Changelog

Release notes.

Every release is structured data in the repo — the same notes Tau shows in the TUI after an upgrade and publishes here.

v0.4.7latest release
33releases published
2026-10-01shipped
1

All releases

v0.4.7Latest

New

  • Archive sessions or whole projects from /resume with Ctrl+Enter without deleting files; use the Archived tab (F2) and Enter to restore them.

Fixed

  • Restore Option/Alt-arrow word navigation in the prompt inside Herdr, including Kitty Alt+b/Alt+f forwarding; Shift+Option/Alt-arrows select by word.
  • Create a fresh provider runtime when a new session selects a different provider instead of reusing the previous session's runtime.
  • Handle Delete correctly in the /resume search field.
v0.4.6

Changed

  • Preview model and thinking shortcuts immediately in the TUI and record only the final selection when sending a prompt; unsent choices remain session-only.
  • Open /model and /scoped-models from cached choices with a virtualized list, then refresh catalogs in the background after the picker renders.
  • Limit GitHub Actions token permissions per job and pin workflow actions to immutable commit SHAs.

Fixed

  • Reduce model-switching and picker latency by avoiding provider rebuilds and session-history writes on each shortcut, cancelling picker-owned refreshes on close, and skipping unnecessary sidebar updates.
  • Keep the active provider working when a refreshed catalog drops its model, and preserve cross-provider preview model and thinking capabilities.
v0.4.5

New

  • Edit prompt templates, context files, and skills from the TUI sidebar: click a resource row to open it in an in-app editor, save with Ctrl+S, and close with Escape.
  • Run `/system` to see where each system-prompt section came from, with prompt files, project instructions, skills, extensions, and runtime values grouped by origin.
  • Isolate credentials, sessions, preferences, and caches per profile with the `TAU_HOME` environment variable.

Changed

  • Show resume history as project columns with the current project first and other directories grouped separately.
  • Add an ask-user-question example extension that demonstrates interactive prompts from extensions.

Fixed

  • Order forked session trees by real history so alternate branches appear after the event where the fork started instead of collecting at the bottom of `/tree`.
  • Resolve `@file` and shell `!` path completions at the caret, including `@` mentions in the middle of a prompt or inside skill and template arguments.
v0.4.4

Changed

  • Open /resume immediately, load project history in the background, and virtualize session rows so large histories remain responsive.

Fixed

  • Restore mouse input in Herdr terminals that advertise pixel mouse support while forwarding cell coordinates.
v0.4.3

New

  • Resume sessions from other projects directly from /resume, with current-directory sessions first and other directories grouped separately.

Changed

  • Accept highlighted non-file autocomplete suggestions with Enter while preserving raw @ file references; Tab continues to accept every completion kind.
v0.4.2

New

  • Bookmark specific moments in the session tree with persistent per-entry labels, including create, edit, clear, filter, and timestamp controls in /tree.
  • Persist extension-injected model context as first-class custom_message entries and expose extension APIs for custom messages and session labels.
  • Discover and persist live OpenAI Codex subscription model catalogs so newly available models can be selected without waiting for a Tau release.
  • Show or hide the TUI sidebar for the current session without changing the saved global preference.

Changed

  • Align session storage more closely with Pi: stop writing redundant leaf pointers, store fixed-size first-kept compaction boundaries, and retain backward-compatible replay for older session files.
  • Include compaction and branch-summary model usage in session analytics, exports, and RPC inspection with provider and routed-provider attribution.
  • Make the prompt newline shortcut configurable and add a scoped-only tab to the scoped-model picker.
  • Expose resolved Tau paths to extensions through ExtensionContext.paths and align Anthropic usage accounting with the provider contract.

Fixed

  • Preserve interleaved Chat Completions reasoning and answer blocks instead of losing or misordering streamed content.
  • Reject or retry incomplete Anthropic-compatible streams rather than accepting empty SSE frames as successful responses.
  • Serialize Z.AI thinking requests correctly.
  • Disconnect shell-command standard input after execution and ignore selection events from detached TUI transcript widgets.
  • Highlight the initial provider correctly when the login picker opens.
v0.4.1

New

  • Set the startup thinking level with a new --thinking/-t CLI flag; it overrides remembered per-model defaults and resumed transcript state without being persisted, and validates against the selected model's supported levels.
  • See per-response performance in the TUI sidebar: average tokens per second and average time to first output, measured from provider stream timing and persisted on assistant messages.

