tau-by-clarity 0.55.6

Python port of pi-mono: coding agent with multi-provider LLM support, TUI, agent loop, and tools

Tau by Clarity

A multi-provider AI coding agent and agent-building framework for Python — an interactive TUI, a headless CLI (tau), an agent loop, file tools, and built-in Tau by Clarity PII tokenization and active context compression.

Run it (tau), embed it (import pi_coding_agent), or build your own agents on it (see Building agents). Tau builds on the PI project — see Credits & lineage.

---

Installation

Install the tau CLI from PyPI (tau-by-clarity):

uv tool install tau-by-clarity      # recommended — puts `tau` on your PATH
# or:  pipx install tau-by-clarity
# or:  pip install tau-by-clarity

Requires Python 3.11+. Verify with tau --help.

To embed the library instead of the CLI, pip install tau-by-clarity then import pi_coding_agent (or the Tau alias import tau_coding_agent).

From source (development)

git clone https://github.com/Nuosis/tau-by-clarity.git
cd tau-by-clarity
uv sync               # workspace + dev deps
uv run tau            # run from the checkout

Dependencies

Installed automatically by uv sync (declared in pyproject.toml):

Required (runtime)	Purpose
pydantic (≥2)	typed models / contracts
anthropic, openai, google-genai, boto3	provider SDKs (Anthropic, OpenAI-compatible, Gemini, Bedrock)
httpx	HTTP transport

Optional (feature-gated, lazy-imported — not required)	Enables
presidio-analyzer	Tau by Clarity's NER detector (regex detection works without it)
Ollama + nomic-embed-text (local, http://localhost:11434)	local semantic recall embeddings for project-local memory

Dev/test extras (pytest, pytest-asyncio, …) install with uv sync --extra dev.

---

Quick Start

1. Set a provider key

Tau reads provider keys from environment variables:

export ANTHROPIC_API_KEY=...     # or  OPENAI_API_KEY · GEMINI_API_KEY · GOOGLE_API_KEY
# AWS Bedrock: export AWS_ACCESS_KEY_ID=...  AWS_SECRET_ACCESS_KEY=...

Add the export to your shell profile to persist it. (A project-local .env is also auto-loaded when running from a source checkout — convenient for development.)

2. Launch the Interactive TUI

tau

This opens the full-featured terminal UI where you can chat with the coding agent.

Keyboard shortcuts:

Key	Action
Enter	Send message
Shift+Enter	New line in input
/	Slash command completion
@	File path completion
Ctrl+P	Cycle to next model
Ctrl+C / Esc	Quit

3. Try a Simple Task

Type in the terminal:

Create a Python function to calculate fibonacci numbers

The agent will write the code and save it to your current directory.

---

Common Use Cases

Single Prompt (Non-Interactive)

For scripting or quick tasks:

tau --print "Write a quicksort in Python"

The agent's response prints to stdout and exits.

Switch Models

# Use a specific model
tau --model gemini-2.5-pro-preview

# Use a provider + model name
tau --provider google --model gemini-2.0-flash

# List all available models
tau --list-models

Resume Previous Sessions

# Continue the most recent session
tau --continue

# Pick from a list of previous sessions
tau --resume

Slash Commands in TUI

Type / in the interactive TUI to see available commands:

Command	Description
/model <name>	Switch to a different model
/thinking <level>	Set thinking detail: minimal · low · medium · high · xhigh
/compact	Compress conversation context to save tokens
/session	Show session statistics (tokens used, cost estimate)
/tools	List all active tools available to the agent

Full CLI Help

tau --help

---

Building agents — tau as an agent framework

Tau isn't only the bundled coding agent — it's a framework for building your own agents and subagents. The agent directory is the deployment unit: prompts, tools, skills, subagents, and evals all live inside it.

