streamflow-pulse-client 2.7.0

Official Python client for StreamFlow Pulse — AI Agent Platform

streamflow-pulse-client — Python SDK for StreamFlow Pulse

Official Python client for the Pulse AI Agent Platform.

Distribution name on PyPI is streamflow-pulse-client; import statement stays the natural from pulse_client import ... (same convention as python-dateutil → import dateutil).

from pulse_client import PulseClient

with PulseClient("http://localhost:9090") as client:
    client.auth.login("alice", "secret")
    for pipeline in client.pipelines.list():
        print(pipeline["name"])

Install

pip install streamflow-pulse-client

Requires Python 3.10+. Pure Python — only depends on httpx.

Why pulse-client

• Pythonic — context-manager friendly, typed exceptions, attribute-style resource access (client.pipelines.list()).
• Lightweight — single dependency (httpx), <500 LoC, no generated bloat.
• Spec-aligned — every method corresponds 1:1 to an endpoint in the Pulse OpenAPI 3.1 spec. Drift is caught at PR time by the in-tree spec invariant tests (B-103).
• Async-ready — the sync client ships today; an AsyncPulseClient (same surface, await everywhere) will follow in v3.0.

Quick start

from pulse_client import PulseClient, PulseAuthError

client = PulseClient("http://localhost:9090")

# Authenticate — the returned JWT is cached on the client automatically
try:
    response = client.auth.login("alice", "secret")
    print(f"Logged in as {response['user']['username']}")
except PulseAuthError as e:
    print(f"Login failed: {e}")

# List + inspect resources
for pipeline in client.pipelines.list():
    print(pipeline["name"], pipeline["status"])

# Create a pipeline from a template
new_pipeline = client.pipelines.create({
    "name": "my-fraud-detector",
    "templateId": "fintech-fraud-detection-realtime",
    "nodes": [
        {"id": "source", "type": "source", "subType": "kafka-source"},
        {"id": "agent", "type": "agent", "subType": "streaming"},
        {"id": "sink", "type": "sink", "subType": "telegram"},
    ],
})

# Inspect deployed agents
for agent in client.agents.list():
    print(f"  {agent['name']} — {agent['engineType']} — {agent['status']}")

client.close()

Supported surfaces (v2.6.0)

Resource	Methods	Notes
client.auth	login(), refresh(), organizations(), switch_org()	Auto-caches JWT on the client after login / refresh / switch_org.
client.pipelines	list(), get(id), create(definition), delete(id)	definition follows the CreatePipelineRequest schema (see OpenAPI spec).
client.agents	list(), get(id)	Read-only — agents are owned by pipelines.
client.templates	list()	The 223+ first-party templates.
client.users	list()	Requires USERS_LIST permission (Owner / Platform Admin personas).
client.version()	top-level	Public — no JWT required.

The full ~112-endpoint surface (admin, audit, backups, chat, workspace, etc.) is documented in the OpenAPI spec at <pulse-server>/api-docs. SDK methods for those land opportunistically as user-facing demand surfaces.

Embedded ML inference & duplex

Score events with an uploaded ONNX model in-process (B-112), and open a bidirectional duplex channel for synchronous decisions (B-114). Full guide: ML inference & duplex.

# Upload + score with an ONNX model (no model-server hop)
client.models.upload(name="fraud", path="./fraud.onnx",
                     input_schema={"amount": "float", "country": "float"})
builder.from_topic("transactions").ml_predict(
    model="fraud", input_fields=["amount", "country"], output_field="prediction"
).filter("prediction.fraud_score > 0.8").to_topic("flagged")

# Upload + run a sandboxed WASM module over each event (B-110, pure-Java
# Chicory on the engine — no host syscalls). Authored from any wasm32
# toolchain (Rust, TinyGo, AssemblyScript, C).
client.wasm.upload(name="pii-redactor", path="./redactor.wasm")
builder.from_topic("events").wasm(module="pii-redactor").to_topic("clean")

# Duplex: send in, receive the correlated output on one connection
# (pip install streamflow-pulse-client[duplex])
async with client.duplex("fraud-detector") as ch:
    await ch.send({"amount": 5000}, correlation_id="tx-1")
    signal = await ch.recv()        # signal["correlation_id"] == "tx-1"

Authentication

Three patterns, pick what fits:

# 1. Username + password (interactive / CLI tools)
client = PulseClient("http://localhost:9090")
client.auth.login("alice", "secret")

# 2. Pre-minted JWT (CI / service accounts)
client = PulseClient("http://localhost:9090", token="ey...")

# 3. JWT from environment (12-factor apps)
import os
client = PulseClient(os.environ["PULSE_URL"], token=os.environ["PULSE_TOKEN"])

For long-running daemons, store the refreshToken from login() and call client.auth.refresh(refresh_token) when the JWT nears expiry (default 1 h TTL).

