tau-coding-agent 0.7.4__tar.gz → 0.7.6__tar.gz

{tau_coding_agent-0.7.4/tau_coding_agent.egg-info → tau_coding_agent-0.7.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tau-coding-agent
-Version: 0.7.4
+Version: 0.7.6
 Summary: A self-extensible agent CLI with terminal UI, multi-provider LLM support, session management with branching, and a plugin system for tools and commands
 Requires-Python: >=3.12
 Description-Content-Type: text/markdown
@@ -32,6 +32,7 @@ Requires-Dist: arrow==1.4.0
 Requires-Dist: kiwisolver==1.5.0
 Requires-Dist: tiktoken==0.13.0
 Requires-Dist: rapidfuzz==3.14.3
+Requires-Dist: pyxclip==0.2.3
 Provides-Extra: tools
 Requires-Dist: ripgrep>=14.0.0; extra == "tools"
 Dynamic: license-file
@@ -108,6 +109,7 @@ tau --resume                             # Resume the latest session
 tau --resume abc123                      # Resume a specific session by ID
 tau --model claude-sonnet-4-6            # Start with a specific model
 tau --model groq/llama-3.3-70b-versatile # provider/model shorthand
+tau --base-url http://localhost:8000/v1 --provider vllm  # point at a local/proxy endpoint
 tau --print "Summarize this repository"  # Run once and print the result
 tau --mode json --prompt "Summarize this repo"  # Emit structured JSON events
 tau --mode rpc                           # Start JSON-RPC mode for IDE clients
@@ -123,6 +125,7 @@ Common flags:
 | `--mode` | | `interactive` (default), `print`, `json`, `rpc` |
 | `--provider` | | Provider to use, e.g. `anthropic`, `openai`, `groq` |
 | `--model` | | Model ID, or `provider/model` shorthand |
+| `--base-url URL` | | Temporarily override the provider's base URL for this run (not persisted) |
 | `--resume [ID]` | `-r` | Resume the most recent or a specified session |
 | `--fork ID` | | Fork a specified session at startup |
 | `--ephemeral` | `-e` | Don't save this session to disk |

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/README.md

@@ -70,6 +70,7 @@ tau --resume                             # Resume the latest session
 tau --resume abc123                      # Resume a specific session by ID
 tau --model claude-sonnet-4-6            # Start with a specific model
 tau --model groq/llama-3.3-70b-versatile # provider/model shorthand
+tau --base-url http://localhost:8000/v1 --provider vllm  # point at a local/proxy endpoint
 tau --print "Summarize this repository"  # Run once and print the result
 tau --mode json --prompt "Summarize this repo"  # Emit structured JSON events
 tau --mode rpc                           # Start JSON-RPC mode for IDE clients
@@ -85,6 +86,7 @@ Common flags:
 | `--mode` | | `interactive` (default), `print`, `json`, `rpc` |
 | `--provider` | | Provider to use, e.g. `anthropic`, `openai`, `groq` |
 | `--model` | | Model ID, or `provider/model` shorthand |
+| `--base-url URL` | | Temporarily override the provider's base URL for this run (not persisted) |
 | `--resume [ID]` | `-r` | Resume the most recent or a specified session |
 | `--fork ID` | | Fork a specified session at startup |
 | `--ephemeral` | `-e` | Don't save this session to disk |

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/docs/cli-reference.md

@@ -21,6 +21,7 @@ augment it.
 | `--quiet` | `-q` | Hide the non-interactive spinner |
 | `--provider` | | Provider to use, e.g. `anthropic`, `openai`, `groq` |
 | `--model` | | Model ID, or `provider/model` shorthand (e.g. `groq/llama-3.3-70b-versatile`) |
+| `--base-url URL` | | Temporarily override the resolved provider's base URL for this run only (not persisted) |
 | `--theme` | `-t` | UI theme: `dark`, `light`, or an installed custom theme |
 | `--resume [ID]` | `-r` | Resume the most recent or a specified session |
 | `--fork ID` | | Fork a specified session at startup |
@@ -347,6 +348,24 @@ tau --model anthropic/claude-sonnet-4-6
 
 An explicit `--provider` always overrides the inferred provider.
 