Agent directory layout

my-agent/
├── OBJECTIVES.md             # user stories, success conditions, I/O artifact contracts
└── .tau/
    ├── settings.json         # provider/model, tool allow-list, extensions, name
    ├── SYSTEM.md             # brief system prompt: identity, hard rules, voice
    ├── extensions/           # your tools (extension_factory) — auto-discovered
    ├── skills/               # agent-local procedural knowledge
    └── subagents/<name>/     # each subagent is itself a full agent dir

Point the runtime at it with PI_CODING_AGENT_DIR=/path/to/my-agent/.tau, then run headless (tau --mode json -p "...") or in the TUI.

Tools are extensions

Register a tool from an extension_factory(pi) in .tau/extensions/*.py, with a typed parameter schema and structured result:

def extension_factory(pi):
    async def execute(tool_call_id, params, signal, on_update, ctx):
        return {"content": [{"type": "text", "text": "..."}], "details": {...}}

    pi.register_tool(
        name="my_tool", label="My Tool", description="What it does.",
        parameters={"type": "object", "properties": {...}, "required": [...]},
        execute=execute,
    )

activate = extension_factory   # loader alias

settings.json tools / extensions lists are structural access control — an empty tools list is a deliberate denial of the default tools, not a hint.

Subagents

A subagent is a full tau agent under .tau/subagents/<name>/. The parent spawns it in isolation (its own settings, SYSTEM.md, extensions), hands it a typed input artifact, and reads back a typed output artifact.

The build discipline

Agent-authoring guidance — directory layout, OBJECTIVES.md contracts, prompt-last sequencing, and the compile → unit-test → live-eval gates — is vendored in skills/agent-build-pattern.

---

Privacy — Tau by Clarity (default ON) — pi_coding_agent.clarity_pii

Tau by Clarity tokenizes personal data before it reaches any model provider, for every LLM call regardless of source (agent sessions, the outer loop, evals, any direct pi_ai use) — installed at the universal pi_ai dispatch hook. Real values never leave the machine: the model sees stable tokens like [PII:EMAIL:1], and the reply is detokenized transparently.

• Reversible per-session vault. Token↔value mappings persist as a lazy, session-referenced artifact at pii_vault/<session>.json — written only when a session actually contains PII, carrying {schema, session_id, created_at, updated_at}. No-PII sessions create no artifact.
• Detection. Built-in high-confidence regex (email, US SSN, phone, credit card with Luhn, IPv4, IBAN, AWS keys) plus optional Presidio NER (lazy-imported — never a hard dependency).
• Control. On by default; disable process-wide with PI_CLARITY_PII_DISABLED=1, or inspect/toggle in the TUI with /pii (status | on | off | vault | reveal <text> | clear).

PII tokenization runs after active compression at the same pi_ai chokepoint, so compressed tool outputs are tokenized before they are sent.

---

Context management & compression

The harness reduces context with one primary mechanism and one fallback. They are mutually exclusive — see design/context-and-memory-management.md §12.

Active compression (default ON) — pi_coding_agent.active_compression

Content-aware, reversible compression of large tool-output payloads, applied universally at the pi_ai dispatch layer (every LLM call, any source; only toolResult messages, never the live prompt). JSON arrays are sampled with error/anomaly items always kept; logs keep error lines; big text keeps head+tail. The original is cached in a local hash-indexed CCR store (SQLite) and is recoverable:

• the model can call the ccr_retrieve tool with a [CCR:<handle>] handle, and
• the harness auto-rehydrates a compressed block in place when the model references its handle (so recovery doesn't depend on the model calling a tool).

Control with the active_compression flag in settings.json — on if the key is absent. Disable per-project with "active_compression": false, or process-wide with the env var PI_ACTIVE_COMPRESSION_DISABLED=1.

// .tau/settings.json
{ "active_compression": true }   // omit entirely for the same (default-on) effect

Summarization compaction — the fallback

The older threshold-based summarization compaction (§7) now runs only when active compression is off. When active compression is on, it owns context reduction and proactive compaction stands down. (Emergency overflow compaction remains unconditional as a hard-limit safety net.)

Position-based working-context compression — removed

The memory module's positional middle-compression (compress_working_context) was dropped (§12); active compression replaces it. Project-local memory (memory_enabled) now only records and recalls atomic facts — it no longer compresses the working set.

