Context Compiler Directive Drafter

Turn natural-language requests into candidate Context Compiler directives.

context-compiler-directive-drafter helps hosts translate user requests like:

Please use Docker for container examples.

into candidate directives, such as:

use docker

This package drafts suggestions for the Context Compiler. Only context-compiler applies directives and updates state.

The drafter suggests candidate directives. context-compiler decides what to do with them.

The drafter owns the human-facing acquisition step between messy user input and canonical directive text. That includes deciding when a message is close enough to propose a canonical directive, when the message is not a directive at all, and when the message is too unclear or malformed to safely interpret without more help. It does not become an authority over state, permissions, or application.

When To Use It

Use this package when you want to:

• Translate user requests into safe, canonical directives.
• Handle near-canonical input, alternate phrasing, and malformed-but-recoverable directive attempts before compiler handoff.
• Distinguish terminal rejection from semantic uncertainty in a stable host-facing contract.
• Avoid accidental candidate drafts from ambiguous input.
• Add a conservative natural-language-to-directive step before applying changes.

This package owns the human-facing acquisition boundary, including exact canonicalization, bounded deterministic rewrites, obvious non-directive rejection, and fallback eligibility for semantically uncertain interpretation. It does not decide whether a drafted candidate is valid, applicable, allowed, contradictory, or executable; Core owns those decisions.

The normative acquisition contract lives in docs/DrafterAcquisitionSpec.md. This package does not own:

• authoritative compiler state
• the decision about whether a canonical directive is allowed in the current context
• directive application or state mutation
• invention of new directive semantics beyond the compiler-owned contract

Installation

Install in your host environment:

pip install "context-compiler-directive-drafter"

For local development:

uv sync --group dev

Basic Usage

Draft a candidate directive:

from context_compiler_directive_drafter import DirectiveDrafter
from context_compiler_directive_drafter import RejectedDirective, UnknownDirective

drafter = DirectiveDrafter()

result = drafter.draft_directive(
    "Please use Docker for container examples.",
)

if hasattr(result.result, "text"):
    print("Candidate directive:", result.result.text)
elif isinstance(result.result, RejectedDirective):
    print("Directive acquisition rejected:", result.result.reason)
elif isinstance(result.result, UnknownDirective):
    print("Need clarification before drafting:", result.result.reason)

The host validates drafted output before passing a CanonicalDirective candidate to engine.apply_directive(...) for authoritative evaluation and application.

For a small runnable example, see examples/basic_usage.py.

English Evaluation Corpus

The reusable English-language acquisition and fallback evaluation corpus lives in evals/corpus/english/directive-drafter-en.jsonl. It contains heuristic contracts, fallback evaluation cases, and selected fixture-backed anchors. The corpus is evaluation data only; existing conformance fixtures remain the executable compatibility authority. See docs/EnglishEvaluationCorpus.md for the schema, classifications, domain scope, and promotion workflow.

OpenAI-Compatible Fallback

Install the optional OpenAI-compatible integration extra:

pip install "context-compiler-directive-drafter[openai]"

For LiteLLM routing across providers, install the separate optional extra:

pip install "context-compiler-directive-drafter[litellm]"

Use it with OpenAI by creating a fallback callback and passing it to the Drafter:

import os

from context_compiler_directive_drafter import DirectiveDrafter
from context_compiler_directive_drafter.fallbacks.openai import create_openai_fallback

fallback = create_openai_fallback(
    model="gpt-4o-mini",
    api_key=os.environ["OPENAI_API_KEY"],
)
drafter = DirectiveDrafter(fallback=fallback, fallback_source="openai")

The asynchronous factory performs the same construction-time probe and must be awaited:

from context_compiler_directive_drafter.fallbacks.openai import create_async_openai_fallback

fallback = await create_async_openai_fallback(model="gpt-4o-mini")

OpenAI-compatible providers use the same helper with a custom endpoint, for example an Ollama server:

fallback = create_openai_fallback(
    model="qwen2.5:7b",
    api_key="ollama",
    base_url="http://localhost:11434/v1",
)

The helper probes structured-output support once while it is constructed, then selects the package-owned converter prompt and transport mode for all callback calls. Structured mode owns the JSON Schema response envelope; unsupported providers use free-text mode, which translates the profile's FallbackProfile.abstention_sentinel abstention value to the generic fallback callback contract's Python None value. Other model text is returned unchanged. The Drafter remains responsible for parsing, validation, and result shaping.

