Client¶

The SDKClient class (also exported as CRPClient) is the main entry point for the CRP progressive SDK. This page combines hand-written examples with auto-generated reference docs from the source code.

Quick example¶

import crp

client = crp.SDKClient()
response = client.complete("Summarise the EU AI Act")
print(response.text)
print(response.crp.risk)   # LOW | MEDIUM | HIGH | CRITICAL
print(response.crp.grounded)
print(response.crp.fabrications)

Auto-generated reference¶

`crp.sdk.client.CRPClient` `dataclass` ¶

Progressive-disclosure SDK client (SPEC-032).

Parameters:

Name	Type	Description	Default
`provider`	`LLMProvider \| None`	An LLMProvider instance (optional - can set later).	`None`
`config`	`CRPConfig \| None`	A CRPConfig instance (optional - loads defaults or crp.config.yaml).	`None`
`safety`	`str \| dict[str, Any]`	Safety profile name ("balanced", "strict", etc.) or dict.	`'balanced'`
`depth`	`str`	Default query depth ("auto", "quick", "standard", "thorough", "exhaustive").	`'auto'`
`app_profile`	`ApplicationProfile \| None`	Optional application capability contract used to auto-select provider, relay strategy, and context behaviour.	`None`

`storage` `property` ¶

Access storage visibility API.

`knowledge` `property` ¶

Access knowledge location.

`audit` `property` ¶

Access the tamper-evident compliance audit trail.

`compliance` `property` ¶

Access EU AI Act / ISO 42001 compliance helpers.

`safety` `property` `writable` ¶

Access the Safety Control Plane (SPEC-033, SPEC-034).

`ckf` `property` ¶

Access the Contextual Knowledge Fabric (SPEC-009, SPEC-025).

`cso` `property` ¶

Access the Cognitive State Object (SPEC-030).

`provenance` `property` ¶

Access the Decision Provenance Engine (SPEC-005).

`reasoning` `property` ¶

Access reasoning scaffolds, CQS, and cross-window validation.

`activation` `property` ¶

Access CRP activation-mode detection (SPEC-017).

`agent` `property` ¶

Access multi-agent safety budget and chain tools (SPEC-012).

`events` `property` ¶

Access the protocol event bus (§9).

`providers` `property` ¶

Access LLM provider registration (SPEC-008).

`extract` `property` ¶

Access the graduated extraction pipeline (§2.5).

`gateway` `property` ¶

Access CRP Gateway helpers (SPEC-016).

`headers` `property` ¶

Access the CRP HTTP header surface (SPEC-002).

`observability` `property` ¶

Access observability subsystems (audit, metrics, telemetry).

`policy` `property` ¶

Access the safety policy engine (SPEC-006).

`scan` `property` ¶

Access CRP Scan helpers (SPEC-013, SPEC-036, SPEC-039).

`comply` `property` ¶

Access CRP Comply helpers (SPEC-040, SPEC-042, SPEC-047, SPEC-048).

`orchestrator` `property` ¶

Access the live CRPOrchestrator instance directly.

This exposes every public orchestrator method and subsystem (e.g. client.orchestrator.dispatch(...), client.orchestrator.ckf).

`modules` `property` ¶

Access any public class or function in the crp package.

This dynamic mirror lets you reach the full CRP API without importing modules manually::

client.modules.envelope.cdr.cdr_rank(...)
client.modules.security.consent.ConsentManager(...)

`core` `property` ¶

Access core orchestrator, session, DAG, ledger, and facilitator.

`continuation` `property` ¶

Access continuation manager and helpers.

`envelope` `property` ¶

Access envelope builder, packer, reranker, CDR, and formatter.

`state` `property` ¶

Access warm store, cold storage, snapshots, event log, and router.

`security` `property` ¶

Access safety manifest, consent, RBAC, checkpoints, and audit trail.

`resources` `property` ¶

Access adaptive allocator, cost model, overhead, and resource manager.

`advanced` `property` ¶

Access curator, feedback, meta-learning, source grounding, and validator.

`cli` `property` ¶

Access CLI sidecar handler and startup result types.

`errors` `property` ¶

Access public CRP exception classes.

`init(provider=None, config=None, safety='balanced', depth='auto', model=None, api_key=None, input_continuation_mode=None, app_profile=None)` ¶

Manual init to allow a safety property on a dataclass.

`configure(**kwargs)` ¶

Apply runtime configuration overrides (Layer 5).

Unknown keys are stored in the unified config. Keys that map to core orchestrator settings are also forwarded when possible.

`complete(prompt, system='You are a helpful assistant.', **kwargs)` ¶

Single-turn completion with automatic governance (Level 0).

