techrevati-runtime

Production-grade runtime primitives for multi-step LLM agent loops — sync and async, with retry classification, circuit-breaker protection, per-model cost tracking, opt-in budget enforcement, role-based tool gating, content guardrails, agent-to-agent handoffs, declarative policy, durable checkpointing, token-aware rate limiting, and OpenTelemetry GenAI semantic conventions out of the box. Beta — 0.2.x; 0.x APIs remain explicitly unstable.

pip install techrevati-runtime
# Or with OpenTelemetry:
pip install 'techrevati-runtime[otel]'

Quick start

from techrevati.runtime import (
    Orchestrator, UsageSnapshot, ModelPricing, register_pricing,
)

register_pricing("model-a", ModelPricing(input_per_million=3.0, output_per_million=15.0))

orch = Orchestrator(
    role="writer", phase="draft", project_id=1,
    budget_usd=10.0, enforce_budget=True, max_iterations=25,
)
with orch.session() as session:
    result, usage = session.run_turn(
        lambda: call_model(prompt),
        model="model-a",
        usage=UsageSnapshot(input_tokens=5000, output_tokens=1200),
        timeout=30.0,
    )
print(session.summary())

The session walks the worker through INITIALIZING → RUNNING → COMPLETED, classifies any exception that bubbles up into a typed failure scenario, attempts recovery once, enforces the budget, gates tool calls behind permissions and guardrails, and emits structured events to any sink you configure — without you wiring any of it by hand.

Async sibling: replace with with async with, session() with asession(), run_turn with arun_turn. Same parameters. asyncio.CancelledError cleanly transitions the worker to CANCELLED.

For an end-to-end example exercising every primitive (permissions + breaker + budget + guardrail + handoff + policy + OTel), see examples/tiny_agent.py and the end-to-end tutorial.

Design goals

• Zero runtime dependencies. Imports are stdlib only. OpenTelemetry is an optional [otel] extra.
• Type-safe. py.typed marker shipped; clean under mypy --strict.
• Composable. Every primitive (CircuitBreaker, AsyncCircuitBreaker, RetryContext, QualityGate, PolicyEngine, UsageTracker, PermissionEnforcer, Guardrail, Handoff) is usable standalone. The Orchestrator is just the wiring.
• Thread-safe and async-safe. threading.Lock in sync paths, asyncio.Lock in async paths. State is per-instance.
• Configuration-free at the edges. Pricing data is empty by default; phase thresholds are not hardcoded; permission roles are caller-defined. The runtime stays opinion-free about what your numbers mean.

Primitives

Module	Provides
orchestrator	Orchestrator, OrchestrationSession, AsyncOrchestrationSession, AgentSession
circuit_breaker	CircuitBreaker, AsyncCircuitBreaker (CLOSED/OPEN/HALF_OPEN with configurable probe permits)
retry_policy	classify_exception, attempt_recovery (sync + async), backoff_delay with full/equal/decorrelated jitter
usage_tracking	UsageTracker, register_pricing, load_pricing_from_file, BudgetExceededError, has_pricing
agent_lifecycle	AgentRegistry, AgentWorker with validated state machine including CANCELLED
agent_events	Typed lifecycle events + OpenTelemetry attribute bridge
permissions	Role × tool authorization, deny-first
guardrails	Pre-call + post-call content gating around run_tool / arun_tool
handoffs	Handoff value + session.handoff_to() agent-to-agent delegation
policy_engine	Composable conditions and rule evaluator with auto-elapsed time
sinks	EventSink / UsageSink Protocols + ring-buffered defaults
otel (optional)	OpenTelemetrySink + OpenTelemetryUsageSink emitting GenAI semconv spans/metrics

Showcase

Async with handoff and guardrails

import asyncio
from techrevati.runtime import (
    AllowAllGuardrail, AsyncCircuitBreaker, Orchestrator, UsageSnapshot,
)

cb = AsyncCircuitBreaker("model-api", failure_threshold=3, recovery_timeout_seconds=30.0)

async def main():
    orch = Orchestrator(
        role="writer", phase="draft",
        async_circuit_breaker=cb,
        guardrails=[AllowAllGuardrail()],
        max_iterations=10,
    )
    async with orch.asession() as session:
        text, _ = await session.arun_turn(
            lambda: acall_model(prompt),
            model="model-a",
            usage=UsageSnapshot(input_tokens=5000, output_tokens=1200),
            timeout=30.0,
        )
        handoff = session.handoff_to("editor", reason="review", context={"draft": text})
        print(f"handed off to {handoff.target_role}")

asyncio.run(main())

OpenTelemetry observability

from techrevati.runtime import Orchestrator
from techrevati.runtime.otel import OpenTelemetrySink, OpenTelemetryUsageSink

orch = Orchestrator(
    role="writer", phase="draft",
    event_sink=OpenTelemetrySink(agent_id="writer-001"),
    usage_sink=OpenTelemetryUsageSink(),
)
# Every AgentEvent now appears as an OTel span with gen_ai.operation.name,
# gen_ai.agent.id, gen_ai.usage.{input,output}_tokens. Drop-in compatible
# with any APM ingest that already understands GenAI semconv.

