deepstrike 0.2.19

No project description provided

DeepStrike Python SDK

Runtime framework built on a Rust kernel. The kernel owns loop control, context compression, governance, signal routing, and memory paging — the SDK owns all I/O (LLM calls, tool execution, disk, long-term memory).

Python is a first-class SDK for the Agent OS native profile: declarative governance and in-kernel signal routing are enabled by default on every run.

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"

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 · DreamStore │
└───────────────────────────┬─────────────────────────────┘
                            │ step(JSON event) ↔ actions / observations
┌───────────────────────────▼─────────────────────────────┐
│  deepstrike._kernel KernelRuntime                       │
│  P1 Syscall · P2 Sched · P3 MM · Proc · IPC             │
└─────────────────────────────────────────────────────────┘

The runner drives a single loop:

1. Kernel returns an action — call_provider, execute_tool, evaluate_milestone, or done.
2. SDK executes the action (stream LLM, run tools, call milestone verifier).
3. SDK feeds the result back as a kernel event (provider_result, tool_results, …).
4. Kernel observations (compression, page-out, spool, signals, …) are drained into SessionLog.

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

What Agent OS gives you

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 (Layer-1 spool + semantic page-out)
Oversized tool results (> 50 KB) stay in context as a preview plus a .spool/ reference — the model reads the full payload on demand via ordinary file tools. When pressure triggers semantic eviction, the SDK summarizes archived content into DreamStore and satisfies page_in_requested on the way back in. Long tasks survive token pressure instead of failing mid-run.

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 (attention_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 DreamStore.commit, 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
Spool, page-out, signals, processes, budgets, and memory events land in SessionLog with categories. Rebuild an OS snapshot (page_out_count, spool_count, process_by_agent, memory counters) from one event stream — replay still strips audit events when reconstructing LLM messages.

You need…	Use…
Policy before tools run	governance_policy (default: allow-all native profile)
External interrupts	signal_source + in-kernel attention_policy
Spawn / memory-write quotas	resource_quota (set_resource_quota)
Huge tool output	Automatic Layer-1 spool; optional custom result_spool
Durable recall across runs	DreamStore + semantic page_out via dream_summarizer
Programmatic memory I/O	runner.write_memory() / runner.query_memory()
Debug / compliance	SessionLog events + OS snapshot helpers

---

Dynamic workflows

Instead of planning and executing a hard task in one long context window, hand the kernel a declarative DAG and let it spawn a fresh-context sub-agent per node. The kernel owns the control flow (gate · budget · suspend-on-join · resume); your SDK runs the agents. See the top-level overview for the full 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:

Node kind	Behavior
spawn (default)	Run the node's agent once
loop (max_iters)	Re-run until the agent signals it's done, capped at max_iters
classify (branches)	The classifier's result selects one branch; the rest are pruned
tournament (entrants)	Generate N entrants, then a pairwise-judge bracket to one winner
reduce (reducer)	Tokenless host-compute — a pure function (dedupe_lines / merge_json_arrays / concat / count, or your own via the reducers option) over the node's dependency outputs

0.2.11 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). Recorded and replayed on resume_workflow.
• 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),
    ),
))

Class	Backend	Notes
OpenAIProvider	OpenAI API	SSE tool-call accumulation
AnthropicProvider	Anthropic API	Native SSE, ThinkingDelta support
QwenProvider	DashScope	enable_thinking via extensions
DeepSeekProvider	DeepSeek API	Reasoner models strip tools automatically
MiniMaxProvider	MiniMax API	M1 reasoning via expose_reasoning
OllamaProvider	Local Ollama	http://localhost:11434 default
KimiProvider	Moonshot API

All providers accept RetryConfig 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.

Slot	Source	Role
system_stable	system partition	Identity, rules — never changes within a run
system_knowledge	knowledge partition	Preloaded memory, skill defs — low frequency
turns[0]	task_state + signals	Goal, plan, progress, compression log, runtime signals
turns[1..N]	history	Conversation transcript

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_ATTENTION_POLICY
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,

    # Agent OS native profile (defaults shown)
    governance_policy=DEFAULT_NATIVE_GOVERNANCE_POLICY,
    attention_policy=DEFAULT_NATIVE_ATTENTION_POLICY,  # SignalRouter queue size 64

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

    # Memory paging & compression (SDK-side I/O)
    compression_store=archive_store,
    dream_provider=dream_llm,
    dream_summarizer=my_dream_summarizer,  # semantic page_out → DreamStore

    # Sub-agents & milestones
    run_spec=AgentRunSpec(
        identity=AgentIdentity(agent_id="my-agent", session_id="session-1"),
        role="orchestrator",
        goal="...",  # overridden by run() goal on start_run
    ),
    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"},
))

Option	Purpose
governance_policy	Declarative deny / ask_user / rate-limit / param rules loaded into the kernel before start_run
attention_policy	In-kernel signal router queue size (default 64)
on_permission_request	Resolves tool_gated + suspended → kernel resume with approved/denied call IDs
compression_store	Writes archived messages on compressed observations
dream_summarizer	Summarizes page_out { tier_hint: "semantic" } into DreamStore during a run
dream_provider	Separate LLM for dream() idle consolidation (falls back to provider)
result_spool	Custom large-result spool (default: .spool/ under cwd)

Validate policies before starting a run:

result = validate_declarative_policy(
    gov_policy=DEFAULT_SANDBOX_POLICY,
    attention_policy=DEFAULT_NATIVE_ATTENTION_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["spool_count"], snap["signals"], …

---

Large result spool (Layer 1)

When a single tool result exceeds 50 KB, the kernel keeps a short preview in context and emits large_result_spooled. The SDK writes the full payload to .spool/ under the process cwd and logs spool_ref in the session.