Returns:

Type	Description
`CRPCompletionResponse`	CRPCompletionResponse with `.crp` governance summary.

`stream(prompt, system='You are a helpful assistant.', *, depth=None, **kwargs)` ¶

Stream a single-turn completion as StreamEvent objects.

    Each event has ``event_type`` (``token``, ``extraction``,
    ``continuation``, ``window_complete``, ``done``, ``error``) and
    ``data``. Concatenating all ``token`` data values produces the same
    text as :meth:`complete`.

    Args:
        depth: Override depth (``quick`` / ``standard`` / ``thorough`` /
            ``exhaustive``).

    Example::

        for event in client.stream("Explain CRP."):
            if event.event_type == "token":
                print(event.data, end="")
            elif event.event_type == "done":
                print("

Done")

`ingest(path)` ¶

Ingest documents into the Contextual Knowledge Fabric.

Accepts a file path, directory path, or list of paths.

`ask(question, system='You are a helpful assistant.', depth=None, **kwargs)` ¶

Multi-turn quality-aware query with source attribution (Level 1–2).

Parameters:

Name	Type	Description	Default
`depth`	`str \| None`	Override depth for this query ("quick", "standard", "thorough", "exhaustive"). Level 2 control.	`None`

Returns:

Type	Description
`CRPAskResponse`	CRPAskResponse with quality tier, sources, completeness, and
`CRPAskResponse`	inspectable reasoning (Level 2).

`tool(fn)` ¶

Decorator: register a tool for tool-mediated dispatch (Level 2).

Usage::

@client.tool
def get_metrics(service: str) -> dict:
    return {"cpu": 0.5}

`call_tool(name, *args, **kwargs)` ¶

Call a registered tool by name (Level 2).

`dispatch_positioned(user_request, *, fabric=None, executor=None, profile=None, governor=None, clarify_handler=None, policy=None, oversight_required=None, context_facts=None, max_operations=12, temperature=0.2, max_tokens=1024, prior_cso=None, max_continuation_windows=1)` ¶

Run the positioned-tool-loop (CRP-SPEC-049/050) with this client's provider.

Positioning, not injection: each operation positions the model on only the 1–3 tools the protocol selected for it - never the full catalogue. Any @client.tool functions are auto-registered as capabilities when no explicit fabric is supplied.

For multi-turn workflows, pass the previous result's .cso as prior_cso so the follow-up turn is positioned with everything the prior turn established (facts, tool observations, decisions).

Returns a PositionedResult exposing .text, .cso, .event_stream, .observation_count, .headers, and .halted.

`conversation(**defaults)` ¶

Start a multi-turn positioned conversation that manages state for you.

Each .say(request) runs the positioned loop with the running CSO relayed forward automatically - no manual prior_cso threading::

convo = client.conversation()
r1 = convo.say("Look up the service on port 443.")
r2 = convo.say("Based on that, is it safe for a login page?")

defaults are passed to every dispatch_positioned call (e.g. profile=, max_continuation_windows=, clarify_handler=).

`derive_profile(messages, tools=None)` ¶

Derive an application capability contract from observed messages.

`session()` ¶

Return a live view of the current CRP session.

`save_config(path)` ¶

Save the current unified config to path.

Parameters:

Name	Type	Description	Default
`path`	`str \| Path`	Destination file path.	required

`config_hash()` ¶

Return a deterministic hash of the loaded unified config.

Returns:

Type	Description
`str`	Hex digest string, or empty string if hashing is unavailable.

`aenter()` `async` ¶

Async context manager entry - not yet implemented.

`aexit(exc_type, exc, tb)` `async` ¶

Async context manager exit - not yet implemented.

`reset()` ¶

Clear accumulated session facts (warm store + CKF).

`close()` ¶

Release orchestrator resources.

Response types¶

`crp.sdk.response.CRPCompletionResponse` `dataclass` ¶

Level 0 response - drop-in governance (SPEC-032 §2).

`crp.sdk.response.CRPAskResponse` `dataclass` ¶

Level 1+2 response - quality-aware with inspectable reasoning (SPEC-032 §3–4).

`crp.sdk.response.CRPResponseMeta` `dataclass` ¶

The one governance object - five fields, not 58 headers (SPEC-032 §2.3).

Provider selection¶

SDKClient auto-detects the LLM provider from the environment:

Source	Provider selected
`OPENAI_API_KEY` set	OpenAI
`ANTHROPIC_API_KEY` set	Anthropic
Ollama reachable on localhost	Ollama

Override explicitly when needed:

client = crp.SDKClient(provider="openai", model="gpt-4o")

Session management¶

s = client.session()
print(s.id)
print(s.status())
print(s.fact_count)
print(s.window_count)

Configuration¶

client.configure(safety_profile="strict")

Safety profiles adjust the trade-off between autonomy and oversight. See Safety and compliance and Configuration for details.

