Repository navigation
Releases
Releases · larsch/xi-agent
v0.9.1
v0.9.0
Added
- Steering and draft navigation: steering messages and draft prompts now share unified navigation.
- F3 context file inspector: inspect files included in the current context.
- Shift+Tab @-file completion: complete @-file paths without injecting them into the prompt.
- Session IPC guidance: skill guidance is included conditionally for sessions using IPC.
- Structured environment notifications: environment changes are surfaced as structured notifications.
- Project rule files: load project-specific rule files.
- Agent selection: choose an agent with a CLI option.
- System prompt context reader: inspect system prompt context from the UI.
- Generic tool presentation: UI shows generic tool names and parameters.
Changed
- Local work timing: the UI displays elapsed runtime for local work.
- Skill loading: the
learntool includes the skill directory, discovers skills in ancestor and home directories, and scopes skills to agent paths. - Session preferences: agent preferences are persisted per session.
- Default agent guidance: tool-writing guidance is platform-neutral and the default prompt is streamlined.
- Session IPC: IPC uses shared runtime sockets, avoids duplicate agent runners during handoff, and improves Codex summaries.
- Provider behavior: print mode defaults to the configured provider; Copilot GPT models use the Responses API; Codex models can be discovered.
- Prompt and input handling: system prompt sections are delimited, pending steering is recalled, and multiline prompts are preserved in history navigation.
Fixed
- Ask-user cancellation: cancellation is honored while awaiting input.
- File writes: successful file writes are trusted.
- File search cancellation:
find_filesresponds cooperatively and promptly to cancellation. - Windows subprocess cancellation and IPC warnings.
- Windows CI: unused test-variable warnings are silenced.
- Model refresh: avoid a retry loop during model refresh.
- Focused-terminal notifications: suppress notifications while the terminal is focused.
Internal
- Agent loop: modularize the agent loop and runner lifecycle.
- Documentation: document the IPC worker skill and clarify tool and skill guidance.
v0.8.0
Added
- Persistent Python REPL: an agent-loop-scoped Python kernel preserves state across executions, supports reset and asynchronous execution, handles timeouts, and can install dependencies through uv when available.
- Desktop notifications: interactive sessions can show best-effort notifications when the agent halts, including the current project and a bounded preview of the question or response.
- Input text selection: chat and shell input fields support selecting and highlighting text, including selections spanning wrapped lines.
- Experimental session IPC: opt-in local IPC provides session control and event subscriptions through a worktree-local socket; this interface is experimental.
Changed
- Windows input: bracketed paste and mouse input now use the updated Crossterm support while retaining compatibility with Windows Terminal and conhost.
- Viewport anchoring: anchor padding is preserved through agent-loop continuations, layout changes, scrolling, and terminal resizing.
- Ask-user presentation: ask-user context now uses the same margin block styling as the question.
Fixed
read_filerange notices: exact offset/limit reads no longer show redundant range notices, while truncation and EOF-clipped ranges continue to be reported.- Unix subprocess tests: the Python REPL context is initialized correctly in Unix-only test helpers.
- Session switching: switching sessions now clears stale log expansion, padding, generation, and scroll state from the previously selected session.
v0.7.0
Added
- Restart host tool: the
restart_hosttool can re-exec the running binary after a rebuild and resume the active session when enabled with therestartfeature. - Resume option:
--resumenow automatically resumes the latest session in the working directory. - Streaming test commands:
test-providernow supportsstream-thinkingandstream-linescommands.
Changed
- Model activity lifecycle: model turns now have explicit lifecycle handling in the UI.
- Model throbber timing: the model activity indicator adapts its timeout to the stream cadence.
Fixed
- Terminal suspend: the default SIGTSTP handler is restored before suspending, so Ctrl-Z works reliably.
- Markdown in lists: fenced code blocks now render correctly inside list items.
- Log markers: log margin markers are rendered consistently.
Performance
- Startup: agent context loading now happens in the background.
- Resume checks: checking for a resumable session avoids reading session files unnecessarily.
v0.6.0
v0.6.0
Added
- Parallel tool execution: independent tool calls in a single assistant batch now run concurrently, with results interleaved back to the model in their original intent order.
- Mouse unfold: click a truncated tool block to expand it in place.
- Braille waiting animations: the activity throbber now animates with Braille characters.
- Model vs. local activity throbbers: the UI distinguishes work done by the model from work done by local tool execution.
- Auto-submit initial prompt: an initial prompt passed on the command line is now submitted automatically in interactive mode.
- DeepSeek thinking mode: the OpenAI-compatible provider supports DeepSeek reasoning/thinking output.
- Refreshed theme palette: the default theme palette and demo were updated.
- Graceful shutdown: SIGINT/SIGTERM handling via
tokio::signalfor clean shutdown.