+## Base URL Override
+
+`--base-url` temporarily points the resolved provider at a different endpoint
+for the current run — a proxy, gateway, or self-hosted deployment. It applies
+to whichever provider ends up in use, whether set via `--provider`, inferred
+from `provider/model` shorthand, or picked up from the saved/default model —
+`--provider` is not required alongside it.
+
+```bash
+tau --base-url http://localhost:8000/v1 --provider vllm
+tau --model groq/llama-3.3-70b-versatile --base-url https://gateway.internal/v1
+tau --base-url https://proxy.example.com/v1   # applies to the saved/default model
+```
+
+The override is in-memory only for that run — it is never written to
+`settings.json` or `auth.json`, and there is no persistent equivalent; pass
+`--base-url` again on the next run.
+
 ## Session Options
 
 ```bash

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tau-coding-agent"
-version = "0.7.4"
+version = "0.7.6"
 description = "A self-extensible agent CLI with terminal UI, multi-provider LLM support, session management with branching, and a plugin system for tools and commands"
 readme = "README.md"
 requires-python = ">=3.12"
@@ -36,6 +36,7 @@ dependencies = [
     "kiwisolver==1.5.0",
     "tiktoken==0.13.0",
     "rapidfuzz==3.14.3",
+    "pyxclip==0.2.3",
 ]
 
 [project.scripts]

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/tau/agent/prompt/builder.py

@@ -146,6 +146,11 @@ _GENERAL_GUIDELINES = [
         "Prioritize accuracy over agreement — investigate before confirming, and "
         "disagree when the evidence calls for it."
     ),
+    (
+        "When asked a question, to explain something, or to review or report on status, "
+        "inspect the relevant files, code, or state before answering — give an "
+        "evidence-backed response, not one from assumption or memory."
+    ),
     (
         "Verify before reporting a task done: run the tests, build, or command that "
         "exercises the change. Don't present untested, partial, or stubbed-out work as complete."
@@ -158,6 +163,12 @@ _GENERAL_GUIDELINES = [
         "Treat git history as shared: never force-push, amend a commit that isn't yours, "
         "or skip hooks (--no-verify) without the user's explicit go-ahead."
     ),
+    (
+        "A new user message that arrives while you're still working may replace the "
+        "active request or add to it — judge from intent. If it replaces, drop the "
+        "prior work; if it adds, address both; if it's just a status question, answer "
+        "it and keep going."
+    )
 ]
 
 _PRECEDENCE_GUIDELINES = """\

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/tau/agent/service.py

@@ -75,6 +75,7 @@ class Agent:
         self._engine.options.before_tool_call = self._before_tool_call
         self._engine.options.after_tool_call = self._after_tool_call
         self._engine.options.transform_context = self._transform_context
+        self._engine.options.ephemeral_injection = self._ephemeral_injection
 
     # -------------------------------------------------------------------------
     # Public interface
@@ -270,6 +271,24 @@ class Agent:
         llm_messages = _to_llm_messages(session_ctx.messages)
         return strip_unusable_trailing_assistant(llm_messages, self._session_manager)
 
+    async def _ephemeral_injection(self) -> list[UserMessage]:
+        """Collect per-turn ephemeral messages from extensions via the "context" hook.
+
+        Called before every LLM inference (see Engine._run). Results are appended
+        to that single request's context only — never persisted to the session —
+        so extensions can keep the model up to date on live state (e.g. a todo
+        list) without that state needing to survive compaction.
+        """
+        from tau.hooks.engine import ContextEvent, ContextEventResult
+
+        session_ctx = self._session_manager.build_session_context()
+        results = await self.hooks.emit(ContextEvent(messages=list(session_ctx.messages)))
+        ephemeral: list[UserMessage] = []
+        for result in results:
+            if isinstance(result, ContextEventResult):
+                ephemeral.extend(result.ephemeral_messages)
+        return ephemeral
+
     # -------------------------------------------------------------------------
     # Internal helpers
     # -------------------------------------------------------------------------
@@ -674,6 +693,7 @@ class Agent:
             list(opts.images) if opts.images else None,
             list(opts.audio) if opts.audio else None,
             list(opts.video) if opts.video else None,
+            list(opts.file) if opts.file else None,
         )
         self._session_manager.append_message(user_message, meta=opts.meta)
 

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/tau/agent/types.py

@@ -46,6 +46,7 @@ class PromptOptions(BaseModel):
     images: list[bytes] = []
     audio: list[bytes] = []
     video: list[bytes] = []
