Skip to content

Dridus/techtonica

Repository files navigation

What are this?

It's a calculator/planner toolbox similar to e.g. Satisfactory Calculator but without a web app and GUI and for Techtonica.

Why?

What, you don't play video games this way?

It was fun is the actual answer. Techtonica isn't (wasn't?) popular enough to have a complete wiki or calculation assistance, so why not way overdo it?

How do?

  1. direnv allow or nix develop to enter the dev shell
  2. cabal repl, import Tech.Ghci

Commands

The Tech.Ghci module implements a simple CLI for mucking about, and maintains state in the form of some global IORefs. Note this means if you :reload or exit the GHCi REPL all your state goes away without warning, so make sure to saveFactory.

General

  • history :: IO ()

    Show up to last 50 changes to state as well as any undone states that could be redo-ed.

  • undo :: IO ()

    Revert to the previous state.

  • redo :: IO ()

    Un-undo.

Factory manipulation

  • printFactory :: IO ()

    Print the current factory's static structure, i.e. the machines and belts specified.

    E.g.:

      ghci> printFactory
      ┏━ 1: cluster of 1 planter/kindlevine: (130.000sec/cycle) 1 kindlevineSeed >-> 1 kindlevine
      ┠┄┄┄ 1 ↘ 2: belt carrying kindlevine
      ┣━ 2: cluster of 1 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      ┠┄┄┄ 2 ↘ 3: belt carrying kindlevineSticks
      ┗━ 3: cluster of 1 thresher/kindlevineSticks: (3.000sec/cycle) 1 kindlevineSticks >-> {4 kindlevineExtract, 4 plantmatterFiber}
      ghci> 
    
  • saveFactory :: FilePath -> IO ()

    Save the current factory out as a YAML file.

    E.g.:

      ghci> saveFactory "test.yaml"
      ghci> 
    
  • loadFactory :: FilePath -> IO ()

    Load a factory back from a YAML file, and also verify it.

    E.g.:

      ghci> clearFactory
      ghci> loadFactory "test.yaml"
      ghci> printFactory
      ┏━ 1: cluster of 1 planter/kindlevine: (130.000sec/cycle) 1 kindlevineSeed >-> 1 kindlevine
      ┠┄┄┄ 1 ↘ 2: belt carrying kindlevine
      ┣━ 2: cluster of 1 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      ┠┄┄┄ 2 ↘ 3: belt carrying kindlevineSticks
      ┗━ 3: cluster of 1 thresher/kindlevineSticks: (3.000sec/cycle) 1 kindlevineSticks >-> {4 kindlevineExtract, 4 plantmatterFiber}
      ghci> 
    
  • setFactory :: FactorySt -> IO ()

    Replace the current factory state in GHCi with some specific factory.

    E.g.:

      ghci> clearFactory
      ghci> setFactory Tech.TestFixtures.linearSt
      ghci> printFactory
      ┏━ 1: cluster of 1 test/a1b1: (1.000sec/cycle) 1 testA >-> 1 testB
      ┠┄┄┄ 1 ↘ 2: belt carrying testB
      ┗━ 2: cluster of 1 test/b1c1: (1.000sec/cycle) 1 testB >-> 1 testC
      ghci> 
    
  • clearFactory :: IO ()

    Big delete.

  • verifyFactory :: IO ()

    Do various sanity checks on the current factory graph. Currently that means ensuring all belt edges have valid connections, both that they connect to valid nodes and that those nodes produce or consume the expected things.

    E.g.:

      ghci> printFactory
      ┏━ 1: cluster of 1 planter/kindlevine: (130.000sec/cycle) 1 kindlevineSeed >-> 1 kindlevine
      ┠┄┄┄ 1 ↘ 2: belt carrying kindlevine
      ┣━ 2: cluster of 1 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      ┠┄┄┄ 2 ↘ 3: belt carrying kindlevineSticks
      ┗━ 3: cluster of 1 thresher/kindlevineSticks: (3.000sec/cycle) 1 kindlevineSticks >-> {4 kindlevineExtract, 4 plantmatterFiber}
      ghci> verifyFactory
      Verify OK!
    