Fixed
- High CPU load: large streaming tool output no longer causes 100% CPU usage.
edit_filediff collapse: a new line that is a prefix extension of the old line no longer silently collapses the diff.- Control characters: literal control characters are sanitized from accumulated tool-call JSON.
- FileTracker on resume: file tracking state is seeded from session events on load/resume, so change detection works across restarts.
- Compaction: plain
compactnow triggers compaction. - Diff placeholders: common-lines placeholders are suppressed in pure addition/removal hunks.
- Keybindings: ToggleFullOutput moved from Ctrl+F to Alt+O; Ctrl+D now acts as delete-char when the input is non-empty.
- Windows build: warnings fixed.
- UI layout: throbber scrolling, anchor padding, and fold/unfold baseline reset now behave consistently during streaming.
Performance
- Virtualized log rendering: only the visible window is materialized.
- Memoized log blocks: streaming re-renders only changed messages.
- Structural hashing: role/phase/JSON hashed structurally to avoid serde allocations.
Arc<str>block identity and borrowed-slice rendering reduce allocations in the hot render path.
Internal
- Terminal backend type centralized.
- Logical log block layout introduced; continuation and panel layout state centralized.
- Tool rendering kept in a single visual block.
- UI render benchmark harness and committed-history streaming benchmark added.
test-providergained aparallel-streamcommand.
v0.5.0
Added
- Ctrl-Z suspend/resume: idle-only Ctrl‑Z suspend and foreground resume support. The TUI is recreated on resume to avoid cursor query races (#4, #5, #6).
- Tab in ask_user: pressing Tab in the ask_user prompt copies the highlighted option into the input field for editing before submitting.
Fixed
- Streaming edit diff stabilisation: prefix-matching and symmetric common-lines computation produce fewer flickering diffs during streaming
edit_fileresults. - Throbber gaps: the bottom gap now shrinks instead of shifting content when the throbber appears, and the throbber is hidden only when tool output visually grows.
- Streaming tool call labels: streaming tool call labels and results use
┆(not╰) for a cleaner flowing look. - Newlines in user messages: newlines are now correctly preserved when rendering user messages in the log.
- Provider menu login: the login option in
/providermenu no longer does nothing. - Provider setup prompt: removed misleading "leave empty to keep current" hint from the API key prompt.
- Ollama context window: context window is now discovered via
/api/psinstead of GGUF metadata; model tags are normalised in the cache. --modelCLI override: the--modelflag is now correctly applied to the provider instance on startup.- Step-back ask_user: the ask_user question text is preserved when stepping back and re-answering.
- Windows build:
fmt,clippy, andchecknow pass with zero warnings on Windows; cross-compile build scripts fixed.
Internal
main.rsdecomposed into modules and handler functions.- Terminal lifecycle extracted into its own module with a
recreate_terminalhelper. - Dimmed hint text removed from provider setup input.
- Python program test command added.
v0.4.0
Added
- Prompt cache support: OpenAI and Anthropic providers now report cached-token counts in the info bar. Anthropic sends automatic ephemeral
cache_controlannotations on every request (matching API spec), OpenAI sends a stable session ID to improve cache routing. Unexpected cache misses (zero cached tokens despite a recent cache-populating turn) surface a⚠️ suffix in the info bar context display. - Session resume context usage: the info bar now shows the last known token utilisation immediately when resuming a session, instead of showing only the context window size until the next turn completes.
- DeepSeek V4 context window: hard-coded 1M token fallback entries for
deepseek-v4-flashanddeepseek-v4-pro(kludge until upstream metadata is available). - Alt+S shortcut: toggle the info bar on and off without
/info. - Theme configuration system: themeable UI with CSS-style color specs and terminal color support via a TOML configuration file.
- Agent-level hooks system: pluggable hook points for tool lifecycle (
OnToolIntent), streaming milestones (OnFirstThinkingToken,OnFirstTextToken), and session events (OnIdle,OnCompacting,OnExternalChange,OnStatusUpdate), with IPC event streaming for external tooling. - Alt+C shortcut: copy the last assistant response to the system clipboard.
- Step-back navigation: step back through ask_user answers with full prompt UI restoration, enabling re-answering or branching from any decision point.
- Mouse text selection: click-drag to select and copy text from the log view.
- Similar session scope: a "similar" scope between
localandforeignin the session resume picker for faster filtering. /newresets FileTracker and reloads skills, ensuring a clean environment for each fresh session.@filebacktick notation:@filenamementions are rewritten to backtick-quoted paths for LLM consumption, reducing confusion.read_skillandedit_skilltools: embedded tools for listing, loading, and editing skills, with filesystem paths and scope indicators.- User message markdown rendering: user prompts are now rendered with markdown formatting in the log view, matching assistant output.
