v0.2.5

v0.2.5

New

• Copy button on assistant responses in the VS Code chat webview.
• Collapsible responses in the webview — each new session starts expanded.
• Token usage footer shows input, cached, fresh, and output token counts per response, compacted for readability (e.g. 818k, 1.2m).

Improved

• Patching accepts more diff dialects — bare context hunks, diff --git prologues, and Begin/End Patch wrappers — placing hunks by context when line numbers are absent.
• File edits recover from stale line ranges: when the target text matches uniquely elsewhere in the file, the edit is applied there instead of failing.
• Reading past the end of a file returns the tail with a note instead of an error, saving the model a corrective round trip.
• Shell output cap raised from 2k to 10k characters — roomy enough for test runs and builds.
• Default system prompt restructured and slimmed down.

Fixed

• Aborting mid-stream no longer loses the partial answer — streamed content and thinking are kept in both the CLI and VS Code.
• Reasoning-only responses are no longer treated as empty; reasoning stays on the thinking channel.
• Ollama context windows are now detected, so per-model context sizes are known.