TheseusWrapper

The main entry point. A transparent proxy that wraps any LLM client with identity protection.

Initialization

from ct_toolkit import TheseusWrapper, WrapperConfig

# Simplest form — provider-only
wrapper = TheseusWrapper(provider="openai")

# With full config
wrapper = TheseusWrapper(
    provider="openai",
    config=WrapperConfig(
        template="finance",
        kernel_name="finance",
    )
)

# Provider string only — client created from env vars
wrapper = TheseusWrapper(provider="openai")

Methods

chat(message, *, model, system, history, **kwargs) → CTResponse

Send a message and receive an enriched response.

response = wrapper.chat(
    "What are your core values?",
    model="gpt-4o-mini",
    system="Additional context.",
    history=[{"role": "user", "content": "Hello"}],
)

validate_user_rule(rule_text) → None

Validate a rule against the kernel. Raises AxiomaticViolationError or PlasticConflictError on conflict.

endorse_rule(rule_text, *, operator_id, approval_channel, commitment_new_value) → EndorsementRecord

Initiate the Reflective Endorsement flow for a plastic conflict.

export_provenance_log() → list[dict]

Verify chain integrity and export all log entries.

resolve_llm_policy(model=None, role=None) → dict[str, Any]

Resolve effective role policy from startup capability registry + environment overrides.

propagate_policy_metadata(model=None, role=None) → dict[str, Any]

Return middleware-friendly policy metadata payload:

{
    "role": "sub",
    "environment": "test",
    "effective": {
        "tool_call": False,
        "reasoning": True,
        "multimodal": False,
        ...
    }
}

propagate_headers(*, model=None, role=None) → dict[str, str]

Generate HTTP headers for sub-agent kernel propagation.

Includes policy headers:

• X-CT-Policy-Role
• X-CT-Policy-Environment
• X-CT-Policy (base64 JSON payload)

Properties

Property	Type	Description
kernel	ConstitutionalKernel	The active kernel
compatibility	CompatibilityResult	Template + kernel compatibility result
divergence_engine	DivergenceEngine	The active divergence engine
compression_guard	ContextCompressionGuard	The active context compression guard
staged_manager	StagedUpdateManager	Tracks staged (cooldown) endorsements

CTResponse fields

Field	Type	Description
content	str	Response text
provider	str	Provider used
model	str	Model used
divergence_score	float \| None	L1 score (0.0–1.0)
divergence_tier	str \| None	ok, l1_warning, l2_judge, l3_icm, critical
provenance_id	str \| None	UUID of the log entry
raw_response	Any	Raw SDK response object
sandbox_divergence	float \| None	Divergence score of the shadow/sandbox agent (if staged)