Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Review the following changes in direct dependencies. Learn more about Socket for GitHub.
|
msullivan
force-pushed
the
tau-agent
branch
2 times, most recently
from
244f134 to
14ac685
Compare
The composer is no longer disabled mid-turn. Submissions go into a pending queue; run_turn is the sole consumer and loops until drained, popping one queued message per turn so user/assistant alternation stays clean.
All interaction with the `ai` library now lives in `chat_loop(app)` at the top of the file. TauApp owns public state (model, agent, messages, pending) that chat_loop reads. run_turn shrinks to a thin worker wrapper around the busy flag.
Streaming deltas only call scroll_end when the transcript was already at the bottom — scroll up to read earlier output and the stream keeps writing offscreen. The scrollbar itself is hidden (scrollbar-size: 0 0) because its per-chunk thumb motion was visually noisy; arrow keys, pageup/pagedown, and mouse wheel still scroll.
tools.py mirrors pi's built-in surface: read, write, edit, bash, grep, find, ls. Same schema shapes and continuation/truncation behavior: - read: 1-indexed offset/limit; head truncation at 2000 lines or 50KB with a 'use offset=N to continue' hint; first-line-too-big escape pointing at sed | head -c. - write/edit/bash: require_approval=True so the agent's default loop gates them behind a ToolApproval hook. - edit: exact-match, must-be-unique str_replace; multiple disjoint edits per call, applied right-to-left against the original file. - bash: tail truncation (errors live at the end), exit-code footer. - grep/find: skip .git/node_modules/etc; respect limit/byte caps. chat_loop now handles ToolEnd, ToolCallResult, and HookEvent — tool calls and their results render as 'tool' bubbles below the streaming text, and pending approvals turn the composer placeholder into a [y/n] prompt. Non-y/n input during an approval falls through to the message queue without resolving the hook.
Replaces the y/n-in-the-composer approval flow with a HookPrompt widget that mounts above the composer when a tool fires its approval hook. Yellow rounded border, shows the tool name + args; single-key shortcuts (y/a approve, n/d deny) resolve the hook. Focus shifts to the prompt automatically and returns to the composer on resolution. Tab cycles between prompt and composer if the user wants to look something up before deciding \u2014 the hook stays pending until y/n. Drops the parallel y/n branch from on_composer_submitted; the prompt owns the decision now.
Only bash still requires approval. File mutations through write/edit are fine — they're targeted, reversible, and the approval prompt became friction more than safety once both fired several times per turn.
… messages When TAU_ADVERTISE=1 is set, appends an instruction to the system prompt asking the model to include a Co-authored-by trailer in any commit messages it writes or suggests. Co-authored-by: anthropic/claude-sonnet-4.6, via tau
Co-authored-by: anthropic/claude-opus-4.6, via tau
Add a Static widget below the composer that displays running totals for input, output, cache-read, and cache-write tokens. Updated on each turn and when restoring a session. Co-authored-by: anthropic/claude-opus-4.6, via tau
Derive context estimate from the last assistant turn's input + output tokens and display it as 'ctx: ~N' in the footer bar. Co-authored-by: anthropic/claude-opus-4.6, via tau
Pass providerOptions.gateway.caching=auto as params to agent.run(). Update the usage footer to show uncached input as 'in' and cache-read tokens separately as 'cached'. Co-authored-by: anthropic/claude-opus-4.6, via tau
Co-authored-by: anthropic/claude-opus-4.6, via tau
Bash approval prompts now offer four options: [y] approve once, [n] deny, [!] always approve this exact command, [a] always approve all commands for the rest of the session. Co-authored-by: anthropic/claude-opus-4.6, via tau
Co-authored-by: anthropic/claude-opus-4.6, via tau
Session replay (iterating messages and creating bubbles) is now a free function alongside chat_loop, following the same pattern of taking the app and calling its rendering methods. Co-authored-by: anthropic/claude-opus-4.6, via tau
Co-authored-by: anthropic/claude-opus-4.6, via tau
…ember Replace the three separate methods (approve_command, approve_directory, approve_all) and the if-chain in on_hook_prompt_decided with a single remember(hook, decision) method on ApprovalTracker. Co-authored-by: anthropic/claude-opus-4.6, via tau
ApprovalTracker.resolve handles both updating approval state and calling ai.resolve_hook. Removes _resolve_hook from TauApp — the app just shows the system message based on the returned bool. Co-authored-by: anthropic/claude-opus-4.6, via tau
All UI code (HookPrompt, ApprovalTracker, TauApp hook plumbing) now works with our own Hook type. The ai.messages.HookPart is converted once in _run_turn via Hook.from_event. The only remaining reference to HookPart is in that constructor. Co-authored-by: anthropic/claude-opus-4.6, via tau
Co-authored-by: anthropic/claude-opus-4.6, via tau
ApprovalTracker.check now returns a list of PromptOption when the operator needs to decide, instead of bare None. HookPrompt renders whatever options it's given and uses _on_key for dynamic dispatch instead of static BINDINGS. The widget no longer knows about tool categories. Co-authored-by: anthropic/claude-opus-4.6, via tau
