DeepStrike Python SDK

Build Python Agents with providers, typed tools, memory, Skills, delegation, workflows, and durable sessions. The SDK keeps the Agent's long-running work explicit through stream events, SessionLog evidence, tool policies, and host-provided integrations.

Install

pip install deepstrike

Requires Python 3.10+. The Rust kernel is distributed as a pre-built wheel (deepstrike._kernel).

When developing against a local kernel build, rebuild the extension from the repo root:

maturin develop --manifest-path crates/deepstrike-py/Cargo.toml

Quick start

import asyncio
from deepstrike import (
    FileSessionLog,
    InMemorySessionLog,
    LocalExecutionPlane,
    OpenAIProvider,
    RuntimeOptions,
    RuntimeRunner,
    collect_text,
    tool,
)

@tool
async def add(x: int, y: int) -> str:
    """Add two numbers."""
    return str(x + y)

plane = LocalExecutionPlane().register(add)
runner = RuntimeRunner(RuntimeOptions(
    provider=OpenAIProvider(api_key="sk-...", model="gpt-5-mini"),
    session_log=FileSessionLog(".deepstrike/sessions"),
    execution_plane=plane,
    max_tokens=4096,
))

asyncio.run(collect_text(runner.run(
    session_id="math-1",
    goal="What is 17 + 28?",
)))
# => "45"

Recipes — the canonical entry points

Most apps need one of two shapes. Start with the facades; drop to RuntimeRunner for streaming, tools, signals, memory, or governance.

from deepstrike import run_agent, run_fanout

# 1) Single agent — one prompt, one model, the text back.
answer = await run_agent(provider=provider, goal="What is 17 + 28?", tools=[add])

# 2) Parallel fan-out → synthesize over the kernel-gated DAG (safe from a stateless handler).
out = await run_fanout(
    provider=provider,
    tasks=["Summarize the auth module", "Summarize the data layer"],
    synthesize="Combine the findings into one summary.",
)
synthesis = out["synthesis"]

Same-session continuity is explicit via session_id:

await collect_text(runner.run(session_id="chat-1", goal="My name is Ada."))
reply = await collect_text(runner.run(session_id="chat-1", goal="What is my name?"))

Use InMemorySessionLog for process-local sessions or FileSessionLog when replay should survive restarts. wake(session_id) resumes from the event log without inserting a duplicate run_started event.

Streaming:

from deepstrike.providers.stream import TextDelta, ToolCallEvent, DoneEvent

async for event in runner.run(session_id="readme-1", goal="Summarize README.md"):
    if isinstance(event, TextDelta):
        print(event.delta, end="", flush=True)
    elif isinstance(event, ToolCallEvent):
        print(f"\n[→ {event.name}]")
    elif isinstance(event, DoneEvent):
        print(f"\ndone in {event.iterations} turns ({event.status})")

Architecture

┌─────────────────────────────────────────────────────────┐
│  RuntimeRunner (Layer 1.5)                              │
│  LLMProvider · ExecutionPlane · SessionLog · MemoryStore │
└───────────────────────────┬─────────────────────────────┘
                            │ durable prepare / append / commit
┌───────────────────────────▼─────────────────────────────┐
│  deepstrike._kernel Canonical Kernel ABI                │
│  P1 Syscall · P2 Sched · P3 MM · Proc · IPC             │
└─────────────────────────────────────────────────────────┘

The runner drives one durable operation loop:

1. The host prepares one canonical envelope and durably appends the exact core-produced record.
2. After commit, the kernel publishes typed effects such as provider, tool, or task work.
3. The SDK executes those effects and returns each outcome through the single resolve_effect input.
4. Typed observations and the terminal disposition are projected into SessionLog and the public stream.

Kernel session events carry an optional category tag (syscall · sched · mm · proc · ipc) for diagnostics and OS snapshot rebuilds.

What this enables

The mechanisms above are not internal refactors — they change what you can build without custom runner code:

Kernel-mediated runtime (M0–M4)
Tool calls, spawns, compression, and signals pass through one kernel gate with an explicit lifecycle (Ready / Running / Blocked / Suspended). You implement I/O; the kernel decides when and whether. Node, Python, and Rust share the same decision path, so wake(session_id) and cross-language tooling see consistent behavior.