- Keyboard shortcuts help: accessible via
?key. - Block content alignment: all block content consistently aligned to column 3 with margin markers.
- Edit diff fillers: adjacent "total lines" and "common lines" fillers in
edit_filediffs are collapsed to reduce noise. - Provider synthesis: built-in provider instances synthesised automatically; explicit provider selection required for ambiguous configurations.
- OSC 52 clipboard: clipboard integration via OSC 52 escape sequences, replacing the
arboardcrate for broader terminal compatibility. - Python 🐍 emoji: the built-in Python tool now shows a Python emoji icon.
Fixed
- Steering race condition: pressing Enter during streaming now defers the user message until the current assistant turn and all its tool calls are committed, preventing transcript corruption and tool-call skipping. Explicit cancellation still interrupts immediately.
- Attachment ordering: synthesized
read_fileevents from@filenameattachments are now placed after the submitted user prompt in the event stream, matching provider expectations. --printmodel override: the--modelflag now correctly overrides the configured model in non-interactive mode.@filemissing-file handling: references to nonexistent files are now silently ignored (no synthetic tool call, no error notice, no provider error). The@filetext remains in the prompt unchanged.- Anthropic
cache_controlplacement: removed the invalid top-levelcache_controlfield and moved it to individual content blocks, fixing 400 errors from the Copilot Anthropic proxy and restoring prompt caching on both direct and proxied routes. - Info bar token persistence: previous-turn token usage (input size, cached size) remains visible when starting a new prompt, instead of clearing at turn launch. Still resets correctly on
/new. - Codex prompt cache hits: parse
input_tokens_details.cached_tokensfrom the OpenAI Responses API in the Codex provider, so cache-hit indicators appear for Copilot GPT-5.x models. - Copilot auth: full token (with metadata) is now stripped correctly before use as Bearer auth, fixing authentication failures.
- ask_user rendering: questions now stream from partial data during tool call rendering and appear in the log; rendering layout improved.
- Provider switching: changing providers no longer uses a stale model list for fetching; model list auto-fetches on startup even when a model is already configured.
- Config load failures are now fatal, preventing silent overwrite of user configuration with defaults.
- Anthropic null tool_args: providers now guard against null tool arguments in Anthropic wire format.
- Auth token expiry: standardized across backends with a new
OAuthBackendtrait and test infrastructure. - Serde error messages: Rust struct names in serde errors are now translated to model-friendly JSON concepts.
- Tool descriptions: reworded to prevent redundant
2>&1and absolute-path annotations in shell commands. - System prompt clarifies that file paths are relative to the working directory.
- Input panel: scroll-to-cursor behavior added when text exceeds the viewport.
- Tool invocation labels: leading and trailing empty lines trimmed; placeholder labels shown consistently during streaming; tool icons render normally even when labels are italic.
- Output trimming: leading and trailing empty wrapped lines removed from output blocks; body line limit enforced on wrapped visual lines, not logical lines.
- Thinking display: final streaming line uses ┆ instead of ╰; blank separator line removed between thinking and response; display stabilized by truncating wrapped lines.
- Throbber: no longer sticks or blocks Escape during token refresh; remains visible during retry; refreshes correctly on tool intent, args delta, and output chunks.
- Streaming blocks: padded during shrink to prevent layout jitter; partial-JSON headline blink eliminated.
- Markdown: extra blank line after pre-formatted text removed; HTML/XML tags now rendered verbatim instead of silently dropped.
- OpenAI reasoning content: always included in assistant wire format; OpenAI-specific parameters guarded by backend type.
- OpenWebUI context window: auto-discovered for OpenAI-compatible backends.
- PowerShell now runs noninteractively.
- Windows build fixed and tests made cross-platform.
- Log redraw now happens before disk I/O on submit for lower perceived latency.
v0.3.0
Added
@filenameattachments: type@in the chat input to get interactive
file completion and attach file contents directly to your message.- Live subprocess output: bash/exec tool output now streams into the UI in
real time instead of appearing only after the command finishes. - Auto model picker: when a provider has no model configured xi automatically
opens the model picker so you can choose one without extra navigation. - Action verb placeholders: while a tool call is still streaming in, the UI
shows a meaningful verb (e.g. "Reading…", "Editing…") instead of a blank line. - Unified truncation indicators: long tool outputs use consistent dimmed
italic placeholder markers to signal hidden content.
Fixed
edit_filenow returns an error ifold_textandnew_textare identical,
preventing silent no-ops.- Tool call pending labels appear immediately before any argument JSON arrives.
ask_userquestion blocks render on the default background, consistent with
agent response blocks; answers appear as normal user message blocks.edit_filediff truncation markers are now coloured by side (add/remove).- Common-line diff placeholders are omitted for pure-addition or pure-removal
hunks, reducing noise.