Client¶

Quick example¶

Auto-generated reference¶

crp.sdk.client.CRPClient dataclass ¶

storage property ¶

knowledge property ¶

audit property ¶

compliance property ¶

safety property writable ¶

ckf property ¶

cso property ¶

provenance property ¶

reasoning property ¶

activation property ¶

agent property ¶

events property ¶

providers property ¶

extract property ¶

gateway property ¶

headers property ¶

observability property ¶

policy property ¶

scan property ¶

comply property ¶

orchestrator property ¶

modules property ¶

core property ¶

continuation property ¶

envelope property ¶

state property ¶

security property ¶

resources property ¶

advanced property ¶

cli property ¶

errors property ¶

__init__(provider=None, config=None, safety='balanced', depth='auto', model=None, api_key=None, input_continuation_mode=None, app_profile=None) ¶

configure(**kwargs) ¶

complete(prompt, system='You are a helpful assistant.', **kwargs) ¶

stream(prompt, system='You are a helpful assistant.', *, depth=None, **kwargs) ¶

ingest(path) ¶

ask(question, system='You are a helpful assistant.', depth=None, **kwargs) ¶

tool(fn) ¶

call_tool(name, *args, **kwargs) ¶

dispatch_positioned(user_request, *, fabric=None, executor=None, profile=None, governor=None, clarify_handler=None, policy=None, oversight_required=None, context_facts=None, max_operations=12, temperature=0.2, max_tokens=1024, prior_cso=None, max_continuation_windows=1) ¶

conversation(**defaults) ¶

derive_profile(messages, tools=None) ¶

session() ¶

save_config(path) ¶

config_hash() ¶

__aenter__() async ¶

__aexit__(exc_type, exc, tb) async ¶

reset() ¶

close() ¶

Response types¶

crp.sdk.response.CRPCompletionResponse dataclass ¶

crp.sdk.response.CRPAskResponse dataclass ¶

crp.sdk.response.CRPResponseMeta dataclass ¶

Provider selection¶

Session management¶

Configuration¶

`crp.sdk.client.CRPClient` `dataclass` ¶

`storage` `property` ¶

`knowledge` `property` ¶

`audit` `property` ¶

`compliance` `property` ¶

`safety` `property` `writable` ¶

`ckf` `property` ¶

`cso` `property` ¶

`provenance` `property` ¶

`reasoning` `property` ¶

`activation` `property` ¶

`agent` `property` ¶

`events` `property` ¶

`providers` `property` ¶

`extract` `property` ¶

`gateway` `property` ¶

`headers` `property` ¶

`observability` `property` ¶

`policy` `property` ¶

`scan` `property` ¶

`comply` `property` ¶

`orchestrator` `property` ¶

`modules` `property` ¶

`core` `property` ¶

`continuation` `property` ¶

`envelope` `property` ¶

`state` `property` ¶

`security` `property` ¶

`resources` `property` ¶

`advanced` `property` ¶

`cli` `property` ¶

`errors` `property` ¶

`init(provider=None, config=None, safety='balanced', depth='auto', model=None, api_key=None, input_continuation_mode=None, app_profile=None)` ¶

`configure(**kwargs)` ¶

`complete(prompt, system='You are a helpful assistant.', **kwargs)` ¶

`stream(prompt, system='You are a helpful assistant.', *, depth=None, **kwargs)` ¶

`ingest(path)` ¶

`ask(question, system='You are a helpful assistant.', depth=None, **kwargs)` ¶

`tool(fn)` ¶

`call_tool(name, *args, **kwargs)` ¶

`dispatch_positioned(user_request, *, fabric=None, executor=None, profile=None, governor=None, clarify_handler=None, policy=None, oversight_required=None, context_facts=None, max_operations=12, temperature=0.2, max_tokens=1024, prior_cso=None, max_continuation_windows=1)` ¶

`conversation(**defaults)` ¶

`derive_profile(messages, tools=None)` ¶

`session()` ¶

`save_config(path)` ¶

`config_hash()` ¶

`aenter()` `async` ¶

`aexit(exc_type, exc, tb)` `async` ¶

`reset()` ¶

`close()` ¶

`crp.sdk.response.CRPCompletionResponse` `dataclass` ¶

`crp.sdk.response.CRPAskResponse` `dataclass` ¶

`crp.sdk.response.CRPResponseMeta` `dataclass` ¶