Changed

  • Accept autocomplete completions with Tab only; Enter now always submits the prompt unchanged.
  • Cycle backward through scoped (favorite) models with Shift+Ctrl+P (remappable via model_cycle_reverse), complementing the existing Ctrl+P forward cycling.
v0.4.0

New

  • Run tool-capable GGUF models through Tau's built-in llama.cpp local backend, with the /local workflow for endpoint configuration, diagnostics, model discovery, downloads, and load or unload management.
  • Keep model metadata current from models.dev with a generated offline snapshot, background /model refreshes, tau update --models, richer capability metadata, and Pi-compatible reasoning levels including max.
  • Recover eligible Hugging Face pre-output route failures automatically, persist successful replacement routes, and expose structured failover diagnostics while preserving explicitly fixed routes.
  • Build dynamic providers from trusted extensions and let extensions add live sidebar sections, structured system-prompt sections, and react to session-name or thinking-level changes.
  • Show active SYSTEM.md and APPEND_SYSTEM.md sources in the TUI sidebar.

Changed

  • Render /system output as transcript Markdown without adding it to model context or session history.
  • Compose all discovered user and project APPEND_SYSTEM.md files from broad to specific before explicit CLI append content, deduplicating overlapping paths across reloads.

Fixed

  • Require explicit Google finish reasons, retry empty clean closes, preserve partial failures, and reject malformed stream chunks instead of treating incomplete streams as successful.
  • Prevent malformed syntax-highlighter spans from crashing the TUI when rendering fenced code blocks.
  • Update extension-owned sidebar content in place to prevent flicker and unnecessary widget remounts.
v0.3.13

New

  • Drive Tau from editors and other subprocess hosts through a Pi-compatible JSONL RPC mode, including prompting, state, models, sessions, compaction, direct shell commands, exports, and session naming.
  • Install trusted local or Git-hosted extensions with tau install, including pinned Git refs, staging validation, and overwrite protection.
  • Use Pi-compatible prompt template arguments such as $1, $@, $ARGUMENTS, defaults, and slices while retaining legacy Tau placeholders.

Changed

  • List Tau's utility commands in tau --help and add Discord community links to the documentation website and README.

Fixed

  • Send agent_settled reliably after completed, failed, and interrupted runs so extensions receive the final idle state after cleanup.
  • Keep TUI prompts readable when the terminal is too narrow for the normal prompt gutter.
v0.3.12

New

  • Keep user-triggered skills out of the model-facing system prompt with disable-model-invocation frontmatter while retaining explicit invocation, picker, autocomplete, reload, and sidebar support.

Changed

  • Refine sidebar skill and prompt headings with bold labels, muted metadata, and no hover or focus background highlights.
v0.3.11

New

  • Resume existing sessions in print mode with --session, restoring saved conversation history, working directory, provider, model, and Hugging Face route while allowing explicit provider and model overrides.
  • Show a startup transcript alert when a loaded skill or prompt template shadows a same-named resource from another location.

Changed

  • Compact adjacent TUI tool activity into selectable batches, group related reads, edits, and writes with their file paths, and keep exact commands and results available through Ctrl+O.
  • Group sidebar skills and prompts by resource origin and make both sections independently collapsible with loaded counts and an estimated skill-index token footprint.
  • Toggle inline tool-result visibility without showing a redundant notification.

Fixed

  • Preserve blank-line boundaries between streamed OpenAI reasoning-summary parts.
v0.3.10

New

  • Edit custom prompt templates directly from the prompts picker with Ctrl+E, then save and reload them with Ctrl+S.

Changed

  • Autocomplete explicit parent-relative file references while limiting external completion to the requested directory.
  • Rename the HTML session export Cache tab to Usage and annotate prompt-input charts with session events and tooltip details.
  • Move Hugging Face route controls out of Tau core and expose route inspection and selection through the public extension API.
v0.3.9

New

  • Autocomplete dot-prefixed project files and directories in file references and shell commands while continuing to hide generated metadata paths.
  • Keep successful /export destinations visible and copyable in the TUI transcript without adding them to model context or session history.
  • Add Kimi K3 to the Hugging Face Inference Providers catalog with image input, a 1M-token context window, pricing, and reasoning effort controls.

Changed

  • Improve Markdown heading and accent contrast in the built-in light theme.
  • Use model-specific output token limits for Anthropic-protocol requests, including explicit GitHub Copilot Claude limits.