LiteLLM provides the same fallback contract across its supported providers:

from context_compiler_directive_drafter.fallbacks.litellm import create_litellm_fallback

fallback = create_litellm_fallback(model="anthropic/claude-sonnet-4-5")

Pass LiteLLM's provider/model identifier unchanged. The adapter uses LiteLLM's structured-schema capability signal first. When that signal is false, it probes the exact package-owned structured request; a successful probe selects structured mode, while an explicitly unsupported probe selects free text. Other provider/runtime errors remain ordinary errors.

Native fallback integrations

Hosts using a provider-specific client can implement the fallback callback directly while reusing the package-owned acquisition contract:

from context_compiler_directive_drafter.fallbacks import (
    FallbackProfile,
    InvalidFallbackResponseError,
    get_fallback_profile,
)

Call get_fallback_profile() for free-text acquisition, or get_fallback_profile(structured_output=True) when the provider supports structured output. The returned FallbackProfile contains the package-owned system prompt, transport mode, and (for structured output) the response schema. Return None for a valid provider rejection, and raise InvalidFallbackResponseError for a malformed or inconsistent structured response. The callback returns raw candidate text; the Drafter passes it through Core parsing and produces the non-authoritative result.

The namespaced fallbacks.DraftFallback and fallbacks.AsyncDraftFallback contracts accept the original user_input and return candidate directive text or None; the asynchronous form returns an awaitable with the same value contract. These callbacks are generic acquisition hooks and do not require prompts or schemas. DirectiveDrafter owns Core grammar validation and DraftResult construction after the callback returns.

Live English Corpus Runner

The live runner executes selected corpus cases through the real DirectiveDrafter path. Install the optional OpenAI dependency first, then run a small subset before a larger evaluation:

# OpenAI, using OPENAI_API_KEY from the environment
uv run python -m evals.runners.directive_drafter_en \
  --model gpt-4o-mini \
  --case-id en-software-preference-001

# Any OpenAI-compatible endpoint
uv run python -m evals.runners.directive_drafter_en \
  --model qwen2.5:7b \
  --api-key ollama \
  --base-url http://localhost:11434/v1 \
  --domain software_development \
  --category preference_statement \
  --limit 3

# LiteLLM provider/model routing
uv run python -m evals.runners.directive_drafter_en \
  --transport litellm \
  --model anthropic/claude-sonnet-4-5 \
  --limit 3

# Write detailed JSONL records to the ignored eval-results directory
uv run python -m evals.runners.directive_drafter_en \
  --model gpt-4o-mini \
  --limit 10 \
  --output eval-results/sample.jsonl

The runner accepts the existing OpenAI-compatible transport by default. Use --transport litellm to pass a LiteLLM provider/model identifier unchanged through the LiteLLM fallback adapter. The runner reports totals, heuristic versus fallback routing, domain and category breakdowns, and failure categories. It does not change the corpus expectations or act as a conformance authority. JSONL records include the actual path, path agreement, fallback invocation count, and raw fallback response; semantic scoring and routing mismatches are reported separately.

