v0.9.0

Kana v0.9.0 adds profile-based asynchronous subagents, configurable built-in Agent tools, and dark, light, and custom TUI themes. It also makes the reusable Kana Agent workflow portable across repositories and providers.

Features

• Delegate bounded work to asynchronous subagents selected from built-in or user-defined profiles, with independent transcripts, approvals, accounting, completion notifications, and /agents management.
• Configure the built-in tool surface exposed to the main Agent.
• Select the built-in dark or light TUI theme, or load a custom semantic color theme.
• Cancel MCP startup and reload from the TUI without exiting Kana.
• Run the reusable Kana Agent GitHub workflow across repositories and supported providers with repository-local configuration.

Bug Fixes

• Frame streamed terminal input consistently with submitted user messages.
• Preserve visible HTML/XML-like tags in Markdown instead of stripping them.
• Correct option handling in the Kana Agent workflow.

v0.8.0

Kana v0.8.0 adds session-owned background jobs, durable session todos, and bounded agent goal execution, and unifies model selection under a new static configuration format.

Features

• Run session-owned background jobs alongside the agent and manage them from the TUI while a run is in progress, with per-session lifetime and full output retrieval.
• Persist session todos durably so checklists survive restarts and session resumes.
• Bound agent goals with explicit goal rounds and status reporting, including headless goal execution for kana exec and opt-in adoption in the Kana GitHub workflow and Harbor Terminal-Bench adapter.
• Run safe slash commands while the Agent is working instead of waiting for the turn to end.
• Configure the agent model statically under agent.model and give the memory agent its own model under memory.agent.model.
• Cap model-facing tool content at 8k tokens by default.

Bug Fixes

• Retry transient Responses stream failures instead of aborting the run.
• Preserve the calibrated context estimate across session resume.
• Remove the 20 KiB page limit from job_output.
• Route escape through focused TUI components and deny tool approval on escape.
• Reject unknown arguments for Kana-owned tools, tolerate Kana handoff preambles, include the source issue in Kana follow-ups, and improve Kana agent run reporting.
• Accept long goal objectives.

Breaking Changes

• The configuration format changed. The top-level model table and provider.active are replaced by agent.model and memory.agent.model; webSearch and imageInput moved from per-model settings to the agent runtime section; maxTokens is renamed maxOutputTokens.
• Old keys are not migrated and do not fail parsing: an upgraded config silently falls back to the default model until it is updated.

Upgrade

If you customized model.* or provider.active in ~/.kana/config.json, run kana reset to regenerate the template (back up the old file first) and reapply your settings in the new format. Verify agent.model points at the intended provider after upgrading; a stale config starts on deepseek with defaults.

v0.7.0

Kana v0.7.0 adds full-fidelity tool inspection and history, maintainer-triggered GitHub issue automation, and reliable time-bounded headless execution.

Features

• Inspect every tool call with Ctrl+O, navigate adjacent calls with [ and ], and review complete commands, paths, arguments, diffs, results, and runtime status without expanding the compact transcript.
• Browse the current session's tool history with /tools and open any selected call in the shared full-detail inspector.
• Trigger scoped Kana work from maintainer-authored GitHub issues and continue it on Kana-owned pull requests, with draft-PR publication, structured handoffs, diagnostics, and preserved partial progress.
• Bound a complete kana exec Agent run with --timeout, graceful cancellation, exit status 124, and structured JSONL termination metadata; prompts now work with both stdin pipes and regular-file redirection.

Bug Fixes

• Preserve complete material values in tool approvals and inspection, including empty write content and edit replacements, while showing the same effective defaults used at execution time.
• Sanitize tool targets before terminal rendering and keep prepared-tool feedback stable while an approval prompt is open.
• Run GitHub automation from the latest verified stable Kana release so a broken target branch cannot prevent the Agent from repairing it, and report structured failures without publishing incomplete error output as a final summary.

v0.6.0

Kana v0.6.0 adds agent image inspection, durable artifacts for oversized tool results, and tool-result policies with repeated tool-call detection.

Features

• Let the agent inspect local images with the new view_image tool and carry native visual observations into model requests.
• Persist oversized tool results as on-disk artifacts with a bounded model-facing preview, so complete bash output is never lost.
• Add tool-result policies with built-in repeated tool-call detection so model loops stop before they burn context.

Bug Fixes

• Preserve complete bash output for artifact storage and harden artifact fallback and replay.
• Bound compact tool transcript rendering to a fixed row budget and stop guessing targets for unknown tools.
• Replace stale runtime context in model input by assembling prompt context per model step.
• Align parallel tool completion and cancellation semantics during bounded scheduling.
• Harden the tool-result policy boundary against invalid policy returns.
• Validate visual tool results, keep the MCP loading transcript append-only across reloads, and capitalize status labels in the TUI.

v0.5.0

Kana v0.5.0 adds custom OpenAI-compatible providers, image prompts for DeepSeek V4 Flash Vision, and terminal-native Mermaid and LaTeX rendering.

Features

• Connect any OpenAI-compatible endpoint with static custom model definitions over HTTP or HTTPS, including streamed thinking events and unified TUI working state for compatible models.
• Attach images to DeepSeek V4 Flash Vision Exp prompts, gated by model metadata and the image_input option so text-only V4 models stay unchanged.
• Render Mermaid diagrams and LaTeX formulas directly in the terminal transcript.
• Align editor shortcuts with readline for consistent text navigation.
• Remove per-request cost estimates from /usage; token usage summaries remain.