Error handling

Every server error becomes a typed exception you can catch precisely:

from pulse_client import (
    PulseClientError,   # base — catches every client-side error
    PulseAuthError,     # 401 — invalid / missing / expired JWT
    PulseNotFoundError, # 404
    PulseValidationError, # 400 — malformed request
    PulseRateLimitError,  # 429 — carries .retry_after_seconds
    PulseAPIError,      # everything else (5xx, etc.)
)

try:
    client.pipelines.get("nope")
except PulseNotFoundError:
    print("Doesn't exist — fine")
except PulseRateLimitError as e:
    print(f"Backing off {e.retry_after_seconds}s")
    time.sleep(e.retry_after_seconds or 60)
except PulseClientError as e:
    print(f"Something else went wrong: {e}")

Every exception carries .status_code, .path, and .body so log lines + bug reports are actionable.

Automatic retry (opt-in)

By default the client makes exactly one attempt per request and surfaces the typed error — retries are off so nothing is retried behind your back. Opt in with max_retries:

client = PulseClient(
    "http://localhost:9090",
    token="ey...",
    max_retries=3,                 # 0 = off (default)
    retry_backoff=0.2,             # base seconds; full-jitter exponential backoff
    retry_max_backoff=10.0,        # cap per attempt
    retry_on_status=(502, 503, 504),
    retry_idempotent_only=True,    # don't retry POST on 5xx (default)
)

Policy:

• 429 (rate limited) is retried for any method (the request was rejected, never processed) and honours retryAfterSeconds / the Retry-After header before falling back to backoff;
• retry_on_status 5xx and transport (connect/read) errors are retried only for idempotent methods (GET/HEAD/PUT/DELETE) unless retry_idempotent_only=False — so a POST create is never silently duplicated;
• terminal 4xx (400/401/404) are never retried; retries are bounded by max_retries.

This opt-in retry policy currently ships in the Python SDK as the reference implementation. Rolling the same policy out to the Rust / Go / JS / Java SDKs is tracked as B-170 (issue #312); those SDKs surface retryAfter on the rate-limit error today so callers can retry manually.

Development

git clone https://github.com/olsisoft/streamflow.git
cd streamflow/pulse-py

# Install in editable mode with dev deps
pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check src tests
mypy src

CI runs the same on every push touching pulse-py/ — see .github/workflows/pulse-py.yaml.

Local pipeline simulation (Python-exclusive)

The streams DSL is client-side declaration, server-side execution — but the Python SDK additionally ships a local, in-process executor: the moral equivalent of Kafka Streams' TopologyTestDriver, with no server and no JVM. Run your pipeline over sample events to see what would reach the sink, before you deploy:

from pulse_client.streams import StreamBuilder, windows, aggs

builder = (
    StreamBuilder("card-velocity-60s")
    .from_topic("card-authorizations")
    .key_by("cardId")
    .window(windows.tumbling("60s"), aggregations={"txCount": aggs.count()})
    .filter("txCount > 5")
    .to_topic("fraud-alerts")
)

# Feed synthetic events through the SAME operator chain deploy() would POST:
survivors = builder.simulate([
    {"cardId": "card-7", "amount": 10, "_ts": 1_000},
    # … more events in the same 60s window …
    {"cardId": "card-7", "amount": 10, "_ts": 70_000},  # advances the watermark, closes the window
])
print(survivors)  # the window emissions that crossed txCount > 5

simulate() supports filter / map / flat_map / key_by / window (tumbling + global) and all 7 aggregators, with a safe ast-based expression evaluator (no eval). Engine-bound operators (enrich, cep, joins, LLM/MCP/ML) raise NotImplementedError — deploy server-side for those.

This is unique to the Python SDK today. The Rust / Go / JS / Java SDKs declare client-side (compile) and execute server-side (deploy) but have no in-process simulator. Cross-language parity is tracked as B-169 (issue #311).

Streaming SQL (compile SQL → pipeline) — B-097

Write a streaming pipeline as SQL; it compiles client-side to the same StreamBuilder pipeline (a KSQL/Flink-SQL-flavoured subset):

from pulse_client import PulseClient, compile_sql

builder = compile_sql(
    """SELECT count(*) AS cnt, sum(amount) AS total
       FROM payments
       WHERE amount > 1000
       GROUP BY customer_id
       WINDOW TUMBLING(60s)
       HAVING cnt > 5
       INTO fraud-alerts""",
    name="fraud-detector",
)

with PulseClient(url, token=jwt) as client:
    client.streams.deploy(builder)          # or: client.streams.from_sql(sql, name=...)

Supported: SELECT aggregates (count(*), sum/avg/min/max(f), distinct_count(f), collect_list(f)) with AS aliases, *, and plain-column projection (→ a map); WHERE/HAVING (SQL =/<>/AND/OR translated to ==/!=/&&/||); GROUP BY; WINDOW TUMBLING/SLIDING/SESSION/COUNT/GLOBAL; INTO. Inspect the result with builder.build() or run it with builder.simulate(events) before deploying.

Roadmap

• v2.5.x — current sync API, 5 core resources (auth, pipelines, agents, templates, users), version().
• v2.6.x — expanded resource coverage: backups, schedules, credentials, settings, approvals, chat.
• v3.0 — AsyncPulseClient with async def everywhere; same surface; one library, two clients.
• B-098 satellite — once olsisoft/pulse-py exists as its own repo, this in-tree code lifts out wholesale. Pip-install will switch to the satellite; in-tree continues to mirror for one release cycle so the migration is non-breaking.

Track progress in docs/STREAMFLOW-BACKLOG.md under item B-098.

License

Apache 2.0 — same as the parent Pulse repository.