AGENT SKILL

fs-skia-evidence-mode

Deterministic render-only evidence mode, visual-proof honesty, and benign/blocking host-warning classification.

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

01 前言 Front Matter

name
fs-skia-evidence-mode
description
Deterministic render-only evidence mode, visual-proof honesty, and benign/blocking host-warning classification.

03 SKILL.md 正文

FS Skia Evidence Mode Capability

Scope

Use this skill for tasks that change deterministic render-only evidence, visual evidence honesty, fallback classification, or readiness host-warning classification. This is the evidence-discipline half of the former fs-skia-layout-evidence skill; the generated-HUD/layout-readability half lives in fs-skia-layout-readability.

Such tasks must declare fs-skia-evidence-mode in tasks.deps.yml and mirror it on the matching tasks.md line. Resolve this skill before implementation starts and record the resolved path in the active feature's readiness evidence.

Public Contract

Public evidence contracts must start in .fsi files before implementation. Keep deterministic render metadata separate from readable layout proof. Use explicit proof levels:

  • DeterministicRenderOnly for hashes, scene metadata, or render readback that does not prove readability.
  • UnsupportedLayoutInspection when host, font, or public API facts cannot be produced. Unsupported facts must be actionable and must not be converted into readability proof.

Readable-layout proof itself (the ReadableLayout level) belongs to fs-skia-layout-readability; this skill governs what is not readable proof.

Runnable example

Classify render evidence honestly: a render hash is deterministic-render-only and must never be reported as readable layout.

open FS.Skia.UI.Scene
open FS.Skia.UI.Testing

let scene = Product.Program.view Product.Program.initialModel
let evidence = LayoutEvidence.capture scene

match LayoutEvidence.proofLevel evidence with
| DeterministicRenderOnly -> printfn "render hash only; not readable-layout proof"
| UnsupportedLayoutInspection -> failwith "unsupported layout inspection: facts not actionable"
| _ -> () // readable layout is proven and classified by fs-skia-layout-readability

Build Commands

Prefer repository targets over ad-hoc command sequences:

./fake.sh build -t GeneratedGuidanceCheck
./fake.sh build -t GeneratedProductCheck
./fake.sh build -t EvidenceGraph
./fake.sh build -t EvidenceAudit

Evidence

Readiness evidence for this capability belongs under the active feature's readiness/ directory and must state whether it proves deterministic rendering only or unsupported layout inspection.

Warning classification evidence must preserve real LaunchFailure, RenderingFailure, LayoutFailure, and PackageFailure diagnostics even when known benign environment warnings are present.

Visual evidence honesty requires screenshot proof, rasterized scene proof, layout readability proof, fallback classification, and unsupported proof to be separate. Accepted visual proof names a decodable image, image dimensions, non-trivial content, renderer mode, fallback classification, and unsupported reason. <!-- BEGIN GENERATED: gov/visual-proof-phrases --> Exact visual proof rejection phrases for scans: metadata-only reports do not satisfy visual proof; 1x1 fallback images do not satisfy visual proof; layout-only bounds claims do not satisfy visual proof. <!-- END GENERATED: gov/visual-proof-phrases -->

Asteroids feedback findings must be classified by owner (framework runtime, generated template workflow, documentation discoverability, consumer authoring), and host feedback must distinguish persistent-window blocking, display/session availability, auto-close smoke needs, benign warning, blocking warning, deferred warning, and name-collision guidance. <!-- BEGIN GENERATED: gov/owner-phrases --> Exact owner phrases for scans: framework runtime; generated template workflow; documentation discoverability; consumer authoring; persistent-window blocking; display/session availability; auto-close smoke; benign warning; blocking warning; deferred warning; name-collision guidance. <!-- END GENERATED: gov/owner-phrases -->

Evidence token parsing reads key=value lines, NOT tables (feature 085, FR-015)

The readiness validators (window-visibility, readiness-contract, generated-validation, the audit-status region) read machine-readable tokens only from key=value lines — e.g. window-visible=observed:true, exact-package-match=true, verdict=.... A markdown table that carries the same token names in cells (| window-visible | observed:true |) does not satisfy the validators — the token must appear as a bare key=value line. Put the authoritative tokens as key=value lines (tables are fine as human-readable supplements alongside them, never as the sole carrier). The exact required-token list per evidence class is single-sourced in docs/evidence-formats.md (generated from FS.Skia.UI.Build.Evidence.EvidenceFormatSchema; read it before authoring, and regenerate via ./fake.sh build -t RefreshSurfaceBaselines — never hand-edit).

--evidence-run deterministic-summary discipline (documented, not a shipped helper)

Generated games expose an --evidence-run command that drives the pure model through a fixed, per-frame held-input script and prints a deterministic summary. The reusable value is the discipline, not a shared function — follow it whenever you add or change an --evidence-run summary:

  • Drive only the pure model/update with a fixed seed and a deterministic per-frame held-input script (no wall-clock, no RNG outside the seeded generator, no host I/O).
  • Format every float with InvariantCulture and a fixed precision (e.g. F3) so the summary is byte-identical across machines and locales.
  • Emit a determinism=byte-identical marker plus the stable core fields status / command / seed / frame-count / score, then the game-specific fields.

Canonical examples: LunarLander1 EvidenceCommands.fs:603-672 emits physics state (final-position / final-velocity / final-rotation / final-fuel, F3); AsteroidsDemo3 EvidenceCommands.fs:670-685 emits entity counts (wave / lives / asteroid-count / bullet-count) and availability probes.

Why this is documented, not shipped (FR-009 deferral). A four-game comparison (LunarLander, Asteroids, Breakout, SpaceInvaders) shows the field set varies materially per game; the only stable core is status/command/seed/frame-count/score/determinism — too thin to justify a shipped record that every consumer then appends 5–10 game-specific fields to. The discipline recurs; the summary shape does not. Next-recurrence bar: ship a reusable summary type only if a stable cross-game field set (beyond that thin core) emerges; until then, write the summary inline and follow the discipline above.

Responds-vs-renders runtime proof (feature 090, FR-006/FR-007)

An interactive-UI story — one whose acceptance involves driving a live host with pointer/keyboard — is not done until it produces a responds-proof: a captured input→visible-change on the running host, where the rendered output before a real dispatched interaction differs from the output after it. This is a distinct evidence class. It is not satisfied by a render-only screenshot of a single frame, nor by an offscreen route probe (runInteractivePointerOnce) that exercises only the model layer. An app that renders but does not respond yields identical before/after frames and an Inert verdict — it cannot pass this proof, so an inert "green-everything-but-dead-window" build can no longer ship as done.

Capture the artifact pair plus the verdict under the feature's readiness/responds-proof/<case>/ (a before and an after frame, plus a responds-proof.txt carrying a bare verdict=Responsive token). No live Vulkan window is required — the before/after frames are honest render-only captures (ControlsElmish.captureRespondsProof reuses the production Control.renderTree path). This proof extends, not replaces, the production-render-path screenshot rule: renders-vs-responds is the gate that distinguishes "it draws" from "it reacts."

Package Boundary

Keep pure evidence classifiers in Scene or Testing contracts. Do not move viewer launch, filesystem, package restore, process, font host, or window-system effects into pure validation helpers. When host warning classification or evidence collection needs I/O, model the request and result explicitly and keep execution at the interpreter or build-target edge.

Related

  • [[fs-skia-layout-readability]]
  • [[fs-skia-scene]]

Sources / links

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.

04 同仓库更多 Skills