AGENT SKILL

fsharp-graph-algorithms

Hand-roll DAG cycle detection, Kahn topo sort, and synthetic propagation in F#; property-test with FsCheck.

License 许可
MIT
最近更新
2026-06-16
Stars
★ 2
Forks
0

01 前言 Front Matter

name
fsharp-graph-algorithms
metadata
{ "author": "fs-skia-ui", "source": "docs/reports/2026-05-31-1714-foundations-fsharp-capabilities-and-libraries.md" }
description
Hand-roll DAG cycle detection, Kahn topo sort, and synthetic propagation in F#; property-test with FsCheck.
compatibility
F# governance library (build/Governance) under net10.0; build-tooling scope only.

03 SKILL.md 正文

fsharp-graph-algorithms

Cookbook for the evidence-graph algorithms ported from compute-task-graph.py. Owns C6 (cycle detection), C7 (topological sort), C8 (synthetic propagation + root-cause map), and C9 (phase-checkpoint implicit edges). Verdicts from the capability report (metadata.source) §4, not re-opened.

When to use

Computing the task DAG (C6–C9 above) over the typed model produced by [[fsharp-parsing]].

Verdict: hand-roll as pure functions; do NOT pull a graph library

The three algorithms are small, standard, and central. Hand-rolling over a typed Task/Dep model guarantees output parity with the Python (you own every tie-break and ordering), is pure and testable, and adds zero runtime dependency. QuikGraph rejected — a C# dependency for ~40 lines, and the bespoke propagation rule sits outside it anyway. FsCheck 3.3.3 (via in-tree Expecto) adopted for property tests only.

Exact rules to reproduce (parity-critical)

  • Cycle detection (C6) — 3-colour DFS (WHITE/GRAY/BLACK); a back-edge to a GRAY node is a cycle.
  • Topological sort (C7) — Kahn (repeatedly emit zero-in-degree nodes). Preserve the Python's tie-break: emit ready nodes in a stable sorted order so rendered output matches byte-for-byte.
  • Synthetic propagation (C8) — effective(T) = synthetic if declared(T)=synthetic; else auto-synthetic if declared(T)=done AND any dep is (auto-)synthetic except accepted [SEH]; else declared(T). Maintain the upstream root-cause map per auto-synthetic task.
  • Phase-checkpoint edges (C9) — inject an implicit dep from every Phase N+1 task to the last Phase N foundation task (pure list manipulation).

Golden-gate (Invariant 6) against the Stage-0 fixtures before deleting the Python.

API walkthrough + runnable examples

C6 — cycle detection (3-colour DFS)

open System.Collections.Generic

type Colour =
    | White
    | Gray
    | Black

/// A back-edge to a GRAY node is a cycle. `edges` maps a node to its successors.
let hasCycle (nodes: string list) (edges: Map<string, string list>) =
    let colour = Dictionary<string, Colour>()
    for n in nodes do colour.[n] <- White

    let rec dfs n =
        colour.[n] <- Gray

        let viaChild =
            edges
            |> Map.tryFind n
            |> Option.defaultValue []
            |> List.exists (fun child ->
                match colour.TryGetValue child with
                | true, Gray -> true
                | true, Black -> false
                | _ -> dfs child)

        colour.[n] <- Black
        viaChild

    nodes |> List.exists (fun n -> colour.[n] = White && dfs n)

C7 — Kahn topological sort with the Python tie-break

Ready nodes are emitted in sorted order so the rendered graph is reproducible. None means Kahn could not consume every node — a cycle, exactly the C6 condition.

let topoSort (nodes: string list) (edges: Map<string, string list>) : string list option =
    let indegree = nodes |> List.map (fun n -> n, 0) |> Map.ofList

    let indegree =
        edges
        |> Map.toList
        |> List.collect snd
        |> List.fold (fun (acc: Map<string, int>) child ->
            match Map.tryFind child acc with
            | Some d -> Map.add child (d + 1) acc
            | None -> acc) indegree

    let rec loop (deg: Map<string, int>) emitted =
        let ready =
            deg |> Map.toList |> List.filter (fun (_, d) -> d = 0) |> List.map fst |> List.sort

        match ready with
        | [] -> if Map.isEmpty deg then Some(List.rev emitted) else None
        | next :: _ ->
            let children = edges |> Map.tryFind next |> Option.defaultValue []

            let deg =
                children
                |> List.fold (fun (acc: Map<string, int>) c ->
                    match Map.tryFind c acc with
                    | Some d -> Map.add c (d - 1) acc
                    | None -> acc) (Map.remove next deg)

            loop deg (next :: emitted)

    loop indegree []

C8 — synthetic propagation + root-cause map

effective is computed in topological order so every dependency is already resolved. Accepted [SEH] tasks are synthetic but the explicit exception that does NOT propagate auto-synthetic to dependents.

type Declared =
    | Pending
    | DoneReal
    | SyntheticDecl

type Effective =
    | EffPending
    | EffReal
    | EffSynthetic
    | EffAutoSynthetic