Fixed

  • Persist interrupted tool results during cancellation so newly interrupted sessions retain valid provider history.
  • Repair malformed tool-call history from older sessions before provider replay so affected sessions can resume normally.
v0.3.8

New

  • Add a reactive cache analytics tab to live self-contained HTML session exports.
  • Pin a Hugging Face model to an explicit backing Inference Provider for the lifetime of a session.
  • Expose the resolved Hugging Face response provider in session metadata and diagnostics.

Changed

  • Update Kimi K3 reasoning effort levels to low, high, and max.
  • Add request timestamps to the cache tab tooltip and cache statistics table.
  • Bill Anthropic 1-hour cache writes at their own independent rate instead of blending with input token rates.

Fixed

  • Retry transient Anthropic stream errors instead of failing the entire turn.
  • Fix tool history reset across provider switches so the active provider sees the correct prior tool results.
v0.3.7

New

  • Decide project-input trust before Tau reads ambient project instructions, skills, prompts, themes, system prompts, or opted-in extensions; save exact/parent decisions or choose run-only scopes.
  • Control automation with run-only `--approve`/`--no-approve` flags and user-global `defaultProjectTrust` ask/always/never policy.
  • Pin a Hugging Face model to an explicit backing Inference Provider for the lifetime of a session.

Changed

  • Unresolved headless `ask` now skips protected project inputs; user/global and explicit CLI resources remain available, and project extensions still additionally require `--project-extensions`.
  • Project trust is explicitly an input-loading guard, not a filesystem, process, network, tool, model, credential, or prompt-injection sandbox.
  • Show latest-request and session prompt-cache hit rates in the sidebar, track OpenAI Responses cache writes, and improve OpenAI prompt-cache affinity.

Fixed

  • Exit Tau when the initial project-trust prompt is cancelled instead of continuing untrusted.
v0.3.6

New

  • Customize the generated system prompt in print or TUI sessions with `--system-prompt` and repeatable `--append-system-prompt` text-or-file inputs.
  • Load Tau-native `SYSTEM.md` and `APPEND_SYSTEM.md` files from user or project `.tau` directories, with CLI and project precedence plus `/reload` support.

Changed

  • Rebuild provider configuration from the current catalog during upgrades while preserving preferences and custom providers, so stale metadata cannot hide new capabilities or restore withdrawn models.
  • Include the current system prompt in live HTML session exports as a separate collapsed section; review exports before sharing because project instructions may be exposed, while JSONL and offline exports remain prompt-free.

Fixed

  • Anchor context estimates to provider-reported token usage, proactively compact large sessions, and present automatic overflow recovery without misleading terminal errors.
  • Keep `@` file-reference completion available in skill arguments and custom prompt text.
  • Prevent direct `!` and `!!` shell commands from cancelling an active agent run.
v0.3.5

New

  • Let print-mode automation choose a safe, durable session ID with `--session-id`, while preserving isolated-session guarantees.

Changed

  • Cache Anthropic prompt prefixes to reduce repeated input billing, with auth-aware retention, compatibility overrides, and a sidebar cache hit rate.
  • Redesign self-contained HTML session exports with compact accordions, entry filters and counts, expand/collapse controls, clearer tool labels, and offline JSONL download.
  • Show normal working indicators and unfocused completion notifications while manual `/compact` runs.

Fixed

  • Fix Anthropic subscription login and serialize OAuth token refreshes so rotating refresh tokens are not spent concurrently.
  • Accept file drops when the terminal is unfocused without duplicating focused paste events.
  • Safely hand off `tau update` on Windows after the running process exits, preserving executable paths and arguments exactly.
v0.3.4

New

  • Send images read from disk to vision-capable models across supported providers, while giving text-only models an explicit omission notice.

Changed

  • Validate and safely process image reads, including BMP-to-PNG conversion and bounded resizing for oversized static images.

Fixed

  • Allow multiline instructions after `/skill:<name>`, including instructions separated from the invocation by a blank line.
v0.3.3

New

  • Add direct Anthropic support for Claude Opus 5, including its 1M-token context window and adaptive-thinking effort mappings.