LocalExecutionPlane transparently resolves read-tool arguments that point at .spool/ paths:

# Kernel context shows a preview + spool reference.
# LLM calls read_file(path=".spool/abc123…") → full content returned.

No configuration is required. Pass a custom result_spool on RuntimeOptions to change the directory (see tests/test_semantic_page_out_dream.py and spool-related tests).

---

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 (also resolves .spool/ refs)
plane.unregister("search")

Execution planes:

Plane	Use case
LocalExecutionPlane	In-process tools (default)
FilteredExecutionPlane	Capability-filtered sub-agent tools
ProcessSandboxPlane	OS subprocess isolation
McpProxyPlane	MCP server tools
RemoteVpcPlane	Remote execution

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. Use initial_memory for durable preload into Slot 2.

Before tool execution the kernel may emit page_in_requested; the SDK satisfies it from DreamStore, KnowledgeSource, and a local semantic page-out cache, then feeds page_in back to the kernel.

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()

DreamStore (long-term memory)

from deepstrike import DreamStore

class MyStore(DreamStore):
    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, top_k): ...
    async def save_session(self, data): ...

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

Three memory paths:

Path	When	What happens
In-session memory(query)	LLM calls meta-tool	DreamStore.search() → history tool result
initial_memory	Run start	Injected into Slot 2 (system_knowledge)
Semantic page_out	Kernel evicts with tier_hint: "semantic"	SDK summarizes via dream_summarizer / dream_provider → DreamStore.commit()
dream(agent_id)	Explicit idle call	IdlePipeline batch-consolidates past sessions

import time
from deepstrike.providers.stream import DoneEvent

async for event in runner.dream("my-agent", now_ms=int(time.time() * 1000)):
    if isinstance(event, DoneEvent):
        print(event.dream_result)

Custom semantic summarizer:

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

runner = RuntimeRunner(RuntimeOptions(
    # ...
    dream_store=MyStore(),
    agent_id="my-agent",
    dream_summarizer=dream_summarizer,
))

Phase-7 memory syscalls (write_memory / query_memory)

await runner.write_memory({
    "metadata": {
        "name": "prefers-small-tests",
        "description": "User prefers focused unit tests",
        "kind": "feedback",
        "created_at": 1,
        "updated_at": 1,
    },
    "content": "User prefers focused unit tests for SDK behavior.",
}, session_id="my-session")

hits = await runner.query_memory({
    "current_context": "Need testing preferences",
    "active_tools": [],
    "already_surfaced": [],
    "top_k": 5,
}, 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):

Urgency	Typical disposition
critical / high	interrupt_now — may yield a new call_provider action
normal / low	queue — buffered; no action until dequeued
queue full	dropped

from deepstrike import SignalGateway, ScheduledPrompt, RuntimeSignal
from deepstrike.runtime import DEFAULT_NATIVE_ATTENTION_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,
    attention_policy=DEFAULT_NATIVE_ATTENTION_POLICY,
))

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

Each routed signal produces a signal_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.

---

Harness (evaluation framework)

from deepstrike import (
    SinglePassHarness, EvalLoopHarness, HarnessLoop, HarnessRequest,
    SubAgentHarnessConfig, QualityGate,
)

outcome = await SinglePassHarness(runner).run(HarnessRequest(goal="Say hello"))

class ContainsHello(QualityGate):
    async def evaluate(self, request, outcome) -> bool:
        return "hello" in outcome.result.lower()

outcome = await EvalLoopHarness(runner, gate=ContainsHello(), max_attempts=3).run(req)

loop = HarnessLoop(runner, eval_provider=eval_provider, max_attempts=3, skill_dir="./skills")

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.run_streaming(HarnessRequest(goal="Write a haiku")):
    if event.type == "done":
        print(event.verdict.passed, event.verdict.feedback)

---

Stream events

Import from deepstrike.providers.stream:

Class	Key fields
TextDelta	delta
ThinkingDelta	delta
ToolCallEvent	id, name, arguments
ToolDeltaEvent	call_id, name, delta, chunk?
ToolSuspendEvent	call_id, name, suspension_id, payload?
ToolResultEvent	call_id, content, is_error
PermissionRequestEvent	tool_name, reason
DoneEvent	iterations, total_tokens, status
ErrorEvent	message

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

---

Further reading

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