      ghci> addBelt 2 3 ironComponents
      ghci> printFactory
      ┏━━━━ 1: cluster of 1 planter/kindlevine: (130.000sec/cycle) 1 kindlevineSeed >-> 1 kindlevine
      ┠┄┄┄┄┄┄ 1 ↘ 2: belt carrying kindlevine
      ┣━━┳━ 2: cluster of 1 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      ┠┄┄┃┄┄┄ 2 ↘ 3: belt carrying kindlevineSticks
      ┃  ┠┄┄┄ 2 ↘ 3: belt carrying ironComponents
      ┗━━┻━ 3: cluster of 1 thresher/kindlevineSticks: (3.000sec/cycle) 1 kindlevineSticks >-> {4 kindlevineExtract, 4 plantmatterFiber}
      ghci> verifyFactory
      Warning: Belt downstream recipe does not accept requested item.
        2 ↘ 3: belt carrying ironComponents
        3: cluster of 1 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      Warning: Belt upstream recipe does not produce requested item.
        2 ↘ 3: belt carrying ironComponents
        2: cluster of 1 thresher/kindlevineSticks: (3.000sec/cycle) 1 kindlevineSticks >-> {4 kindlevineExtract, 4 plantmatterFiber}
      Verified with 2 warning(s).
    
      ghci> 
    
  • addCluster :: MachineIdentifier -> RecipeIdentifier -> Quantity -> IO Node

    Add a cluster (of machines) to the current factory. The first parameter is IO Recipe so that findRecipe fits in there. You can pure some recipe you already have let-bound also.

    E.g.:

      ghci> addCluster "planter" "kindlevine" 1
      4
    
  • editCluster :: Node -> (ClusterSt -> ClusterSt) -> IO ()

    Edit an existing cluster in-place without detaching any belts as deleting it and recreating it would.

    E.g.:

      ghci> loadFactory "test.yaml"
      ghci> printFactory
      ┏━ 1: cluster of 1 planter/kindlevine: (130.000sec/cycle) 1 kindlevineSeed >-> 1 kindlevine
      ┠┄┄┄ 1 ↘ 2: belt carrying kindlevine
      ┣━ 2: cluster of 1 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      ┠┄┄┄ 2 ↘ 3: belt carrying kindlevineSticks
      ┗━ 3: cluster of 1 thresher/kindlevineSticks: (3.000sec/cycle) 1 kindlevineSticks >-> {4 kindlevineExtract, 4 plantmatterFiber}
      ghci> editCluster 2 (set quantity 2)
      ghci> printFactory
      ┏━ 1: cluster of 1 planter/kindlevine: (130.000sec/cycle) 1 kindlevineSeed >-> 1 kindlevine
      ┠┄┄┄ 1 ↘ 2: belt carrying kindlevine
      ┣━ 2: cluster of 2 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      ┠┄┄┄ 2 ↘ 3: belt carrying kindlevineSticks
      ┗━ 3: cluster of 1 thresher/kindlevineSticks: (3.000sec/cycle) 1 kindlevineSticks >-> {4 kindlevineExtract, 4 plantmatterFiber}
      ghci> 
    
  • delCluster :: Node -> IO ()

    Remove a cluster from the current factory.

  • addBelt :: Node -> Node -> Item -> IO ()

    addBelt upstream downstream item

    Add a belt carrying the given item between an upstream cluster and downstream cluster.

    E.g.:

      ghci> clearFactory
      ghci> addCluster (findRecipe planter "kindlevine") 1
      1
      ghci> addCluster (findRecipe thresher "kindlevine") 1
      2
      ghci> addBelt 1 2 kindlevine
      ghci> printFactory
      ┏━ 1: cluster of 1 planter/kindlevine: (130.000sec/cycle) 1 kindlevineSeed >-> 1 kindlevine
      ┠┄┄┄ 1 ↘ 2: belt carrying kindlevine
      ┗━ 2: cluster of 1 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      ghci> 
    

Factory Planning

  • estimateFactory :: IO ()

    Estimate the flows of the current factory configuration and output them.

