ADR 2: Steppable CEK machine

Date: 2022-10


Marty Stumpf
Ziyang Liu




In order to have a minimal viable product of a debugger for Plutus, we need a CEK machine that will give us more information for debugging than our current one.

In order to provide debugging information for each evaluation step, we need a steppable CEK machine. Implementing the steppable CEK machine is a non-trivial task and involves some design decisions. One decision to make is about whether we can share the code between the production and the debugging machine. That is not the scope of this ADR. See the next ADR for that.

This ADR proposes a design for an implementation of a steppable CEK machine. Of course, this doesn’t mean that this is the final decision. This means that the next step for us is to prototype the machine in this way - which we have reasons to believe will go well. We may adjust our proposed approach depending on how the prototyping goes.


This section describes the proposed implementation of the debugging machine.

We first abstract out the computation to “steps” on our current machine. We then implement a coroutine system to add the debugging functionalities.

Abstracting out the computation to “steps”

This abstraction has been implemented in PR#4909.

The current machine inlined the steps. We separate each steps into separate functions. They all return a CekState:

data CekState uni fun =
    -- the next state is computing
    Computing WordArray (Context uni fun) (Closure uni fun)
    -- the next state is returning
    | Returning WordArray (Context uni fun) (CekValue uni fun)
    -- evaluation finished
    | Terminating (Term NamedDeBruijn uni fun ())

data Closure uni fun = 
  Closure (Term NamedDeBruijn uni fun ()) (CekValEnv uni fun)

The computing step is computeCekStep with the following signature:

    :: forall uni fun s
    . (Ix fun, PrettyUni uni fun, GivenCekReqs uni fun s)
    => WordArray
    -> Context uni fun
    -> Closure uni fun
    -> CekM uni fun s (CekState uni fun)

Similarly for the returning step (returnCekStep). Then we link up all the steps with continue, and the machine behaves very similar to our current one:

continue :: forall uni fun s
    . (Ix fun, PrettyUni uni fun, GivenCekReqs uni fun s)
    => CekState uni fun
    -> CekM uni fun s (Term NamedDeBruijn uni fun ())
continue (Computing !unbudgetedSteps ctx (Closure term env)) = do
    state <- computeCekStep unbudgetedSteps ctx env term
    continue state
continue (Returning !unbudgetedSteps ctx val) = do
    state <- returnCekStep unbudgetedSteps ctx val
    continue state
continue (Terminating term) = pure term

Coroutines in Haskell

The next step is to add debugging capabilities between each step. To do so, we implement it as a coroutine system. A detailed introduction to coroutines in Haskell can be found in Coroutine Pipelines. This section gives a brief summary.

A coroutine system is composed of multiple computations cooperatively passing data and control to one another. In this instance, one computation is the user issuing commands like “step forward”, and the other is the CEK machine processing the commands and performing actions like interpreting the script being debugged. We’ll refer to them as the “user computation” and the “machine computation” respectively.

Coroutines in Haskell can be implemented using the free monad transformer, FreeT. The Coroutine type used in the above article is isomorphic to FreeT.

To use FreeT f m, we need two things: a suspension functor f, and a base monad m.

The suspension functor is a pattern functor that describes the ways the user computation can suspend and pass control to the machine computation. Each constructor of the suspension functor should thus represent a user request, such as “step forward”. Constructors generally follow a RequestType request (response -> a) pattern.

As an example, consider the following suspension functor (the uni and fun parameters are omitted for readability):

data RequestF a
  = StepF CekState (CekState -> a)
  | LogF Text a
  | InputF (Command -> a)
  deriving Functor

StepF passes a CekState to the machine computation and suspends, requesting the machine computation to progress one step, and send a CekState back. LogF sends a Text to the machine computation (its response type is () and is omitted). InputF requests a Command from the user.

Note that this pattern is not limited to a single suspension functor and two computations. Multiple suspension functors and computations can be composed using coproducts.

The base monad m is the monad the machine computation runs in. The machine computation interprets each request into an m action. It is essentially a natural transformation from the suspension functor to m. This m will replace our current monad CekM. Although we can actually just use CekM in the steppable CEK machine when we add IO capabilities for debugging. This is because we can convert it to/from IO via unsafeSTToIO and unsafeIOToST.

Suppose we define a type SteppableCekM a as our base monad m. Then the machine computation can be implemented as the following request handler function:

handle :: RequestF a -> SteppableCekM a
handle = \case
  StepF state k -> step state >>= pure . k
  LogF text k  -> log text >> pure k
  InputF k     -> input >>= pure . k

where step state, log text and input return SteppableCekM actions. step will likely correspond to computeCekStep and returnCekStep depending on the states.

We can then use handle to construct a monad morphism, interpreting the user computation (a FreeT structure) into a SteppableCekM action:

runSteppableCek :: FreeT RequestF SteppableCekM a -> SteppableCekM a
runSteppableCek userAction = do
  runFreeT userAction >>= \case
    Pure res -> pure res
    Free req -> handle req >>= runSteppableCek

To construct the user computation, FreeT RequestF SteppableCekM, we first provide helper functions for constructing RequestFs and lifting them into the FreeT:

stepF :: Monad m => CekState -> FreeT RequestF m CekState
stepF state = liftF (StepF state id)

logF :: Monad m => Text -> FreeT RequestF m ()
logF text = liftF (LogF text ())

inputF :: Monad m => FreeT RequestF m Command
inputF = liftF (InputF id)

then we can implement the user computation like this:

userComputation :: CekState -> FreeT RequestF SteppableCekM ()
    userComputation currentState = do
      cmd <- inputF
      case cmd of
        Step -> do
          logF "Received Step command"
          mState <- stepF currentState
          userComputation mState

We enter the debugging mode with the input UPLC program or term to debug with enterDebug:

enterDebug :: UPLCTerm -> FreeT RequestF SteppableCekM ()
enterDebug termToDebug = do
  state <- stepF (Computing (toWordArray 0) NoFrame (Closure term Env.empty))
  userComputation state

Argument: coroutine system

Why a coroutine system? In short, structuring the code this way will ease our future work. Some of the advantages are mentioned above already. Here is a summary:

  • The debugger is naturally a coroutine, where one routine is the user and the other is the CEK machine, and they take turns to suspend and pass data and control to each other in a debugging session. The literature has contributed a good way to design/implement a coroutine. It makes sense to implement a well studied design.

  • We can probably reuse the same monad (CekM) in the steppable CEK machine, because we can convert it to/from IO via unsafeSTToIO and unsafeIOToST.

  • It should be easier when we add more functionalities because multiple suspension functors and computations can be composed using coproducts.

  • This should also play nicely when we implement Debug Adapter Protocol for the debugger later on.


In summary, we proposed to implement the debugging machine as a coroutine system with “steps”. This implies that:

  • We have to maintain the CEK machine. E.g., we need to check its conformance.

  • We will add a debugger for our users. We can give users more information at each evaluation step.

  • We will need to write some tests to ensure that the debugging machine continuously output reasonable information.