mira-sdk

Client SDK for agent processes in the Mira ecosystem (MiraRun/MiraGen). Ships two deliberately separated faces (see docs/architecture.md for why they must never share a process):

• mira_sdk.telemetry — run-scoped OTLP telemetry export, correlated with the run/routine/agent identity a Mira-orchestrated process already has. For embedding in any ordinary service.
• mira_sdk.driver — a standalone, privileged, read-only infrastructure observer: bounded Docker discovery (list/inspect/logs/stats), host metrics with OpenTelemetry semantic-convention names, poll-based container death/OOM detection with post-mortem log tails, and a report loop (mira-driver) that publishes to MiraRun's reported-target endpoint (ADR-022) and/or movingfirm-admin's infra collector.

Install

pip install mira-sdk            # telemetry only
pip install 'mira-sdk[driver]'  # + the driver toolkit (httpx)

Quickstart

from mira_sdk import MiraTelemetry

telemetry = MiraTelemetry(
    endpoint="https://mira.example/otlp/v1/traces",
    run_id=run_id,
    routine_id=routine_id,
    routine_revision_id=routine_revision_id,
    agent_id=agent_id,
    token=run_credential,  # same run-scoped credential used for MCP auth
)

with telemetry.span("resource.query", **{"mira.resource.uri": target_uri}):
    ...  # do the thing; exceptions are recorded on the span and re-raised

telemetry.shutdown()  # once, at process exit

The driver as a process

MIRA_DRIVER_TARGET_REFERENCE=vps1 \
MIRA_DRIVER_ADMIN_COLLECTOR_URL=http://10.8.0.4:6767/api/ops/infra/collector \
MIRA_DRIVER_ADMIN_COLLECTOR_TOKEN=... \
mira-driver

Configuration is env vars only, documented in mira_sdk/driver/process.py. The reference deployment — a read-only Docker socket proxy plus the driver, as one Compose stack — lives in deploy/; the socket proxy is the single privileged component, and it doubles as MiraRun's direct-pull endpoint on WireGuard-reachable hosts. As a library, the same pieces compose explicitly: DriverRunner(DockerDriver(...), [MirarunReportSink(...), AdminCollectorSink(...)], ...).

Design

• Non-blocking by construction. Spans are queued and exported on a background thread (BatchSpanProcessor); this module never wires a synchronous exporter. A telemetry backend being down must not fail or stall the operation that emitted the span.
• Identity set once. run_id/agent_id/routine_id/routine_revision_id/service.*/deployment.environment are OTel resource attributes — set once at construction, present on every span this instance emits. Attaching them per-span instead is the mistake that breaks filtering and aggregation once there's a real backend behind the endpoint.
• OTLP, not a vendor SDK. endpoint is a plain OTLP/HTTP traces endpoint. This package has no opinion on what's behind it — an OTel Collector, a vendor backend, or a bare receiver during development.
• export_stats for observability, not a promise of completeness. It counts export batches that reached the exporter and failed — the actionable half of "did my telemetry get out." Queue-overflow drops (a producer outpacing the export thread) are logged by OpenTelemetry's own opentelemetry.sdk.trace.export logger, not duplicated here.

Development

pip install -e ".[dev]"
pytest

Releasing

Publishing to PyPI happens on GitHub Release (Trusted Publishing / OIDC — no token stored in this repo). Bump version in pyproject.toml, merge, then cut a GitHub Release; .github/workflows/release.yml builds and publishes.