+    file: list[bytes] = []
 
 
 @dataclass

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/tau/builtins/extensions/sandbox/sandbox_tool.py

@@ -45,7 +45,7 @@ def _render_result(content: str, opts: Any) -> list[str]:
     metadata = opts.metadata or {}
     exit_code = metadata.get("exit_code", 0)
     timed_out = metadata.get("timed_out", False)
-    failed = timed_out or exit_code != 0
+    failed = opts.is_error or timed_out or exit_code != 0
 
     if timed_out:
         return [f"{RED}Timed out{RESET}"]

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/tau/builtins/extensions/todo/__init__.py

@@ -6,9 +6,9 @@ rebuilt by replaying the session's custom entries (so it naturally follows
 branch/fork navigation), plus a `/todos` command for a quick look.
 
 The task list itself is surfaced through a board widget above the input box
-rather than the tool call/result in the transcript — the board only appears
-while there is at least one pending task, and disappears once the list is
-empty or fully done.
+rather than the tool call/result in the transcript — the board appears as
+soon as a task exists and stays up (showing done/failed tasks too) until the
+list is explicitly cleared or every task is deleted.
 """
 
 from __future__ import annotations
@@ -60,7 +60,7 @@ class TodoBoard:
 
     def _lines(self, theme: Any) -> list[str]:
         from tau.modes.interactive.components.message_list import apply_render_shell
-        from tau.tui.style import apply_style
+        from tau.tui.style import Style, apply_style
 
         glyphs = {
             "pending": ("☐", theme.muted),
@@ -73,16 +73,15 @@ class TodoBoard:
             glyph, style = glyphs[item.status]
             subject = item.subject
             if item.status == "done":
-                subject = apply_style(style.patch(theme.strikethrough), subject)
+                subject = apply_style(theme.muted.patch(Style().dim().strikethrough()), subject)
             items.append(f"{apply_style(style, glyph)} {item.id}. {subject}")
-        return apply_render_shell(items, theme)
+        return apply_render_shell(items, theme.message)
 
     def sync(self, ctx: ExtensionContext) -> None:
         ui = ctx.ui
         if ui is None:
             return
-        pending = self._state.remaining()
-        if not pending:
+        if not self._state.items:
             if self._shown:
                 ui.remove_widget(WIDGET_KEY)
                 self._shown = False
@@ -132,6 +131,30 @@ def register(tau: ExtensionAPI) -> None:
 
     tau.on("session_shutdown", _on_shutdown)
 
+    def _inject_todo_state(_event: Any, _ctx: ExtensionContext) -> Any:
+        """Re-assert the live todo list into context on every turn.
+
+        Old `todo` tool calls/results fall out of view once compaction (or
+        just enough turns) pushes them out of the kept window, and the
+        board is TUI-only — the LLM otherwise has no way to recover the
+        list. Injecting it fresh, ephemerally, each turn means it never
+        depends on surviving compaction or being noticed by the
+        summarizer: it's simply re-told every time, so it's never stale.
+        """
+        from tau.hooks.engine import ContextEventResult
+        from tau.message.types import UserMessage
+
+        remaining = state.remaining()
+        if not remaining:
+            return None
+        lines = "\n".join(f"- #{i.id} [{i.status}] {i.subject}" for i in remaining)
+        reminder = UserMessage.from_text(
+            f"[Todo list reminder — not from the user, injected each turn]\n{lines}"
+        )
+        return ContextEventResult(ephemeral_messages=[reminder])
+
+    tau.on("context", _inject_todo_state)
+
     tau.register_tool(TodoTool(state, tau._runtime_ref, on_mutate=board.sync))
 
     async def cmd_todos(ctx: ExtensionContext, _args: list[str]) -> None:

{tau_coding_agent-0.7.4 → tau_coding_agent-0.7.6}/tau/builtins/extensions/todo/todo_tool.py

@@ -305,7 +305,10 @@ class TodoTool(Tool):
                 "for one task (requires 'id'); 'delete' removes a task (requires 'id'); "
                 "'clear' removes every task. The list is shown to the user in a board above "
                 "the input, not in the transcript, so don't repeat task contents back to the "
-                "user after calling this tool. State persists across branches and restarts. "
+                "user after calling this tool. The board stays visible — even once every task "
+                "is 'done'/'failed' — until 'clear' is called, so once the whole plan is "
+                "complete, call 'clear' to dismiss it instead of leaving it up. State persists "
+                "across branches and restarts. "
                 "Never tell the user a task was created, updated, or completed unless you "
                 "actually made the matching 'create'/'update' call in this turn — calling "
                 "'list' alone does not change anything. If it's unclear which task or field "