Longer, sturdier sessions (external payloads + semantic page-out) The host atomically persists oversized tool results before submitting an External result. Core journals only the opaque locator, digest, size, and preview; read_result becomes a correlated LoadPayload effect. When pressure triggers semantic eviction, the SDK summarizes archived content into MemoryStore.

Safety and governance by default (OS native profile)
Every run loads declarative governance_policy (deny / ask_user / rate-limit / param rules) and in-kernel signal routing (signal_policy, default queue 64). Dangerous tools, external interrupts, and approval flows are policy — not ad-hoc checks in your handlers.

Long-term memory as syscalls (Phase-7)
write_memory and query_memory run outside the main tool loop: kernel validation before MemoryStore.put, search → select_memories → memory_retrieval_result on query. Failed writes emit memory_validation_failed for audit; good memory is durable without polluting history.

Multi-agent and multi-signal orchestration
Sub-agents register in the kernel process table (agent_process_changed); parent runs suspend explicitly until sub_agent_completed. Signals get disposition (Interrupt / Queue / Observe / Dropped) in-kernel, so gateways, cron, and heartbeats compose with the main loop instead of racing it.

Observable like an OS log
Page-out, signals, processes, budgets, and memory events land in SessionLog with categories. Rebuild an OS snapshot (page_out_count, process_by_agent, memory counters) from one event stream; payload residency stays in the canonical journal.

Dynamic workflows

For a task that needs more than one Agent, describe a DAG and let the runtime run fresh-context specialists under the same budgets, policies, and session evidence as the parent. See the workflow guide for the complete pattern catalog.

from deepstrike import WorkflowSpec, WorkflowNodeSpec

# One fresh-context verifier per rule (no inherited author context → can't rubber-stamp),
# then a skeptic that reviews their flags. The kernel spawns the 3 verifiers as one gated
# batch, suspends on the join, and runs the skeptic once they complete.
outcome = await runner.run_workflow(WorkflowSpec(nodes=[
    WorkflowNodeSpec(task="Rule: money is integer cents — violated?", role="verify"),
    WorkflowNodeSpec(task="Rule: all errors propagate — violated?",    role="verify"),
    WorkflowNodeSpec(task="Rule: timestamps are UTC — violated?",       role="verify"),
    WorkflowNodeSpec(task="Skeptic: which flags are real violations?",  role="verify", depends_on=[0, 1, 2]),
]))
# => {"completed": ["wf-node0", …], "failed": []}

A node's kind selects the control-flow shape; the same executor drives them all, every spawn passing the syscall gate:

Workflow capabilities

• Runtime fan-out — give a node the submit_workflow_nodes_tool and its agent can append nodes to the live DAG mid-run (true loop-until-done; one verifier per claim it discovers). Submission events remain audit projections; checkpoint state owns recovery. Governance rejection fails the submitting node instead of acknowledging work that was never appended.
• Quarantine, no escape — set trust="quarantined" on a node that reads untrusted content; it's denied write-capable isolation in-kernel, and any nodes it submits are coerced to quarantined too (no privilege escalation).
• Structured output — set output_schema on a node; the runner instructs the agent, validates the result against the JSON-Schema subset, and re-runs once with the errors on mismatch. A node that never conforms fails (its dependents starve).
• Budget as signal — with a max_workflow_nodes / max_concurrent_subagents quota installed, each spawned node's goal carries its remaining headroom so a coordinator can size its fan-out to fit.

Providers

Resource quotas are opt-in and flow through the same replayable kernel event ABI:

from deepstrike import MemoryWriteRateLimit, ResourceQuota

runner = RuntimeRunner(RuntimeOptions(
    # ...
    resource_quota=ResourceQuota(
        max_concurrent_subagents=4,
        max_spawn_depth=2,
        memory_writes_per_window=MemoryWriteRateLimit(max_writes=20, window_ms=60_000),
    ),
))

The top-level package exports the base providers OpenAIProvider, OpenAIResponsesProvider, and AnthropicProvider. Every other backend is a factory function in deepstrike.providers, with a protocol argument where a backend speaks both the OpenAI- and Anthropic-compatible wire:

from deepstrike.providers import deepseek, kimi, minimax

ds = deepseek(api_key="...")                          # OpenAI-compatible wire (default)
dsA = deepseek(api_key="...", protocol="anthropic")   # Anthropic-compatible wire
mm = minimax(api_key="...")                           # MiniMax defaults to the Anthropic wire

All providers accept retry_config for exponential backoff and share a CircuitBreaker.

extensions are forwarded to the provider while SDK-owned structural fields remain protected.

