regista-harness 0.3.0

Observability-first agent harness: every session is a structured, deterministically replayable trace.

regista

The agent harness with a flight recorder.

regista is an observability-first agent harness for Python: the runtime layer that turns a stateless LLM API into an agent that can act. Every session is recorded as a structured, append-only trace — complete enough to deterministically replay the entire session with zero API calls and zero cost.

regista (Italian): the deep-lying playmaker who directs the game. Also: "director."

pip install regista-harness        # the import name is `regista`

Status: v0.2 released; v0.3 feature-complete on main. Loop, tools, environment, policy, Anthropic + OpenAI-compatible providers, streaming, compaction, deterministic replay, OTel export, Agent.resume, MCP client, eval/regression runner — plus v0.3's subagents, Skills, and ContainerEnvironment. 180+ tests, strict mypy, every subsystem traced.

Why another harness?

Every serious agent harness treats observability as an add-on or a paid platform. regista treats the trace as the product:

• 🔍 Structured event log — every LLM call, tool execution, permission decision, token count, and dollar of cost, as typed JSONL events. If a behavior isn't in the trace, it's a bug.
• ⏪ Deterministic replay — re-run any recorded session with the LLM and tools served from the recording: hermetic, keyless, $0. Strict mode hash-verifies every request and fails with a structural diff — zero-cost regression tests for CI.
• 📊 OpenTelemetry export — the same trace as session/turn/llm/tool spans in Jaeger/Grafana, post-hoc, with the recorded timestamps.
• 🔌 Provider-neutral — native Anthropic adapter plus an OpenAI-compatible adapter (OpenAI, Ollama, vLLM, gateways). Streaming on both.
• 🛡️ Permission gate — Allow/Deny/Ask policies on every tool call, workspace-scoped execution, honestly documented as a gate (not a sandbox).

A coding agent in ~30 lines

import asyncio

from regista import Agent
from regista.environment import LocalEnvironment
from regista.policy import PermissionRequest, workspace
from regista.providers import AnthropicProvider
from regista.tools.builtin import builtin_tools

env = LocalEnvironment("./sandbox")          # every file op is pinned here

async def approve(request: PermissionRequest) -> bool:
    return input(f"allow {request.tool_input}? [y/N] ") == "y"

agent = Agent(
    provider=AnthropicProvider("claude-sonnet-4-6"),   # the model is always explicit
    instructions="You are a careful coding agent. Run the tests after every change.",
    tools=builtin_tools(env),                # read/write/list/glob/search/shell/fetch
    policy=workspace(),                      # file tools allowed; shell escalates to Ask
    ask_handler=approve,
    max_turns=30,
    max_cost_usd=1.00,                       # hard budget, from provider-reported usage
)

result = asyncio.run(agent.run("pytest is failing in this repo — find the bug and fix it"))
print(result.output, f"${result.cost_usd:.2f}", result.trace_path)

The same Agent streams: async for event in agent.stream(task) yields text deltas, tool start/finish, and per-turn usage as they happen — and writes the identical trace.

The flight recorder

Every run writes a JSONL trace: session.start, llm.request (with a SHA-256 request_hash), llm.response, tool.call, permission.decision, tool.result, context.compaction, session.end. That log is complete enough to re-run the session:

from regista import replay

replayed = await replay(result.trace_path)   # no API key, no network, no side effects
assert replayed.cost_usd == 0.0
assert replayed.output == result.output

Replay is hash-verified call by call. Change anything that shaped a request — the prompt, a tool schema, a sampling param — and strict mode fails with a diff pointing at exactly what drifted:

ReplayDivergence: request 1 diverged from the recording (trace seq 1)
request.system: recorded 'You are a greeter.' != live 'You are a greeter. Always answer in French.'

mode="warn" keeps serving positionally for time-travel debugging; mode="hybrid" falls through to a live provider mid-session. agent.resume(trace_path) packages that: the recorded prefix replays for $0 (tool effects never re-run), and the session continues live from wherever the crash left it.

Zero-cost testing

FakeProvider is public API. Script the model, run the real loop, assert on the real trace:

from regista.providers import FakeProvider, text_response, tool_use_response

provider = FakeProvider([
    tool_use_response(("tu_1", "greet", {"name": "world"})),
    text_response("done"),
])

~90% of regista's own test suite runs on it — and your agent's tests can too. Committed traces replayed strictly in CI are your end-to-end tests, forever, for free.

regista.evals packages that into a regression runner: an EvalSuite of tasks with checks on outcome and trace shape (output_contains, tool_never_called("shell"), max_cost_usd(0.50), …). suite.record(agent) saves passing runs as fixtures; suite.replay() re-judges them in CI at $0, with any divergence from the recording failing the task.

Architecture

regista is organized around nine harness primitives — instructions, context management, tool interface, execution environment, durable state, orchestration, subagents, skills, verification & observability — each mapping 1:1 to a module with one interface. ARCHITECTURE.md explains the whole system from first principles, including the dependency diagram and the life of a request; it's written for people who have never built an agent.

Try it

On PyPI the distribution is named regista-harness (the bare name was squatted); the import name is regista everywhere. From source:

git clone https://github.com/mohakwagh/regista && cd regista
uv sync
uv run python examples/01_hello_agent.py      # $0: scripted model, real loop, real trace
uv run python examples/05_replay.py           # $0: replay + a divergence diff
uv run --extra mcp python examples/06_mcp.py  # $0: MCP tools, replayed with the server off
ANTHROPIC_API_KEY=... uv run python examples/04_real_provider.py   # ~1 cent

Roadmap

• v0.2 — all on main: Agent.resume() · MCP client · eval/regression runner (replay-powered $0 CI)
• v0.3 — all on main: subagents (Agent.as_tool) · Skills · ContainerEnvironment

Safety, honestly

The permission layer is a policy gate, not a sandbox: an allowed shell command can do anything your user account can. The environment scopes paths (symlinks included), strips secrets from subprocess env, and enforces timeouts — and every decision is in the trace. For untrusted tasks, ContainerEnvironment runs commands inside Docker with a one-line swap. SECURITY.md has the full threat model.

License

MIT