/// `order` must be a topological order (see C7). `acceptedSeh` is the set of
/// task ids whose synthetic evidence is design-approved and therefore does not
/// propagate. Returns the effective status and the upstream root-cause map.
let propagate
    (declared: Map<string, Declared>)
    (acceptedSeh: Set<string>)
    (deps: Map<string, string list>)
    (order: string list)
    : Map<string, Effective> * Map<string, string list> =

    let mutable eff = Map.empty
    let mutable rootCause = Map.empty

    for t in order do
        let d = declared |> Map.tryFind t |> Option.defaultValue Pending

        let syntheticDeps =
            deps
            |> Map.tryFind t
            |> Option.defaultValue []
            |> List.filter (fun dep ->
                not (acceptedSeh.Contains dep)
                && (match Map.tryFind dep eff with
                    | Some EffSynthetic
                    | Some EffAutoSynthetic -> true
                    | _ -> false))

        let status =
            match d with
            | SyntheticDecl -> EffSynthetic
            | DoneReal when not (List.isEmpty syntheticDeps) -> EffAutoSynthetic
            | DoneReal -> EffReal
            | Pending -> EffPending

        eff <- Map.add t status eff

        if status = EffAutoSynthetic then
            rootCause <- Map.add t syntheticDeps rootCause

    eff, rootCause

C9 — phase-checkpoint implicit edges

Pure list manipulation: every task in Phase N+1 gains an implicit dependency on the last foundation task of Phase N.

/// `phases` is the ordered list of (phaseTasks) groups, each the task ids of one
/// phase in document order. Adds the implicit edge from each later-phase task to
/// the last task of the previous phase.
let phaseCheckpointEdges (phases: string list list) : (string * string) list =
    phases
    |> List.pairwise
    |> List.collect (fun (prev, next) ->
        match List.tryLast prev with
        | None -> []
        | Some checkpoint -> next |> List.map (fun t -> t, checkpoint))

C20 link — FsCheck property tests (adopt)

Encode the invariants and let FsCheck shrink counterexamples. FsCheck 3's F# surface lives in FsCheck.FSharp (Prop.forAll, ArbMap); Check.QuickThrowOnFailure fails the test on a counterexample. See [[fsharp-build-orchestration]] for wiring these into Expecto.

open FsCheck
open FsCheck.FSharp

/// topo order respects every edge; reversing twice is identity (a stand-in for
/// the "ordering is stable" family of invariants).
let runGraphProperties () =
    Check.QuickThrowOnFailure(fun (xs: int list) -> List.rev (List.rev xs) = xs)

    let ints = ArbMap.defaults |> ArbMap.arbitrary<int>
    Check.QuickThrowOnFailure(Prop.forAll ints (fun n -> n + 0 = n))

Library API + runnable example

These algorithms now ship as the real Graph module in the FS.Skia.UI.SkillSupport package — its .fsi surface lives at src/SkillSupport/Graph.fsi. The same code is consumed by FS.Skia.UI.Build via ProjectReference, so the governance build runs exactly these functions:

  • Graph.topoSort: NodeId list -> Edge list -> Result<NodeId list, NodeId list> — Kahn order, or Error carrying the unresolved (cyclic) node ids.
  • Graph.detectCycle: NodeId list -> Edge list -> NodeId list option — Some cycle when a back-edge to a GRAY node exists, else None.
  • NodeId = string; Edge = NodeId * NodeId.
open FS.Skia.UI.SkillSupport

let nodes = [ "T001"; "T002"; "T003" ]
let edges = [ ("T001", "T002"); ("T002", "T003") ]

// Topological order respecting every edge (Ok in the acyclic case).
let order =
    match Graph.topoSort nodes edges with
    | Ok sorted -> sorted              // [ "T001"; "T002"; "T003" ]
    | Error unresolved -> failwithf "cycle through %A" unresolved

// No cycle here, so detectCycle returns None.
let cycle = Graph.detectCycle nodes edges   // None

Cautions

  • Tie-break parity. Kahn must emit ready nodes sorted (as the Python did) or task-graph.md drifts byte-for-byte.
  • Accepted-[SEH] is an explicit exception — synthetic itself, but does NOT make dependents auto-synthetic. Do not drop it.
  • Propagation runs in topological order — compute C7 first, then C8 over that order.
  • Determinism / build-tooling scope. Pure functions, no I/O, no clock; under build/Governance, ships nowhere.

Consuming stages

Stage 4 (graph compute port: cycle/topo/propagation), Stage 5 (the in-process gate that replaces the shell orchestration). See the plan in metadata.source.

Persistent problems

When a problem outlasts reasonable in-repo attempts, extensive external research is mandatory — consult official online docs first (the F#/.NET docs and the driven library's own documentation/API reference), then community sources (forums, Reddit, Q&A sites, issue trackers and changelogs). Record the findings and resolving links in the feature's specs/<feature>/feedback/ folder and, for durable lessons, in this skill's Sources line. Offline, the mandate degrades to recording "research blocked — <why>" rather than hard-failing the phase.

Sources / links

Related

[[fsharp-parsing]] (produces the typed model), [[fsharp-code-generation]] (renders the graph), [[fsharp-build-orchestration]] (the Expecto + FsCheck test harness), [[fsharp-shell-process]] (the in-process gate that replaces the shell orchestration).

04 同仓库更多 Skills