Context model (four slots)

The kernel renders context as four LLM API slots — only history is compressed.

runner = RuntimeRunner(RuntimeOptions(
    initial_memory=["User prefers chartreuse."],  # → Slot 2
    system_prompt="You are a helpful assistant.",  # → Slot 1
    # ...
))

• memory(query) / knowledge(query) meta-tool results → history (tool results)
• Inbound signals are routed by the in-kernel attention policy and rendered into Slot 3

Full reference: docs/concepts/context-slots-compression.md

Runtime options

from deepstrike import (
    DEFAULT_NATIVE_GOVERNANCE_POLICY,
    DEFAULT_SANDBOX_POLICY,
    validate_declarative_policy,
    AgentIdentity,
    AgentRunSpec,
)
from deepstrike.runtime import DEFAULT_NATIVE_SIGNAL_POLICY, PromptBudget
from deepstrike.governance import GovernancePolicy, GovernancePolicyRule

runner = RuntimeRunner(RuntimeOptions(
    provider=provider,
    session_log=FileSessionLog(".deepstrike/sessions"),
    execution_plane=plane,

    # Scheduler budget
    max_tokens=128_000,
    max_turns=25,
    timeout_ms=60_000,

    # Default governance and signal policy
    governance_policy=DEFAULT_NATIVE_GOVERNANCE_POLICY,
    signal_policy=DEFAULT_NATIVE_SIGNAL_POLICY,  # SignalRouter queue size 64
    prompt_budget=PromptBudget(
        prompt_overhead_tokens=20,
        output_reserve_tokens=4096,
        safety_margin_tokens=256,
    ),

    # Host I/O
    extensions={"temperature": 0.1},
    skill_dir="./skills",
    knowledge_source=my_ks,
    signal_source=gw,
    memory_store=my_store,
    agent_id="my-agent",
    initial_memory=["..."],

    # Memory paging & compression (SDK-side I/O)
    compression_store=archive_store,
    memory_provider=memory_llm,
    memory_summarizer=my_memory_summarizer,  # semantic page_out → MemoryStore

    # Sub-agents & milestones
    run_spec=AgentRunSpec(
        identity=AgentIdentity(agent_id="my-agent", session_id="session-1"),
        role="orchestrator",
        goal="...",  # overridden by the run() goal on canonical root start
    ),
    milestone_contract=my_contract,
    milestone_policy="require_verifier",
    on_milestone_evaluate=my_verifier,
    sub_agent_harness=SubAgentHarnessConfig(eval_provider=eval_provider, max_attempts=3),

    # Governance UX (AskUser path)
    on_permission_request=lambda req: {"approved": True, "responder": "user"},
))

Validate policies before starting a run:

result = validate_declarative_policy(
    gov_policy=DEFAULT_SANDBOX_POLICY,
    signal_policy=DEFAULT_NATIVE_SIGNAL_POLICY,
)
assert result["valid"], result["errors"]

Rebuild an OS diagnostics snapshot from session events:

from deepstrike.runtime.os_snapshot import rebuild_os_snapshot_from_session_events

events = [e.event for e in await session_log.read(session_id)]
snap = rebuild_os_snapshot_from_session_events(events)
# snap["page_out_count"], snap["signals"], …

External tool payloads

When a tool result exceeds the configured inline threshold, the SDK persists the full body before sending the canonical External result. The kernel receives only payload_ref, digest, original_size, and a bounded preview.

payload_ref is opaque and never passed to ordinary file tools. The model calls read_result; core authorizes the reachable handle and emits LoadPayload, which the runner resolves through PayloadStore.

No configuration is required. Pass a PayloadStore through RuntimeOptions.payload_store to use a different filesystem root or storage adapter.