Conformance layers

The checked-in contract fixtures use the existing pytest.mark.contract and fixture model. high-level-drafting-v1.json defines shared API concepts and observable drafting behavior for compatible ports, including the DirectiveDrafter, result variants, fallback routing, source metadata, public rejection reasons, and the native fallback integration capabilities. The native integration contract defines the exact rendered artifacts for each contractual fallback profile: the system prompt, structured schema when applicable, and abstention representation. Restricted-kind variants are rendered deterministically from the canonical rules. Free-text adapters obtain the exact abstention value from FallbackProfile.abstention_sentinel; it is not a separately exported sentinel symbol. Structured adapters represent abstention as rejected with a JSON null output. Provider adapters may wrap these artifacts in provider-specific transport, but must not rewrite them. Changes to a contractual artifact require a conformance/version change. The contract also requires fallback callbacks to receive original user input and return candidate text or no candidate. It does not require an OpenAI-specific factory. public-api-v1.json defines Python package mechanics such as root exports, reflection-visible signatures, descriptors, and Python exception behavior; those details are not requirements for other languages.

Heuristic fixture inputs and portable expected outcomes are shared behavior. Private Python entry points and detailed internal diagnostic reasons are Python implementation coverage. The English evaluation corpus remains evaluation-only except for cases explicitly linked to promoted heuristic fixtures.

Public API

Public interface:

• DirectiveDrafter(): Synchronous orchestration over heuristic preprocessing, optional fallback acquisition, fallback output parsing and validation, and final result construction.
• DraftResult: Structured non-authoritative result returned by DirectiveDrafter.draft_directive(...).
• context_compiler_directive_drafter.fallbacks: Namespaced public callback contracts, FallbackProfile, and get_fallback_profile(...) for generic fallback acquisition.
• RejectedDirective and UnknownDirective: Non-canonical drafting result variants with preserved reasons.
• create_openai_fallback(...): Create a synchronous OpenAI-compatible fallback callback.
• create_async_openai_fallback(...): Create an asynchronous OpenAI-compatible fallback callback.
• context_compiler_directive_drafter.fallbacks: Public fallback profiles, callback contracts, structured-response parsing, and invalid-response helpers for native fallback integrations.
• context_compiler_directive_drafter.fallbacks.openai: Optional OpenAI-compatible fallback factories.
• context_compiler_directive_drafter.fallbacks.litellm: Optional LiteLLM fallback factories using LiteLLM's provider/model routing.

Output Contract

The intended drafting boundary is:

• input: user text
• output: DraftResult(source=<final producer>, result=<drafting-layer variant>)

Every drafting path should end in one of three host-visible result variants:

• CanonicalDirective: a proposed canonical directive that is ready for compiler review and independent policy checks
• RejectedDirective(reason=...): acquisition is terminally rejected and must not reach fallback
• UnknownDirective(reason=...): the input appears directive-related or interpretation failed, but the drafter should not guess

The heuristic uses these outcomes deliberately:

• RejectedDirective means acquisition is terminally rejected, including ordinary prose, questions, quoted or reported commands, incomplete directives, and compound or malformed input. It is never fallback-eligible. Its stable public reasons are non_directive, incomplete, multiple_directives, and invalid_candidate.
• UnknownDirective means semantic interpretation remains plausible but the heuristic cannot confidently produce one candidate. This result alone is eligible for host fallback interpretation.

Failure to recognize canonical syntax or a bounded rewrite is not, by itself, evidence for RejectedDirective.

As a bounded deterministic rewrite, a clear whole-message I prefer X form is treated like please use X and produces the proposed candidate use X when the payload reduces to one canonical directive. Questions, mixed explanations, and multiple preference statements remain unresolved. This does not apply to general evaluative language such as Docker would be better.

The source field records only the final producer of the returned drafting result, such as heuristic or the source metadata configured for a host-provided fallback acquisition callback. It does not track fallback history.

A returned CanonicalDirective means "this is a proposed canonical directive," not "this directive is permitted" and not "this directive has been applied."

