Update Skia

Update Skia

Use this skill as the source of truth for both a developer-run update and the automated sync. The workflow only supplies resolved inputs, a prepared host, and automated PR delivery.

Skia C++ -> SkiaSharp C API -> generated P/Invoke -> managed wrappers

Run from the mono/SkiaSharp repository root. Each phase lives in a separate reference so only the current work enters context.

Start state

Automation exports the SKIA_SYNC_* variables below. A local run resolves them in Phase 01. Never replace supplied automation values with assumptions about main, skiasharp, or branch names.

SKIA_SYNC_AUTOMATION=1 selects the automation handoff in Phase 11. Without it, use the repository PR templates and normal developer pushes.

Initialize the artifact directory once:

ARTIFACT_DIR="${SKIA_SYNC_ARTIFACT_DIR:-${TMPDIR:-/tmp}/skia-sync-agent}"
mkdir -p "$ARTIFACT_DIR"

Definition of success

An update is complete only when:

• The authoritative old-upstream-to-target range was analyzed before merging.
• The mono/skia result is a genuine two-parent merge with every fork patch and dependency decision accounted for.
• The updated native source builds; downloaded old native artifacts were never substituted.
• Bindings were regenerated and every required managed wrapper was reviewed.
• The final unfiltered tests/SkiaSharp.Tests.Console.slnx run passes every host, including every GPU backend required by GpuPolicy on the validation host.
• The parent points to the exact tested mono/skia commit.
• Both PR descriptions identify untested platforms and are ready for human review.

Working rules

• Create feature branches in both repositories before changes; never commit to protected branches.
• Use a genuine two-parent merge in mono/skia; never use a tree-override merge.
• Preserve every fork patch unless upstream contains an equivalent or improved form.
• Mark a removed patch upstreamed only after enumerating every independent behavior in its old delta and locating each one in the target; one upstreamed hunk does not cover another lost hunk.
• Classify dependency revisions against the fork base, not only the prior upstream milestone.
• Treat final DEPS as ground truth: every enabled revision that differs from the fork base must have a matching final decision, exact-SHA evidence, and reconciled Component Governance metadata. A dependency recorded as preserved must still equal the fork-base revision.
• Let update_versions.py identify dependency changes from exact base/final DEPS. For every tracked changed dependency, and every legacy tracked registration missing verification evidence, derive the semantic version from checked-out source and complete the skia_dependency fields before the helper can pass. Never update a manifest version when its DEPS identity did not change.
• Components not sourced from Skia DEPS, including ANGLE and its dependencies, are updated in separate dependency PRs rather than bundled into a Skia upstream sync.
• Never use externals-download after a submodule/native/C API change.
• Never manually edit generated declarations or interop code. Direct edits to source-controlled /// comment trivia are allowed when regeneration preserves them.
• Keep public managed ABI additive.
• A focused project test is diagnostic only; it never satisfies the final gate.
• GPU bring-up failures are test failures. The validation environment must provide every backend required by GpuPolicy for its platform.
• Diagnose failures from repository evidence. Do not add one-off compiler/GN flags, skip tests, weaken assertions, or encode milestone-specific answers to make one run green.
• Execute the update phases in this agent. Do not delegate the full update, or any mutating, build, test, or delivery phase, to a general-purpose/background agent. Only Phase 03's explicitly read-only discrepancy review may be delegated.
• Run long builds and full test suites in the foreground as one shell invocation. Do not use nohup, &, or agent-turn progress polling. If the shell tool keeps a command running, block on that existing session and inspect its output once after it exits.
• Keep complete logs and broad diffs in $ARTIFACT_DIR; inspect bounded excerpts instead of streaming large generated files, manifests, or successful build/test logs into model context.
• Do not create PRs, write automation handoff files, or report completion while any gate fails.
• In automation, no-work is handled before the agent starts. A started agent that cannot complete must fail rather than return success-shaped output.