What's New

The chat panel now operates across three explicit modes: Edit, Plan, and Knowledge. Edit routes directly to file modification. Plan generates a structured multi-file plan rendered as a reviewable artifact in the code panel. Knowledge answers questions about the codebase without making changes. Mode routing is automatic in Auto mode or can be set manually per request.

Edit and Plan requests now support an extended reasoning phase before execution. The model's internal reasoning is streamed live to the UI in a dedicated thinking block that expands in real time, then collapses once the response begins. This runs as a distinct phase in the pipeline and is separate from the final output stream.

When Plan mode generates a multi-file plan, the result is rendered as a structured artifact directly in the code panel alongside the file tree. A floating action bar provides approve and reject controls. Approving executes all file operations in the plan sequentially. A live per-file checklist updates in real time as each operation completes.

Auto mode now reads the active file, recent chat history, and the full content of the request before deciding whether to route to Edit or Plan. The previous implementation used keyword matching. The new implementation passes context to the model and uses the model's own classification, eliminating the class of misroutes caused by ambiguous phrasing.

The thinking display was redesigned. Reasoning stages now render outside the chat bubble as a sequential reveal with typing effects. A roadmap timeline view shows each stage as a discrete step with its own label and status indicator. This applies to both the Edit ReAct loop and the Plan generation phase.

The AI backend was migrated from the DeepSeek native API to OpenRouter. This enables runtime provider switching and removes the hard dependency on a single inference endpoint. The DeepSeek R1 reasoner model ID was updated to the correct OpenRouter path (deepseek/deepseek-r1).

The admin panel now includes a dedicated viewer for generated file outputs. Each generation record links to a full file tree preview showing every file produced in that generation pass, alongside the existing tabular generation log.

AbortController and AbortSignal timeout wrappers on DeepSeek API calls were triggering premature stream termination mid-response. The timeouts were removed. Stream lifetime is now managed by the Vercel Edge runtime and the upstream provider connection rather than a client-side timer.

The previous intent classifier was producing incorrect Edit/Plan splits on requests with mixed phrasing. The classifier was rewritten to pass full context to the model rather than using surface-level keyword heuristics.

The SSE chunk processing loop was blocking the React render thread between token deliveries, causing the UI to batch updates rather than reflect each token. A yield point was added between chunk processing cycles to allow React to commit intermediate states.

The retry path on failed modifications was requesting up to 32k tokens, which exceeded the Vercel function timeout budget on complex files. The retry token cap was reduced to 16k. The modify route maxDuration was also made explicit to guarantee completion on large files.