Bug Fixes

• Detect the Ghostty notification backend instead of silently falling back to a generic terminal notification.
• Dismiss the usage-scope prompt when the selection fails, and align /usage summary columns.
• Reject duplicate pending message identities and serialize inbox claims during shutdown.
• Pin the Bun toolchain version so local tooling, CI, and release builds stay aligned.

v0.4.0

Kana v0.4.0 adds image prompts for OpenAI Codex, hosted web search for DeepSeek V4, and clearer TUI feedback for large inputs and tool execution.

Features

• Attach up to 10 images to OpenAI Codex prompts from the macOS clipboard or local paths, with safe normalization, session persistence, and context compaction.
• Run DeepSeek V4 Flash and Pro through the Responses API with hosted web search and low, high, or max reasoning effort.
• Collapse long pasted input into a compact editor item while preserving the complete text for submission and history.
• Group streamed tool-call preparation into one progress block, then show running, completed, canceled, and failed states at their actual lifecycle boundaries.
• Expand /help with the current editor, image, cancellation, tool-output, and local-shell shortcuts.

Bug Fixes

• Keep Codex tools and image inputs on the compatible classic Responses contract and surface provider stream error details.
• Distinguish non-zero Bash exits from execution failures, and stop tool or hosted-search timers cleanly after cancellation.
• Estimate context usage from replayable conversation state and preserve image-bearing history safely during compaction.
• Keep prompt placeholders inside the editor frame and preserve one thinking timer across adjacent reasoning items.
• Stop background memory consolidation cleanly during shutdown.

v0.3.0

Kana v0.3.0 adds queued input and scheduled-message management, Codex hosted web search, and clickable terminal hyperlinks.

Features

• Add OpenAI Codex hosted web search with visible provider activity rendering.
• Add /schedule to view, create, refresh, and delete in-process scheduled messages.
• Queue input during Agent runs and deliver queued submissions and due wakes in FIFO order.
• Render safe Markdown links as terminal hyperlinks when supported, with readable fallbacks.

Bug Fixes

• Fix queued-input delivery and duplicate display after deferred steering.

v0.2.2

Kana v0.2.2 makes kana update more reliable by streaming asset downloads to disk instead of buffering them in memory.

Bug Fixes

• Stream self-update downloads to disk so large releases no longer stall on slow or proxied responses.

v0.2.1

Kana v0.2.1 sharpens session control and reliability with a new clean launch mode, per-session approval overrides, and smarter context handling.

Highlights

Clean launch mode

• Launch with --clean for a temporary session that touches nothing — no session files, no memory, discarded on exit.
• /skills, /mcp, /memory stay discoverable in clean mode but clearly report unavailability instead of failing silently.

Per-session tool approval

• /approval switches the current session between Always ask, Ask unless trusted, and Never ask (with a confirm step).
• The override applies only to the current session and resets on new/fork/resume or exit — nothing written to config or journals.

Smarter context and tool behavior

• Parallel tool calls are gated by model capability instead of being allowed blindly.
• Tool output is dynamically capped by remaining context to protect the budget near the limit.
• Reasoning effort shows in the TUI status line; unsupported Codex ultra effort removed.
• Bash timeout ceiling raised; buffered TUI text renders before tool results.

Also included

• Harbor evaluation adapter with proxied evaluation support.
• Clean-session and rendering reliability fixes.

v0.2.0

Kana v0.2.0 is the largest Kana release so far. It adds a second model provider, a complete MCP stack, and a shared runtime for interactive and automated work.

Highlights

OpenAI Codex and live model switching

• Authenticate with kana auth login openai-codex.
• Switch provider, model, and reasoning effort from /model without losing the current conversation.
• DeepSeek remains the default, and legacy [model] configuration stays compatible.

MCP from local processes to authenticated HTTP

• Connect stdio servers and MCP 2025-11-25 Streamable HTTP servers.
• Manage servers from /mcp, with OAuth 2.0/PKCE, per-server proxies, scoped authorization, and bounded tool results.
• Optional servers fail independently; required servers fail closed.

Headless agent execution

• Run complete agent turns with kana exec, including stdin prompts and session resume.
• Use --json for a versioned JSONL event stream suitable for automation.
• --allow-all-tools authorizes every available tool without creating file-system or process isolation. Use it only in controlled environments.

Also included

• Automatic and manual context compaction with durable checkpoints.
• Incremental V3 session journals and interrupted-run recovery.
• Parallel tool execution with deadlines, cancellation, and hardened lifecycle handling.
• New list, glob, and grep tools, plus explicit file overwrite support.
• A redesigned TUI command flow, improved bottom layout, and interactive memory management.
• Skills installation, synchronization, and recovery improvements.
• Built-in update checks and kana update for future upgrades.
• A reorganized internal architecture and fully rewritten Chinese and English READMEs.

Breaking changes

• Kana now reads V3 sessions only. Existing V1/V2 session files are preserved but no longer appear or load in v0.2.0. Keep an older binary if you still need to access them.
• kana install --skills is now kana skills install.
• kana install --force is now kana reset; use kana skills reinstall when only bundled Skills need to be restored.
• /memory show and /memory compact are replaced by the interactive /memory flow.

Upgrade

Kana v0.1.3 does not include the self-updater, so upgrade once with the install script:

curl -fsSL https://raw.githubusercontent.com/longyijdos/kana/main/install.sh | sh

After installing v0.2.0, future releases can be installed with kana update.

Full changelog: v0.1.3...v0.2.0