context-compiler 0.5.0

Deterministic conversational state engine for LLM applications.

Context Compiler

A deterministic directive engine that converts explicit user instructions into structured conversational state for LLM applications.

Modern language models reason well but are unreliable at maintaining consistent state across interactions.

Corrections compete with earlier statements, constraints disappear, and long conversations accumulate contradictions.

The Context Compiler introduces a deterministic state layer that governs authoritative conversational state independently of the model.

The model performs reasoning and generation while the compiler manages premise and policies. Once accepted, directives remain authoritative until explicitly corrected or reset.

Evidence (cross-model runs)

• Models tested: llama3.1:8b, gpt-4o-mini, gpt-4.1, gpt-5, claude-sonnet-4, claude-opus-4
• Baseline path: 2–4 / 6 pass across runs
• compiler path: 6 / 6 pass across runs
• compiler+compact path: 6 / 6 pass across runs (after compact-path fixes)
• Demo 6 context reduction: up to 99%
• Demo 6 prompt reduction: about 50%

Why “Compiler”?

Context Compiler treats explicit user directives as inputs to a deterministic process.

Instead of relying on the LLM to remember constraints across a conversation, user instructions are compiled into structured state before the model runs.

The idea is similar to a traditional compiler: user directives are translated into a structured representation that the rest of the system can rely on.

Installation

• Python 3.11+
• pip install context-compiler
• Dev/test: uv sync --group dev and uv run pytest

---

10-Second Example

User sets a constraint once:

User: prohibit peanuts

Outcome: policy state includes "peanuts": "prohibit".

Later in the conversation:

User: how should I make this curry?

The host supplies the authoritative state to the model so the constraint persists across turns.

---

Architecture

User Input
     │
     ▼
Context Compiler
     │
     ▼
Decision
     │
     ▼
Host Application
 ├─ clarify → ask user
 ├─ passthrough → call LLM
 └─ update → call LLM with compiled state

The compiler governs authoritative conversational state and never calls the LLM. The host decides whether to call the model based on the returned Decision.

---

Decision API

Each user message produces a Decision.

class Decision(TypedDict):
    kind: Literal["passthrough", "update", "clarify"]
    state: dict | None
    prompt_to_user: str | None

Meaning:

kind	host behavior
passthrough	forward user input to LLM
update	forward input with updated state
clarify	show prompt_to_user and do not call the LLM

Host Integration Example

engine = create_engine()
decision = engine.step(user_input)

if decision["kind"] == "clarify":
    show_to_user(decision["prompt_to_user"])
else:
    state = decision["state"] or engine.state
    messages = build_messages(state, user_input)
    render(call_llm(messages))

API Reference

API	Description
create_engine(state=None)	Create a new compiler engine; optional state provides initial authoritative state (validated/canonicalized).
step(user_input)	Parse one user turn and return a deterministic Decision.
compile_transcript(messages)	Replay a transcript from a fresh engine and return either final state or a confirmation prompt.
engine.apply_transcript(messages)	Replay a transcript onto the current engine state and return either final state or a confirmation prompt.
engine.state	Read current authoritative in-memory state snapshot.
get_premise_value(state)	Read the current premise value from a state snapshot.
get_policy_items(state, value=None)	Read policy items from a state snapshot (all, use, or prohibit).
engine.export_json()	Export current state as JSON for persistence/transport.
engine.import_json(payload)	Load/restore state from exported JSON payload.

---

State Model

The compiler maintains an authoritative state snapshot. Hosts should treat this state as structured application data and avoid coupling to internal field names or nested layout.

State Access and Persistence

Hosts can provide initial state at engine creation (create_engine(state=...)), read current in-memory state via engine.state, and persist/restore via engine.export_json() and engine.import_json(). Semantic state mutations occur through directives processed by step(). Storage is managed by the host application.

Use the returned state snapshot as structured host input for prompt construction, policy enforcement, or replay/storage workflows. For host code that needs typed reads without direct nested key lookups, use get_premise_value(state) and get_policy_items(state, value=...).

Transcript Replay

Transcript replay compiles conversational history by reusing the same deterministic directive path:

• Only messages with role == "user" are processed.
• Assistant/system/non-user messages are ignored.
• Replay calls step() for each user message in order.
• Replay stops on the first clarification and returns a confirmation prompt.
• compile_transcript(messages) starts from a fresh engine.
• engine.apply_transcript(messages) applies replay onto the current engine state.

State Properties

• Premise is a single slot (set once, then replaced via change premise to ...)
• Policies are per-item and exclusive by item (use or prohibit)
• No inference or semantic reasoning

Identical input sequences always produce identical compiler state. LLM responses may still vary unless deterministic decoding is used by the host.

Premise replacement and policy mutation are always explicit via directives. The engine does not infer intent from conversational wording outside the documented grammar.

---

Directive Examples

Set and change premise explicitly:

User: set premise concise replies
User: change premise to concise bullet points

Result: premise is updated deterministically via explicit premise lifecycle commands. Narrow near-miss premise to variants clarify (no mutation): set premise to <value> and change premise <value>.

Per-item policy directives:

User: use docker
User: prohibit peanuts

Result: policies store authoritative per-item states (use / prohibit). Recognized empty payloads clarify without mutation: use, use , prohibit, and prohibit .

Explicit replacement:

User: use podman instead of docker

Result: remove docker use policy and set podman to use (or clarify if blocked). Incomplete replacement payloads also clarify without mutation (for example: use x instead of, use instead of y, use instead of y).

Policy contradiction (clarify):

User: use peanuts
User: prohibit peanuts

Result: compiler asks for clarification and leaves state unchanged.

Policy Removal and Reset Commands

Policy cleanup commands are:

• remove policy <item> removes one policy item if present (idempotent update if absent)
• reset policies clears all policy items
• clear state resets premise and policies to initial values

Single-policy correction flow:

User: prohibit peanuts
User: remove policy peanuts
User: use peanuts

Example:

• If policies include {"docker": "use", "peanuts": "prohibit"}:
• after remove policy docker, policies become {"peanuts": "prohibit"}.
• after reset policies, policies become {}.

---

Examples

• examples
• demos
• integrations

---

Quickstart

Run the interactive REPL:

context-compiler

Run an example:

python examples/01_persistent_guardrails.py

Run tests:

uv run pytest

---

Guarantees

• State changes only through explicit user directives or confirmation.
• Identical input sequences produce identical compiler state.
• Model responses never modify compiler state.
• Ambiguous directives trigger clarification instead of changing state.

These invariants are verified through behavioral tests and Hypothesis-based property tests.

---

Design Notes

More detailed design and milestone documents are available in:

• Project overview
• Directive grammar specification

---

License

Apache-2.0.