Changed

  • Highlight terminal-command prompt input, border, and prefix with the active theme's running-tool color.
v0.3.2

Fixed

  • Hide the unsupported `gpt-5.6` API alias from the OpenAI Codex subscription catalog while retaining it for direct OpenAI API users.
v0.3.1

New

  • Add one-line installers for macOS, Linux, and Windows that install Tau through uv and verify the resulting command.
  • Add configurable desktop or bell notifications when an agent run finishes while Tau is unfocused.
  • Add searchable TUI pickers for prompt templates and skills, plus a searchable tools reference with full descriptions.

Changed

  • Keep dense TUI sidebar sections compact by limiting rendered skills, context files, tools, prompts, and extensions while showing hidden-item counts.
  • Make user settings forward-compatible by ignoring unknown fields while continuing to validate recognized settings.

Fixed

  • Clear the prompt after cancelling the `/skills` picker instead of restoring the handled command.
v0.3.0

Changed

  • Breaking: mirror Pi's non-interactive CLI flags. `--resume` is now `--session`, `-p`/`--prompt <text>` is now `-p`/`--print` with a positional prompt, `-o`/`--output` is now `--mode`, and `-x` is now `-e` for `--extension`; the old flags now error with a migration hint instead of running.
  • Add a top-level `--export <session-id> [out]` flag alongside the existing `tau export` subcommand, and add `-v` as a short flag for `--version`, both matching Pi's flag names.
  • Merge piped stdin into the print-mode prompt (e.g. `cat file | tau -p "..."`), matching Pi's behavior.
v0.2.4

New

  • Add 28 live-routable Hugging Face Inference Providers models from Pi's built-in catalog.
  • Insert files dragged onto the terminal as normalized prompt paths while preserving existing text and quoting paths that contain spaces.

Changed

  • Shorten TUI sidebar context paths under the home directory with ~/ and give provider metadata quieter styling.
  • Raise the minimum supported Pydantic version to 2.11 for explicit wire-model alias validation and serialization behavior.

Fixed

  • Show expanded prompt-template text before automatic session naming starts, and avoid redundant naming attempts after provider retries are exhausted.
  • Fix the resume picker search to avoid matching workspace paths while retaining session-name and model matches.
v0.2.3

Changed

  • Default the TUI sidebar to the right side while preserving explicit left, right, and off preferences.
  • Label TUI sidebar token totals as cumulative session usage and document why cumulative provider usage can exceed the active context estimate.
  • Keep Tau's internal extension and model-catalog maintenance guidance in packaged self-documentation so it no longer appears in or competes with user skills.

Fixed

  • Fix session JSONL reads splitting records on Unicode line separators (U+2028, U+2029, U+0085), which made affected sessions and the session index fail to load; the read-side fix also repairs already-affected session files.
  • Fix picker highlight text contrast so focused labels use each theme's configured highlight foreground and background colors.
v0.2.2

New

  • Add data-driven JSON themes with built-in palettes plus user and project theme discovery.
  • Add live search to the resume session picker across session titles, models, and working directories.

Changed

  • Keep long TUI sessions responsive by mounting a bounded transcript window while preserving access to the complete conversation.
  • Redesign the TUI sidebar around session identity, lifetime activity, token usage, estimated cost, compaction status, and loaded capabilities.
v0.2.1

New

  • Persist thinking/reasoning blocks and provider replay metadata across Tau's Pi-compatible event streams, session storage, exports, and TUI rendering.
  • Bundle Tau self-knowledge docs, examples, and skills so installed copies can answer Tau-specific questions without a source checkout.

Changed

  • Discover Codex context and compaction limits at runtime from the authenticated model catalog, with conservative static fallbacks and a visible limit source in session diagnostics.
  • Remove the redundant animated tool-row spinner while keeping static status-colored markers and elapsed-time feedback.

Fixed

  • Fix provider-error recovery so a failed or aborted assistant turn no longer poisons the next prompt, while keeping the error visible in the TUI and session history.
  • Fix tool-call labels in the session tree for assistant turns without visible text.
  • Fix light theme code block background in fenced code blocks.
  • Coerce the startup thinking level to a mode the remembered default model supports, preventing crashes on incompatible saved preferences.
  • Fix tests launching the real Anthropic OAuth browser during test runs.
v0.2.0

New

  • Add Pi-compatible agent, provider, coding-session, and extension event streams with canonical lifecycle and settlement events.
  • Add OAuth login parity for Anthropic, GitHub Copilot, and OpenCode providers.
  • Add Kimi K3 model support.

Changed

  • Replace the legacy runtime message, event, tool, session, and extension protocol with strict Pi-compatible contracts.
  • Enrich extension turn events with turn indexes and timestamps.
  • Keep tau_agent independent of tau_ai by moving portable provider contracts into the agent package.