See docs/api/otel.md for the full attribute list and span name mapping.

Standalone primitives

Pick just what you need. Each primitive is usable on its own without Orchestrator.

from techrevati.runtime import (
    CircuitBreaker, CircuitOpenError,
    UsageTracker, UsageSnapshot,
    classify_exception, attempt_recovery, RecoveryContext,
)

cb = CircuitBreaker("downstream", failure_threshold=5, recovery_timeout_seconds=60.0)
result = cb.call(fetch, url, timeout=10)  # raises CircuitOpenError if tripped

ctx = RecoveryContext()
scenario = classify_exception(my_error)
recovery = attempt_recovery(scenario, ctx)  # returns RecoveryResult with steps to retry

tracker = UsageTracker()
tracker.record_turn("model-a", UsageSnapshot(input_tokens=5000, output_tokens=1200))
print(tracker.format_cost())

Why not LangGraph / OpenAI Agents SDK?

techrevati-runtime is intentionally smaller and narrower than either:

• LangGraph is a workflow engine with durable execution, checkpointer protocols, and a graph model. Use it when your agent flow is a graph that needs to survive restarts and you're OK with the LangChain ecosystem footprint.
• OpenAI Agents SDK is a cohesive runtime tied to OpenAI's models, with default tracing through their dashboards. Use it when you're committed to OpenAI and want the smoothest path.
• techrevati-runtime is a zero-dep primitive set. Sync + async. Vendor-neutral. Emits OpenTelemetry GenAI semantic conventions so the same APM dashboards that consume OpenAI Agents SDK telemetry will pick us up too. Bring your own model client and your own persistence — the runtime stays opinion-free.

The runtime ships a pluggable CheckpointSaver protocol with InMemorySaver and SqliteSaver implementations (0.2.0) — enough for resume-from-checkpoint replay across restarts. It is still not a full durable workflow engine in the Temporal sense; pair with Temporal, dbos, or LangGraph's checkpointer if you need cross-host scheduling, retries-as-history, or a durable timer service.

Limitations (be honest with yourself before adopting)

• Pricing must be registered. The bundled pricing.json is intentionally empty. Without register_pricing() or load_pricing_from_file(), every cost calculation returns $0.00 (you will see a one-time warning per model).
• Budget enforcement is opt-in. Set Orchestrator(enforce_budget=True) to raise BudgetExceededError; the default merely records an event and continues.
• Permissions are advisory. OrchestrationSession.run_tool() enforces; run_turn() does not gate model calls. There is no sandbox — pair with OS-level isolation if needed.
• Durable execution is opt-in. Default sessions are in-memory; pass a CheckpointSaver (e.g. SqliteSaver) plus a stable thread_id to get resume-from-checkpoint replay. Pair with Temporal/dbos if you need cross-host scheduling or durable timers.
• Default sinks are in-memory ring buffers. Long-running sessions need a durable EventSink and UsageSink (e.g. OpenTelemetrySink, or your own).
• CircuitBreaker state is per-process. Each replica counts its own failures. Add a shared coordinator if you need fleet-wide breaker state.

Status

techrevati-runtime is at version 0.2.0 (beta). This release ships durable execution (CheckpointSaver + SqliteSaver), token-aware rate limiting (RateLimiter / AsyncRateLimiter), provider routing, per-session UsageLimits, nested OTel agent spans, persistent SQLite sinks, and supply-chain hardening (CycloneDX SBOM + CodeQL + zero-deps smoke). The AgentSession rename and OTel wire-format change are the two soft-breaking items — see docs/migrating-from-0.1.x.md. 0.x APIs remain unstable; breaking changes will continue to be gated by deprecation warnings. Pinning Python 3.11+ for from __future__ import annotations ergonomics and modern asyncio.

See CHANGELOG.md for the per-sprint release notes and docs/tutorials/end-to-end.md for a guided tour of every primitive.

Issues and PRs welcome — see CONTRIBUTING.md and SECURITY.md.

License

MIT — copyright © 2026 TechRevati doo. See LICENSE.