Tools

from deepstrike import tool, read_file

plane.register(tool(name="search", description="Search.", parameters=schema)(my_fn))
plane.register(read_file)     # built-in: read files explicitly named by the caller
plane.unregister("search")

Execution planes:

Mount capabilities on an active run:

runner.mount_tool(schema)
runner.mount_skill("summarize", "Summarize text")
runner.unmount_capability("tool", "search")

Skills

Set skill_dir — the kernel auto-injects a skill meta-tool, and the LLM loads skills by name on demand.

runner = RuntimeRunner(RuntimeOptions(
    provider=provider,
    session_log=InMemorySessionLog(),
    execution_plane=plane,
    max_tokens=4096,
    skill_dir="./skills",
))

---
name: summarize
description: Summarize text into 2-3 concise bullet points
when_to_use: When you need to condense long text
effort: 1
---
1. Identify the 2-3 most important points
2. Express each as a concise bullet

Knowledge

Implement KnowledgeSource — the kernel injects a knowledge meta-tool. Runtime retrieval results land in history as tool results (single-use fact content that decays with compaction), not in the durable knowledge partition. Use initial_memory for durable preload into Slot 2.

For durable content at runtime use runner.push_knowledge(content, key=..., pinned=...) — a keyed entry upserts on a repeated key and runner.remove_knowledge(key) removes it, both applied at the next compaction/renewal boundary. knowledge_budget_ratio (default 0.25, 0 disables) caps the partition: over budget, the oldest unpinned, non-skill entries are evicted at boundaries while pinned=True entries survive. A loaded skill's body is pinned here as skill:<name> and unpinned by runner.deactivate_skill(name) or a skill_lease_turns expiry.

from deepstrike import KnowledgeSource

class VectorSearch(KnowledgeSource):
    async def init(self) -> None:
        await vector_db.connect()

    async def retrieve(self, query: str, top_k: int = 5) -> list[str]:
        return await vector_db.search(query, top_k)

runner = RuntimeRunner(RuntimeOptions(
    provider=provider,
    session_log=InMemorySessionLog(),
    execution_plane=plane,
    knowledge_source=VectorSearch(),
))

Memory

WorkingMemory (SDK-side scratch pad)

WorkingMemory is an SDK helper — not the kernel working partition. Kernel task state renders into Slot 3 (turns[0]).

from deepstrike import WorkingMemory

mem = WorkingMemory()
mem.set("step", 1)
mem.get("step")  # 1
mem.clear()

MemoryStore (long-term memory)

from deepstrike import MemoryStore

class MyStore(MemoryStore):
    async def load_sessions(self, agent_id): ...
    async def load_memories(self, agent_id): ...
    async def commit(self, agent_id, result, existing): ...
    async def search(self, agent_id, query): ...  # -> list[MemoryRecall]
    async def save_session(self, data): ...

runner = RuntimeRunner(RuntimeOptions(
    provider=provider,
    session_log=InMemorySessionLog(),
    execution_plane=plane,
    memory_store=MyStore(),
    agent_id="my-agent",  # enables memory meta-tool + semantic page-out archival
))

Three memory paths:

Custom semantic summarizer:

async def memory_summarizer(archived, ctx):
    return f"Long-term summary for action={ctx.get('action')}"

runner = RuntimeRunner(RuntimeOptions(
    # ...
    memory_store=MyStore(),
    agent_id="my-agent",
    memory_summarizer=memory_summarizer,
))

Phase-7 memory syscalls (write_memory / query_memory)

from deepstrike import MemoryProvenance, MemoryQuery, MemoryRecord, MemoryScope

scope = MemoryScope(tenant_id="acme", namespace="assistant")
await runner.write_memory(MemoryRecord(
    record_id="prefers-small-tests",
    scope=scope,
    name="prefers-small-tests",
    kind="feedback",
    content="User prefers focused unit tests for SDK behavior.",
    description="Testing preference",
    provenance=MemoryProvenance(author="host", trust="user_asserted"),
    created_at=1,
    updated_at=1,
), session_id="my-session")

hits = await runner.query_memory(MemoryQuery(
    scope=scope,
    query="Need testing preferences",
    top_k=5,
    kinds=["feedback"],
), session_id="my-session")

Session events: memory_written, memory_queried, memory_validation_failed, memory_retrieval_result.

