Tested execution and compatibility

The 0.2.0.dev0 preview extends the existing circuit breaker. Package dependency minimums remain unchanged; they do not guarantee that every permitted framework release supports these execution contracts. Bare import longguard keeps framework dependencies optional.

Execution modes

Interface	Supported behavior	Boundary or unsupported mode
CircuitBreaker.check	Existing synchronous policy and additive explicit resume	Application owns model/tool dispatch and checkpoint integrity
guarded, provider converters, GuardedAgentExecutor	Existing contracts retained and regression tested	Decorator advice must be consumed by the application
add_guard_to_graph	Existing wrapping; excluded tools are gated after OPEN	One mutable guard; reset for a new run; use controlled runner for thread isolation/resume
compile_guarded_graph	Synchronous START -> agent -> tools/END, tools -> agent; saved node boundaries	Exactly two nodes; no async, streaming, subgraphs, parallel nodes, namespaces or checkpoint branching
GuardedGraph.resume	Same controller/process; matching stopped checkpoint; explicit correction and finite reflection grant	Hard caps, completed/failed runs, unknown threads, concurrent continuation and changed checkpoints rejected
add_guard_to_crew	Sequential, synchronous ReAct Crew.kickoff; public CrewAgentExecutor; separate report per kickoff	Async/batch/streaming/hierarchical/native tool batches, delegation, reasoning, planning, memory and CrewAI checkpoint continuation rejected
Callback-only CrewGuard	Existing callback API, reports and last_reflection	No automatic advice delivery or full execution guarantee; attach through add_guard_to_crew for tested recovery
Provider cleanup helper	Valid OpenAI/Anthropic dictionaries and LangChain message histories	Caller must block dispatch; running tools retain unresolved status; invalid IDs rejected

CrewAI's public CrewAgentExecutor is deprecated in the tested release. Its alternative executor is outside this preview's proven mode. A future compatibility change needs real contract tests before widening support. The adapter installs execution-scoped hooks rather than registering global handlers.

Version evidence

Local environment	Python	LangGraph	LangChain / core	CrewAI
Lockfile graph verification	3.12.14	1.2.11	1.3.18 / 1.6.1	Separately tested 1.15.22
Combined environment before CrewAI upgrade	3.12.14	1.2.14	1.4.3 / 1.6.7	1.15.22
Latest stable verification	3.12.14	1.2.14	1.4.3 / 1.6.7	1.15.25

Latest-version results and exact commands are recorded in the development evidence after verification. Only executed checks establish support. CI runs the existing core Python 3.11–3.12 Ubuntu/macOS matrix plus separate locked/newest framework jobs; configuration of a CI job is not evidence that those platforms passed locally.

Usage and side effects

Guards observe completed requests. A token/step cap trips on > and a known estimated-cost cap on >=, as before. The crossing request has already incurred usage. CrewAI's complete action callback follows its tool result: that turn's tool may already have executed before the decision. Subsequent guarded requests stop after a trip.

Actual model usage is used when available. Graph text estimates and CrewAI missing-usage zero defaults are identified in usage_warnings. Cost retains the approximate half-input, half-output token split; it is not provider billing. Guarded continuation retains previous consumption. Changing the budget or importing a report does not authorize execution.

Callbacks and arbitrary tools remain application code: LongGuard cannot undo their side effects or guarantee exactly-once behavior after an application crash. The supported graph continuation avoids replaying completed tools at its saved boundary in the same process.

Newest NumPy 2.5 typing stubs require Python 3.12 syntax. Framework jobs use strict MyPy with a Python 3.12 target; core jobs use the Python 3.11 target and locked dependencies.

Controlled continuation was verified with LangGraph InMemorySaver; other saver implementations are unverified. Pass an unwrapped graph to the controller. Do not share guarded agents or the same guard between crews. Shared provider usage counters across crews are outside tested accounting.