Fixed

  • Migrate legacy Tau v1 sessions, including null usage costs, at the persistence boundary.
  • Improve interactive OAuth login modal navigation.
v0.1.6

New

  • Add Python extensions for custom tools, slash commands, lifecycle hooks, dialogs, and message rendering.
  • Add Kimi K2.7 Code, Kimi Code subscription, and MiniMax-M3 model catalog entries.
  • Add tiered model pricing metadata for models whose rates change with input size.

Changed

  • Require Textual 8.2.8 or newer so CJK input methods work correctly.
  • Pass unknown slash-prefixed input through as prompts instead of rejecting it as an unknown command.
  • Retry Anthropic HTTP 425 and all 5xx responses consistently with other providers.

Fixed

  • Fix extension reload lifecycle ordering.
  • Ignore saved provider preferences when their providers are no longer configured.
  • Prevent missing or malformed bundled release notes from crashing startup.
v0.1.5

New

  • Add OpenAI GPT-5.6, GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna model catalog entries for OpenAI and OpenAI Codex.
  • Add NVIDIA NIM as a built-in provider.
  • Redesign session HTML export with a cleaner, icon-based interface.

Changed

  • Improve the documentation website with a roadmap page, clearer Pi links, and refreshed navigation and homepage polish.

Fixed

  • Fix release-note packaging so installed wheels include the bundled release notes file.
  • Fix Windows test isolation by avoiding accidental writes to a real ~/.tau directory.
  • Fix TUI theme selection edge cases.
v0.1.4

New

  • Add automatic session naming for new conversations.
  • Add terminal tab titles with the Tau mark and active session name.
  • Add Claude Fable 5 and Claude Sonnet 5 model catalog entries.
  • Add a TUI sidebar_position setting for left or right sidebar placement.

Changed

  • Unify skill discovery around the Agent Skills specification.
  • Read tau --version from installed package metadata.
  • Compact large pasted prompts in the TUI while preserving pasted content.

Fixed

  • Fix Windows crashes caused by non-ASCII CLI output on non-UTF-8 streams.
  • Fix Gemini tool-call compatibility by preserving thought signatures and stripping unsupported JSON Schema keywords.
  • Fix skill discovery edge cases by ignoring .agents root directories and bare Markdown files under .agents/skills.
  • Fix SOCKS proxy environment URL handling.
  • Fix queued steering recall, TUI completion height stability, pasted placeholder expansion, auto-naming indexing, and terminal-title refreshes.
v0.1.3

New

  • Add a config-driven Pi API provider catalog with user overlay support.

Changed

  • Optimistically render ordinary TUI prompt submissions to reduce perceived latency.
  • Improve scoped provider/model switching performance and responsiveness.
  • Migrate the documentation site to Hugo.

Fixed

  • Prevent custom prompt slash commands from appearing twice in the TUI transcript.
  • Persist interrupted tool repairs correctly when resuming sessions.
  • Show bad --model values as clean errors instead of tracebacks.
  • Fix provider/model selection and resume mismatch edge cases.
  • Hide code block scrollbars while assistant messages are streaming.
v0.1.2

New

  • Show one-time release highlights in the TUI transcript after Tau is upgraded.
  • Show the active session name in the TUI header.

Changed

  • Render assistant and thinking messages without role blocks for a cleaner transcript.
  • Apply role foreground colors to streaming transcript messages.
  • Clarify the PyPI release process in the developer notes.

Fixed

  • Fix autocomplete suggestion window shrinking.
  • Keep pre-branch context when branching with a summary.
  • Keep session state on the active branch when persisting messages.
v0.1.1

New

  • Check PyPI for Tau updates on startup.
  • Route OpenAI gpt-5.5, gpt-5.4, and codex models to the /v1/responses API.
  • Add the /system command for viewing the active system prompt.
  • Add prompt history recall with the Up key.

Changed

  • Lower the Python requirement to 3.12.
  • Render transcript messages as full-height role blocks.

Fixed

  • Restore streaming transcript scrollback.
  • Fix TUI code block horizontal scrolling.
  • Fix autocomplete suggestion viewport sizing.
  • Disable TUI text selection while streaming.
  • Guard Markdown selection until blocks are mounted.
  • Validate conflicting session flags in the CLI.
  • Stabilize CI checks.