    E.g.:

      ghci> loadFactory "test.yaml"
      ghci> printFactory
      ┏━ 1: cluster of 1 planter/kindlevine: (130.000sec/cycle) 1 kindlevineSeed >-> 1 kindlevine
      ┠┄┄┄ 1 ↘ 2: belt carrying kindlevine
      ┣━ 2: cluster of 1 thresher/kindlevine: (6.000sec/cycle) 1 kindlevine >-> {1 kindlevineSeed, 4 kindlevineSticks}
      ┠┄┄┄ 2 ↘ 3: belt carrying kindlevineSticks
      ┗━ 3: cluster of 1 thresher/kindlevineSticks: (3.000sec/cycle) 1 kindlevineSticks >-> {4 kindlevineExtract, 4 plantmatterFiber}
      ghci> estimateFactory
      ┏━━━━━━━ external resources ↓ 0.461/min kindlevineSeed
      ┠┄┄┄┄┄┄┄┄┄ external ↘ 1: belt carrying kindlevineSeed (0.461/min >-> 0.461/min)
      ┣━━━━━━━ 1: cluster of 1 planter/kindlevine: 0.461/min kindlevineSeed >-> 0.461/min kindlevine
      ┠┄┄┄┄┄┄┄┄┄ 1 ↘ 2: belt carrying kindlevine (0.461/min >-> 0.461/min)
      ┣━━┳━━━━ 2: cluster of 1 thresher/kindlevine: 9.999/min kindlevine >-> {9.999/min kindlevineSeed, 39.999/min kindlevineSticks}
      ┃  ┠┄┄┄┄┄┄ 2 ↘ 3: belt carrying kindlevineSticks (39.999/min >- short 39.999/min -> 79.999/min)
      ┠┄┄┃┄┄┄┄┄┄ 2 ↘ external: belt carrying kindlevineSeed (9.999/min >-> 9.999/min)
      ┃  ┣━━┳━ 3: cluster of 1 thresher/kindlevineSticks: 19.999/min kindlevineSticks >-> {79.999/min kindlevineExtract, 79.999/min plantmatterFiber}
      ┃  ┠┄┄┃┄┄┄ 3 ↘ external: belt carrying plantmatterFiber (79.999/min >-> 79.999/min)
      ┃  ┃  ┠┄┄┄ 3 ↘ external: belt carrying kindlevineExtract (79.999/min >-> 79.999/min)
      ┗━━┻━━┻━ byproducts ↓ {79.999/min kindlevineExtract, 9.999/min kindlevineSeed, 79.999/min plantmatterFiber}
      ghci> 
    
  • proposeFactory :: Image Rate -> ProposalConstraints Last -> IO ()

    Propose additions to the factory structure such that the resulting factory produces as byproducts all of the items given.

    E.g.:

      ghci> clearFactory 
      ghci> proposeFactory (One ("ironComponents", 1)) (assumeGiven (Just "ironIngot"))
      Generated 1 proposals, 1 successful, 0 not.
      Proposal 1 succeeded after 1 steps (add assembler/ironComponents to get 60.000/min ironComponents for goal)
        
          Structure:
              new 1: cluster of 2 assembler/ironComponents: (4.000sec/cycle) 1 ironIngot >-> 2 ironComponents
          Estimated:
             ┏━ external-1 ↓ 30.000/min ironIngot
             ┠┄┄┄ external-1 ↘ 1: belt carrying ironIngot (30.000/min >-> 30.000/min)
             ┣━ 1: cluster of 2 assembler/ironComponents: 30.000/min ironIngot >-> 60.000/min ironComponents
             ┠┄┄┄ 1 ↘ external-2: belt carrying ironComponents (60.000/min >-> 60.000/min)
             ┗━ byproducts-2 ↓ 60.000/min ironComponents
    

    As implied by the example, additional constraints can be given to control how much is added versus assumed and whether overflows in the factory are proposed to be wired into new construction.

    ProposalConstraints Last is a Monoid, so you use <> to combine constraints. Constraints currently understood:

    • assumeGiven :: Foldable f => f Item -> ProposalConstraints Last

      Assume the named item(s) are available as feedstock. If an item isn't assumed given, then the planner will try to find recipes to produce the intermediate items required, and fail if there are no such recipes.

    • doUseOverflows, dontUseOverflows :: ProposalConstraints Last

      When looking for required intermediate items do or do not (the default is do not) consider overflowing belts and propose splitting those belts to consume the overflow in newly proposed machines.

      This is not enabled by default because the resulting factory might be messier. It's also possibly very wrong as at the moment the estimator naïvely splits products evenly among consumers which means some belts will be over and some under, when in fact all product is being used.

    Multiple proposals might be generated in the case where multiple recipes are found which produce a required item. proposeFactory only prints the first but also prints how many proposals in total were generated. Use printProposals to print them all, or printProposal to print proposals one at a time.

