The `intent_guard.py` module#

Summary#

Classes

`GuardFinding`	Single signature hit on a snippet.
`GuardReport`	Collection of findings for a snippet.

Functions

`is_enabled`	Honor `FLUIDS_MCP_INTENT_GUARD` (default ON).
`evaluate`	Evaluate code against intent-guard risk signatures.

Description#

Static, Fluent-specific risk classifier for run_code snippets.

This module is stateless. It does NOT plan, schedule, or track history. It only inspects a single snippet’s AST against a fixed table of Fluent-specific crash signatures and (optionally) performs a single intra-snippet topological reorder when use-before-create is detected for a small set of well-known dependencies (such as named expressions and materials).

By design this module:

Contains no model orchestration and no rule-pack DSL. Only a fixed table of code shapes that have empirically crashed PyFluent/the gRPC channel during in-house sessions.
Is opt-out via FLUIDS_MCP_INTENT_GUARD=0 so a host that prefers its own planner can disable it entirely.
Never imports ansys.fluent.core. It pure ast walks only so that it is safe to evaluate before a solver session exists.
Returns the same RunCodeResult envelope the rest of the leaf uses, with one of three new error_code values: risk_blocked, sequence_error, and solver_disconnected. (The third is set by pyfluent backend, not here.)

This is a defense-in-depth layer at the execution boundary, in scope for the standalone leaf. It is NOT a planner. Higher-level planning, recipes, and journaling remain out of scope.

Module detail#

intent_guard.is_enabled(env: dict[str, str] | None = None) → bool#

Honor FLUIDS_MCP_INTENT_GUARD (default ON).

Parameters:

envdict[str, str], optional: Environment mapping to inspect. When omitted, os.environ is used.

Returns:

bool: False only when the environment value is 0, false, no, or off. Otherwise, True.

intent_guard.evaluate(code: str, *, live_named_objects: Mapping[str, Iterable[str]] | None = None, iterating: bool = False) → GuardReport#

Evaluate code against intent-guard risk signatures.