Put the arrow marker before the newline, not after it, so tool result bubbles don't start with a blank line. Co-authored-by: anthropic/claude-opus-4.6, via tau
…omment - Move _bell from TauApp staticmethod to module-level function. - Drop unnecessary async from on_text_area_changed. - Update tools.py docstring to explain why all tools use require_approval=True. Co-authored-by: anthropic/claude-opus-4.6, via tau
Handle ai.events.PartialToolCallResult in the agent loop to stream tool output (e.g. bash) into the transcript as it arrives, instead of waiting for the final ToolCallResult. Each tool_call_id gets its own bubble; final results are skipped for tools that were already streamed. Co-authored-by: anthropic/claude-opus-4.6, via tau
Prevents child processes from inheriting the terminal's stdin, which would conflict with Textual's input handling. Co-authored-by: anthropic/claude-opus-4.6, via tau
Iterate assistant message parts instead of just using msg.text, so ReasoningPart (thinking), ToolCallPart (tool invocations), and TextPart are all rendered into the transcript on resume. Co-authored-by: anthropic/claude-opus-4.6, via tau
Walk up from cwd looking for AGENTS.md and append its contents to the system prompt. Follows the convention used by other coding agents for project-level instructions. Co-authored-by: anthropic/claude-opus-4.6, via tau
Detect image files (jpg, png, gif, webp) via magic bytes and return them as base64-encoded image content parts instead of dumping binary as garbled text. Uses the Vercel AI SDK multi-part content format (text + image) so the gateway can pass images through to vision models. Co-authored-by: anthropic/claude-opus-4.6, via tau
The hand-rolled follow_scroll / at_bottom approach had a stale-state bug:
after an async gap (e.g. between stream end and error handler), Textual's
layout would update and at_bottom would become False even though the user
never scrolled.
Textual's built-in Widget.anchor() handles all of this correctly:
- New content auto-scrolls to the bottom
- User scrolls up → anchor releases, no forced scrolling
- User scrolls back to bottom → anchor restores
Removed: follow_scroll(), at_bottom property, auto_scroll parameter,
_following flag, and all manual scroll_end calls.
Added: single transcript.anchor() call at mount time.
Co-authored-by: anthropic/claude-opus-4.6, via tau
Co-authored-by: anthropic/claude-opus-4.6, via tau
Switch from Textual's default dark theme (which forces #121212 background and #1e1e1e surface) to ansi-dark, which uses the terminal's own background color. Removes explicit background overrides from composer and hook prompt widgets. Co-authored-by: anthropic/claude-opus-4.6, via tau
Tools can return [str, FilePart] (e.g. image reads) and all three providers (gateway, anthropic, openai) now convert them to proper image/document content blocks instead of stringifying them. Provider changes: - Gateway: _has_file_parts(), _file_part_to_v3_inline(), _multipart_to_v3_content(), _tool_result_output() route between json, content, and error-text output types - Anthropic: _tool_result_to_anthropic() expands FileParts into image/document blocks via existing _file_part_to_anthropic() - OpenAI: _tool_result_to_openai() converts images to image_url with data URLs for chat-completions FilePart serialization fixes: - Replace ser_json_bytes/val_json_bytes with a field_serializer that uses standard base-64 (+, /) instead of pydantic's URL-safe encoding (-, _) which LLM APIs reject - Add model_validator on ToolResultPart that rehydrates FilePart dicts back into FilePart instances after JSON round-trip (the result: Any field loses type info during deserialization) Co-authored-by: anthropic/claude-opus-4.6, via tau
Ctrl+J inserts a newline (works on all terminals). Trailing backslash before Enter also inserts a newline. Shift/Alt+Enter work on terminals with Kitty keyboard protocol support. Co-authored-by: anthropic/claude-opus-4.6, via tau
Detect the underlying provider from MODEL_ID and add the appropriate server-side web search tool: - anthropic -> anthropic.tools.web_search() - openai/xai -> openai.tools.web_search() Also adds Ctrl+J and backslash newline shortcuts to the composer, and fixes a docstring escape warning. Co-authored-by: anthropic/claude-opus-4.6, via tau
The model wasn't using web_search because the system prompt explicitly listed only the coding tools. Now it includes a mention of web_search when provider tools are configured. Co-authored-by: anthropic/claude-opus-4.6, via tau
Display BuiltinToolEnd and BuiltinToolResult events (from provider-executed tools like web_search) in the transcript alongside regular tool call/result output. Also updates system prompt on session resume so new tools and AGENTS.md changes take effect without a fresh session. Co-authored-by: anthropic/claude-opus-4.6, via tau
MODEL_ID includes the 'gateway:' prefix, which shouldn't appear in commit trailers. Use _raw_model instead to get the user-provided value. Co-authored-by: anthropic/claude-opus-4.6, via tau
- Import detect_image_media_type from ai.types.media instead of ai.messages.media (not an explicit export). - Rename shadowed variable to avoid ToolCallPart/BuiltinToolCallPart type conflict. - Guard against None tool_call_id in PartialToolCallResult. - Add mypy to dev dependencies. Co-authored-by: anthropic/claude-opus-4.6, via tau
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
tau, a crappy coding agent using Textual.
(Even though I don't really like alternate screen coding agents, but.)