    A proposal can be accepted with acceptProposal. Proposals can't be cleared (rejected) explicitly, but will become invalidated if other changes are made to the factory state. If changes are made and the proposals invalidated, you have to proposeFactory again if you want revised proposals.

  • printProposals :: IO ()

    Print all proposals previously generated by proposeFactory.

  • printProposal :: Int -> IO ()

    Print a single 1-indexed proposal previously generated by proposeFactory.

  • acceptProposal :: Int -> IO ()

    Accept a single 1-indexed proposal previously generated by proposeFactory.

    If you regret, see also undo.

Items, Machines, and Recipes

  • loadEnv :: FilePath -> IO ()

    Load items, machines, and recipes from a YAML file.

  • saveEnv :: FilePath -> IO ()

    Save current items, machines, and recipes to a YAML file.

  • listItems :: IO ()

    List all registered items.

  • addItem :: Item -> IO ()

    Add an item to the set of known items.

  • delItem :: Item -> IO ()

    Delete an item to the set of known items.

  • listMachines :: IO ()

    List all machines currently in state.

  • addMachine :: MachineIdentifier -> IO ()

    Add a machine with the standard parallelism of 1.

  • addMachine' :: MachineIdentifier -> Rational -> IO ()

    Add a machine with a specific parallelism.

  • editMachine :: MachineIdentifier -> (Machine -> Machine) -> IO ()

    Edit a machine in-place. Not different from delMachine and addMachine, just more convenient.

  • delMachine :: MachineIdentifier -> IO ()

    Delete a machine along with all recipes associated with it.

  • findRecipes :: (Recipe -> Bool) -> IO ()

    List all recipes which match some predicate.

  • listAllRecipes :: IO ()

    List all recipes currently in state, for all machines.

  • listRecipes :: Machine -> IO ()

    List all recipes currently in state, for a single machine.

  • addRecipe :: Machine -> RecipeIdentifier -> NominalDiffTime -> Transfer Quantity -> IO ()

    addRecipe machine recipeId cycleTime transferFunction

    Register a recipe so you can look it up by identifier later.

    You don't need to addRecipe a recipe, it just registers the name. You can also just mention the recipe when adding a cluster.

  • editRecipe :: Machine -> RecipeIdentifier -> (Recipe -> Recipe) -> IO ()

    Edit a recipe in-place. Not different from delRecipe and addRecipe other than being more convenient.

  • delRecipe :: Machine -> RecipeIdentifier -> IO ()

    Like addRecipe, but more violent. And with fewer questions.

Machines?

:browse Tech.Machine

Items?

listItems. They come from recipes.yaml.

Recipes?

listAllRecipes or listRecipe. They come from recipes.yaml.

Data Model?

Scalars

  • Item represents a single type of item, e.g. ironIngot. They are defined in Tech.Items though you can make up new ones as you like. newtype over Text.

  • Machine represents a single type of machine, e.g. assembler. They are defined in Tech.Machines. newtype over Text.

  • RecipeIdentifier identifies a particular recipe. They are unique within the scope of a particular machine. Nothing will prevent you from having two recipes with the same identifier and machine but distinct parameters, other than zalgo. newtype over Text.

  • Quantity represents a unit quantity of an item, usually per machine cycle. newtype over Data.Fixed.Micro.

  • Rate represents a quantity of items per unit time, typically per second internally. The prettyprinter shows them divided to per minute rates usually.

FactorySt, FactoryProp, and FactoryDy

Factories come in several flavors:

  • FactorySt: Static structure which describes the clusters (machines) and connections between them, but doesn't include information about flow.
  • FactoryDy: Dynamic flow which augments FactorySt with flow information. Tech.Planner.estimate is the primary way to go from static to dynamic. There isn't a corresponding reverse, though the FactoryDy is a strict superset of the corresponding FactorySt.
  • FactoryProp: Proposed static structure which augments FactorySt with flags indicating that clusters or belts are new relative to some original FactorySt. factoryStFromProp discards the IsNew flag.

Factories are FGL directed acyclic graphs with ClusterSt (or ClusterDy) nodes representing clusters of identical machines and BeltSt (BeltDy) edges which represent belts carrying items between clusters.

BeltSt and BeltDy

Belts are fairly straightforward at the moment. Statically they identify which item they carry, and belts with mixed items aren't supported at the moment.

Dynamically they also have the estimated item flow rate measured as entering items/sec (printed as /min) and exiting, so they show overpressure and underpressure.