---

Running Tests

All tests

uv run pytest

Per-package

uv run pytest packages/tui/tests/          # TUI components
uv run pytest packages/ai/tests/           # AI providers
uv run pytest packages/agent/tests/        # Agent core
uv run pytest packages/coding-agent/tests/ # CLI + coding agent

Live API tests (requires GEMINI_API_KEY)

uv run pytest packages/ai/tests/ --live -v

# Or via environment variable
LIVE_TESTS=1 uv run pytest packages/ai/tests/ -v

All tests run against mocks by default — no API key required, no quota consumed.

---

Test Status

Package	Tests	Status
pi_tui	135	✅ passed
pi_ai + pi_agent	156	✅ passed (7 skipped = live-only)
pi_coding_agent	287	✅ passed
Total	578	✅ all passing

---

Project Structure

tau-by-clarity/
├── .env                          ← API keys (never commit)
├── pyproject.toml                ← uv workspace root
├── conftest.py                   ← global pytest config (.env loader)
└── packages/
    ├── ai/                       ← LLM provider layer
    │   └── src/pi_ai/
    │       ├── providers/        ← google.py, openai.py, anthropic.py, …
    │       ├── stream.py         ← unified stream_simple() / complete_simple()
    │       └── utils/            ← overflow detection, JSON parse, …
    ├── agent/                    ← core agent loop
    │   └── src/pi_agent/
    │       ├── agent.py          ← main run loop
    │       ├── tools/            ← tool registry & execution
    │       └── session.py        ← session state
    ├── coding-agent/             ← CLI entry point & extensions
    │   └── src/pi_coding_agent/
    │       ├── cli.py            ← `pi` command
    │       ├── core/             ← AgentSession, system prompt, tools
    │       └── modes/interactive/← TUI interactive mode
    └── tui/                      ← terminal UI library
        └── src/pi_tui/
            ├── components/       ← Editor, SelectList, Markdown, …
            ├── tui.py            ← differential rendering engine
            └── keys.py           ← Kitty keyboard protocol parser

---

TypeScript → Python Mapping

TypeScript	Python
interface X {}	class X(BaseModel): or @dataclass
type X = A | B	X = Union[A, B]
async function f()	async def f()
AsyncIterable<T>	AsyncGenerator[T, None]
AbortSignal	asyncio.Event (cancellation token)
EventEmitter	dict[str, list[Callable]]
TypeBox schema	pydantic.BaseModel
vitest	pytest + pytest-asyncio

---

FAQ

Problem	Solution
uv: command not found	Run the install script: curl -LsSf https://astral.sh/uv/install.sh | sh
GEMINI_API_KEY not set	Add your key to .env
ModuleNotFoundError: pi_tui	Use uv run tau instead of python directly
TUI shows garbled characters	Ensure your terminal supports UTF-8 (iTerm2, Warp, or any modern terminal)
Tests are skipped	Add --live to run real API tests
400 thought_signature error	Upgrade to the latest version — this is fixed in the google provider

---

Credits & lineage

Tau stands on the shoulders of the PI project. Its architecture, algorithms, package boundaries, and the pi_* import namespaces come from there:

• PI (pi-mono) — the original TypeScript coding-agent monorepo by Mario Zechner (@badlogic, @mariozechner/*). Tau mirrors its design directly.
• PI for Python — Tau is forked from the Python port at openxjarvis/pi-mono-python.

Package lineage (and why the import names are pi_*):

PI (TypeScript)	Tau (Python)	Layer
@mariozechner/pi-ai	pi_ai	Unified LLM streaming (Google, Anthropic, OpenAI, Bedrock, …)
@mariozechner/pi-agent-core	pi_agent	Agent loop, tool execution, state
@mariozechner/pi-coding-agent	pi_coding_agent	Coding agent + file tools
@mariozechner/pi-tui	pi_tui	Terminal UI rendering engine

With gratitude to the PI authors and contributors.