Recommended Host Flow

1. Run DirectiveDrafter().draft_directive(message) as the high-level drafting API. It always tries heuristic drafting first and may optionally call a non-authoritative fallback acquisition callback when the heuristic returns UnknownDirective.
2. If you configure a fallback, have it return canonical directive text or None, and register the source metadata you want preserved on any fallback-produced DraftResult.
3. If the result yields a CanonicalDirective, pass that canonical directive to context-compiler for authoritative review and application.
4. If the result yields RejectedDirective, continue the host flow without a directive handoff.
5. If the result yields UnknownDirective, preserve the boundary: ask for clarification, show resubmission guidance, or retry drafting in a safer workflow.

Custom fallback callbacks receive the original user input and return raw text or None. Provider callbacks own provider response parsing; DirectiveDrafter owns candidate normalization, Core validation, and result construction.

Safety Guidance:

• Always validate drafting output before compiler handoff.
• Never pass raw model output directly to the compiler.
• Bypass drafting when clarification is pending.
• Do not drive authoritative transitions from package-owned drafting code.
• Do not read or mutate engine.state directly from package-owned drafting code.
• Prefer abstaining when the input cannot be confidently reduced to one directive.
• Output validation checks the canonical directive contract, not whether the directive is allowed in context.
• A structurally valid drafted directive may still be the wrong interpretation of the user's meaning.
• Reviewed semantic drafting belongs in a separate higher-level workflow such as preview, approval, or engine application.

Hosts may use UnknownDirective to trigger clarification, confirmation, or resubmission guidance. That interaction is part of the human-input drafting boundary, but any eventual canonical directive must still be revalidated before compiler handoff.

Do not pass raw model output to the compiler.

Current Limits

This package is intentionally conservative. It returns UnknownDirective only when a single acquisition unit is directive-adjacent but remains semantically uncertain:

• The input may be ambiguous or not safely interpretable as one canonical directive.
• The input may be embedded in prose, markdown, or code when the heuristic cannot establish a terminal non-directive boundary.

Boundary rules:

• Process the full message, not fragments.
• Treat one sentence or one directive request as the acquisition unit. The drafter does not perform conversational sentence segmentation.
• Emit at most one canonical directive.
• Reject when one message contains multiple directive-shaped instructions.
• Do not mine surrounding prose for commands.
• Do not split one message into multiple drafted directives.
• Do not invent new directive semantics.
• Apply only bounded deterministic rewrites that preserve one apparent directive operation.
• Defer ambiguous semantic interpretation to the host fallback when available.

Obvious multi-sentence conversational input is outside the acquisition unit. The heuristic returns rejected for that input rather than sending it to fallback. The host is responsible for sentence segmentation and may resubmit the resulting units individually. unknown is reserved for an eligible single acquisition unit that is directive-adjacent but cannot be confidently reduced to one canonical candidate and may require fallback. Core remains responsible for canonical directive validity, applicability, authorization, and execution; it does not own host sentence segmentation.

Questions are terminal acquisition rejections. Forms such as allow docker?, do not use peanuts?, please use docker?, and can you help with lunch? return rejected; they never reach fallback or become heuristic candidates.

Quoting is intentionally distinct from quoting an operand: a full-message command such as "use docker" is treated as quoted or reported text and is terminally rejected, while use "docker" is a canonical use candidate whose operand is the literal quoted text. The heuristic does not strip operand quotes.

context-compiler-directive-drafter only proposes at most one candidate directive. context-compiler remains responsible for independently enforcing the single-directive invariant before any authoritative application.

The drafter should consume the compiler-owned grammar contract once that extracted contract is available. This package should not duplicate or become the normative owner of grammar rules in its own documentation or prompt resources.

Hosts that want broader proposal behavior should implement it explicitly.

CLI

The CLI command is directive-drafter. The CLI currently supports a limited set of behaviors:

uv run directive-drafter "please make replies concise"

It returns a non-zero exit status because the public high-level drafting API requires a host-provided engine context.

Development

Run local checks:

uv run pre-commit run --all-files
uv run pytest

License

Apache-2.0