ClusterSt and ClusterDy

Clusters are described by a parallelism quantity representing the number of instances of the machine in question, along with a Recipe which controls how much they consume of what items and identifies the machine.

Dynamically, a cluster has its item rates per time calculated versus the recipe expressing it in terms of full unit quantities and a cycle time.

Recipe, Transfer, Image

Recipe names a Transfer function and how long the cycle time of that transfer function in a particular machine.

Recipes are keyed with the RecipeKey type, which has a Machine and RecipeIdentifier. Note that equality and ordering for Recipe is defined and in terms of this key. Zalgo might come to visit if you change a recipe's definition but not its key.

Transfer functions are Images of their input and outputs and typically created with pattern synonyms. Images are in turn maps of Item to some quantity q, which is typically Quantity in Recipes and Rate for ClusterDy or BeltDy.

Transfers and thus Recipes are typically constructed using a pattern synonyms to simplify down from the general case of Map Item Quantity -> Map Item Quantity:

ghci> putDocLn . ppTransfer ppQuantity $ [(ironIngot, 1)] :>>: [(ironComponents, 1)]
1 ironIngot >-> 1 ironComponents
ghci> putDocLn . ppTransfer ppQuantity $ [(kindlevine, 1)] :>>: [(kindlevineSticks, 1), (kindlevineSeed, 1)]
1 kindlevine >-> {1 kindlevineSeed, 1 kindlevineSticks}
recipe :: RecipeIdentifier -> NominalDiffTime -> Transfer Quantity -> Recipe
recipe rid = Recipe (RecipeKey thresher rid)

recipe' :: (Item, Quantity) -> NominalDiffTime -> [(Item, Quantity)] -> Recipe
recipe' src@(i, _) time outs = recipe (RecipeIdentifier . unItem $ i) time ([src] :>>: outs)

kindlevine, kindlevineSticks :: Recipe
kindlevine = recipe' (I.kindlevine, 1) 6.00 (Pair (I.kindlevineSeed, 1) (I.kindlevineSticks, 4))
kindlevineSticks = recipe' (I.kindlevineSticks, 1) 3.00 (Pair (I.kindlevineExtract, 4) (I.plantmatterFiber, 4))

The Transfer synonym is :>>: and takes a list of pairs as input and output.

  • (i, q) :->-: (i, q): one item (type) in, one item out.
  • (i, q) :->=: image: one item in, potentially many out.
  • image :=>-: (i, q): potentially many in, one out.
  • image :=>=: image: many in, many out. Equivalent to the Transfer constructor.

And there are synonyms for Image as well:

  • One (i, q) = Map.singleton i q
  • Pair (i1, q1) (i2, q2) = Map.fromList [(i1, q1), (i2, q2)]
  • Many [(i1, q1), .., (ix, qx)] = Map.fromList [(i1, q1), .., (ix, qx)]

Other

Lens?

Very yes. In particular using makeLensesWith underscoreFields so accessors are punned. E.g. view quantity works on both ClusterSt and ClusterDy by way of the auto-generated Has_quantity typeclass.

Pretty printing?

Also very yes. Most of the pretty printing is implemented in Tech.Pretty. Uses prettyprinter and prettyprinter-ansi-terminal which are based on the famous Wadler/Leijen formulatio

YAML?

By way of Aeson. A set of *Spec types mirrors the *St types to represent the serialized version. Spec types are internally keyed as opposed to externally keyed and have symbolic references to recipes rather than direct inclusion.

E.g.:

data ClusterSpec = ClusterSpec
  { _clusterSpec_node :: Node
  , _clusterSpec_recipeKey :: RecipeKey
  , _clusterSpec_quantity :: Quantity
  }

data ClusterSt = ClusterSt
  { _clusterSt_recipe :: Recipe
  , _clusterSt_quantity :: Quantity
  }

See Tech.Store.

Automated tests?

Yep. Uses tasty, tasty-hunit, tasty-quickcheck, as one do.

Nix

For sure. Uses haskell-flake and thus the flake-parts modules, based on on the example haskell-template.

h/t to @srid et al.

This uses flakes, so you'll need a decently recent nix, and maybe to enable the flakes and nix-command experimental-features.

TODO

https://github.com/Dridus/techtonica/issues

Help?

If that's an offer, sure. https://github.com/Dridus/techtonica/issues

About

Techtonica calculator

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published