Changelog
Kite 1.0.10, released 1 November 2026. 47 shell patterns are refused unconditionally in the version this page describes. This page summarises the release from the repository changelog; the full text, including every earlier release, is in the repo.
Kite 1.0.10: faster hot paths, session fork, and safer traces.
Added
- Bench and tracing. `kite bench` compares baselines, and `KITE_TRACE_JSONL` records runtime events. Isolated worktrees, fast gates, offline end-to-end QA, and CLI profiling are documented in the contributor workflows.
- Session fork. `kite sessions --fork <id>` starts a separate resumable conversation that inherits runtime settings.
- Subagent profiles. Unknown profile names suggest close matches, and profile details show the model and tool settings.
Changed
- Hot paths. Controlled runs (not portable timing guarantees) show faster streaming and reasoning accumulation, cheaper UI import and model construction, and quicker path completion, picker, directory sketch, and grep.
- Memory and resume. Redaction of large plain text and process-output peak memory dropped sharply. Session tail resume of a 50 MB row went from seconds to tens of milliseconds, and recent-event reads and semantic-memory rendering got cheaper.
- Install and prompt. A warm development install is faster. The system prompt uses bounded Kite reads and search on every platform and documents scoped-worker write limits. The bundled progress interval matches the 2-second runtime default. Named behavioral tests replace count-driven batching.
Fixed
- Prompt cache. Anthropic `cache_control` blocks go only to Anthropic-compatible routes. Cache-hit ratios no longer double-count normalized cache-creation tokens, and a failed LiteLLM prewarm import can retry.
- Git push. An ordinary `git push` reaches the serious-action approval flow. Force-push stays blocked.
- CLI and history. CLI and REPL rendering, completion, jobs, session recovery, verification, prompt-history correctness, and episodic-store lock handling were corrected. See kite_commands.md and architecture.md.
Security
- Output and traces. Output capture is bounded, redaction happens before truncation, trace files stay private, and durable writes are atomic. Details are in SECURITY.md.