affect-kernel for Python

This directory contains the zero-runtime-dependency Python implementation of the deterministic affect-state kernel for long-lived AI characters.

python -m pip install affect-kernel

from affect_kernel import AffectState, appraise_turn

state = AffectState()
result = appraise_turn(state, "CURIOSITY")
print(result.state.mood)

The repository README contains runnable headless examples, benchmark results, limitations, and the cross-runtime contract. From a source checkout, run ./scripts/setup.sh and ./scripts/check.sh at the repository root.

The package deliberately leaves model inference, persistence, and retrieval I/O behind injected protocols. Its cross-runtime guarantee covers deterministic, non-habituating affect, appraisal, memory scoring, and presence transforms only. Prompt wording is supplied by the caller and is not part of production parity.

Adapter security and transaction contract

AffectEngine sends trusted instructions in GenerateInput.system_prompt and bounded retrieved/carried evidence in GenerateInput.untrusted_context. Model adapters must place untrusted_context in a user/tool-data channel and obey its evidence-only rule; they must never concatenate it, GenerateInput.state, or retriever output into the system prompt. Gate, retrieval, and generation request states have carried_thought=None, so the carried text is available only through the explicit untrusted field on generation.

Gate failures and unsupported intents raise by default. Applications may opt in to gate_error_mode="respond" or "silent". Built-in intents are normalized to uppercase. A caller can register custom_intents; custom labels are also normalized, must match [A-Z][A-Z0-9_]{0,63}, and intentionally receive no built-in appraisal impulse.

Domain kernels can inject a synchronous AppraisalPolicy into AffectEngine. It receives an AppraisalPolicyInput containing the current state, normalized intent/event, message, and expectation, and must return an AppraisalResult. default_appraisal_policy explicitly preserves the reference English conversational mapping implemented by appraise_turn.

EngineLimits bounds each message (16,000 characters), loaded history (200 messages / 128,000 characters), assembled system prompt (128,000 characters), and buffered model output (32,000 characters) by default. Pass a replacement EngineLimits value to tighten those ceilings for your deployment. The engine checks an offending output chunk before invoking its token callback or committing the turn.

Stores implement StateStore.transaction(conversation_id). That context must serialize the complete turn for the conversation across every engine instance, and its ConversationTransaction.commit(...) must atomically persist the state update and message batch. If gate, retrieval, generation, a token callback, or cancellation fails before commit, the store must leave both state and transcript unchanged. InMemoryStateStore is the dependency-free reference implementation.