Governance

In-kernel declarative policy (preferred)

Every run loads governance_policy into the kernel via load_governance_policy:

from deepstrike import DEFAULT_SANDBOX_POLICY
from deepstrike.governance import GovernancePolicy, GovernancePolicyRule, GovernanceRateLimit

policy = GovernancePolicy(
    rules=[
        GovernancePolicyRule(pattern="read_file", action="allow"),
        GovernancePolicyRule(pattern="write_file", action="ask_user"),
        GovernancePolicyRule(pattern="*", action="deny"),
    ],
    rate_limits=[GovernanceRateLimit(tool="api_call", max_calls=10, window_ms=60_000)],
)

runner = RuntimeRunner(RuntimeOptions(
    provider=provider,
    session_log=InMemorySessionLog(),
    execution_plane=plane,
    governance_policy=policy,
    on_permission_request=lambda req: {"approved": True, "responder": "cli"},
))

• deny → tool rejected with tool_denied
• ask_user → tool_gated + suspended; resolve via on_permission_request, then kernel resume

Default when omitted: allow-all (DEFAULT_NATIVE_GOVERNANCE_POLICY).

Standalone Governance class

Governance wraps the native governance evaluator for SDK-side use (tests, custom gates). It is not wired automatically into RuntimeRunner — use governance_policy for run-time enforcement.

from deepstrike import Governance

gov = Governance("allow")
gov.add_permission_rule("danger.*", "deny")
gov.block_tool("rm_rf")
gov.evaluate("read_file", '{"path":"x"}')

SDK PermissionManager

PermissionManager is a separate SDK-side permission layer for apps that manage their own approval UX outside the kernel loop.

from deepstrike import PermissionManager, PermissionMode

pm = PermissionManager(PermissionMode.DEFAULT)
pm.grant("fs", "read")
pm.evaluate("fs", "read")

Signals

Inbound signals are routed by the in-kernel attention policy (default queue size 64):

from deepstrike import SignalGateway, ScheduledPrompt, RuntimeSignal
from deepstrike.runtime import DEFAULT_NATIVE_SIGNAL_POLICY

gw = SignalGateway()
gw.schedule(ScheduledPrompt(goal="standup", run_at_ms=target_time))
gw.ingest(RuntimeSignal(kind="alert", payload={}, urgency="normal"))

runner = RuntimeRunner(RuntimeOptions(
    provider=provider,
    session_log=InMemorySessionLog(),
    execution_plane=plane,
    signal_source=gw,
    signal_policy=DEFAULT_NATIVE_SIGNAL_POLICY,
))

runner.interrupt()  # cooperative abort → kernel timeout path
gw.destroy()

Each routed signal produces a correlated signal_delivery_disposed session event (category: "ipc").

Sub-agents

Spawn isolated child agents through the kernel process table:

from deepstrike import AgentRunSpec, AgentIdentity
from deepstrike.providers.stream import DoneEvent

async for event in runner.spawn_sub_agent(AgentRunSpec(
    identity=AgentIdentity(agent_id="researcher-1", session_id="child-session"),
    role="explore",
    goal="Find three sources on topic X",
    isolation="worktree",
)):
    if isinstance(event, DoneEvent):
        print(event.status)

Requires an active parent run (run() / wake() in progress). The kernel emits agent_process_changed; the default SubAgentOrchestrator runs the child with a filtered execution plane and feeds sub_agent_completed back.

Attempt loop

from deepstrike import (
    AttemptLoop, AttemptRequest, RuntimeAttemptBody, StopPolicy,
    LlmEvalJudge, SubAgentHarnessConfig,
)

loop = AttemptLoop(
    body=RuntimeAttemptBody(runner),
    judge=LlmEvalJudge(eval_provider),
    stop=StopPolicy(max_attempts=3),
)
outcome = await loop.run(AttemptRequest(goal="Write a haiku"))
print(outcome.run_status, outcome.outcome, outcome.verdict)

runner = RuntimeRunner(RuntimeOptions(
    provider=provider,
    session_log=InMemorySessionLog(),
    execution_plane=plane,
    sub_agent_harness=SubAgentHarnessConfig(eval_provider=eval_provider, max_attempts=3),
))
async for event in loop.stream(AttemptRequest(goal="Write a haiku")):
    if event.type == "completed":
        print(event.outcome.run_status, event.outcome.verdict)

The default carry policy keeps one session and injects judge feedback as context. Use fresh_with_feedback or fresh_with_digest(...) only when attempt isolation is intentional.

Stream events

Import from deepstrike.providers.stream:

status: completed · max_turns · token_budget · timeout · user_abort · error · milestone_pending

Further reading

• SDK OS parity matrix
• Kernel ABI reference
• Context slots & compression