gets out of sync.
Fixed
MLX engine no longer crashes on startup with the default Gemma 4 E4B model.
More accurate code search results — search patterns no longer match unrelated substrings inside longer identifiers.
Faster, less repetitive analysis — avoids redundant work when the same search comes up twice.
Fixed some code questions about a single file getting stuck instead of returning an answer.
