Skia Analyst

You analyze Skia features for SkiaSharp from two angles simultaneously:

1. What shipped — upstream engine benefits, PR links, migration guides
2. What's missing — gap analysis with impact/priority/effort scoring, hidden API scan, action items

Every run produces both. Output is structured JSON and rendered GitHub-flavored Markdown.

This skill always runs in a SkiaSharp checkout. It uses:

• externals/skia/ submodule for the C API (our fork at mono/skia)
• binding/SkiaSharp/SkiaApi.generated.cs for the C API reflected as P/Invoke externs
• binding/SkiaSharp/*.cs for the C# wrappers
• Upstream google/skia headers fetched via GitHub for hidden API comparison

Key References

• references/schema-cheatsheet.md — Human-readable schema
• references/skia-analyst-schema.json — JSON Schema (Draft 2020-12)
• references/analysis-instructions.md — Classification criteria

Input Flexibility

The user may specify anything — infer the scan mode:

If unclear, ask. But try to infer first.

Workflow

Phase 1: Setup (determine scan range, prepare sources)
Phase 2: Launch dual-model agents (Opus + GPT in parallel)
Phase 3: Synthesize — merge findings, dedupe, both lenses
Phase 4: Generate outputs (JSON → validate → Markdown)
Phase 5: Present results

Phase 1: Setup

1a. Determine scan range

Based on user input, establish:

• scanMode: full, windowed, or diff
• For windowed: milestoneFrom and milestoneTo
• For diff: refFrom and refTo (resolve SHAs, dates, commit count)

1b. Determine current milestone

cat externals/skia/include/core/SkMilestone.h

Or check commit messages for the latest Skia bump PR.

1c. Fetch Skia release notes

Fetch RELEASE_NOTES.md from google/skia using the GitHub MCP tool or gh api:

gh api repos/google/skia/contents/RELEASE_NOTES.md -H "Accept: application/vnd.github.raw" > skia-release-notes.md

Save to a file in the working directory for agents to reference.

1d. Prepare upstream headers for hidden API scan

Agents will fetch upstream C++ headers directly from GitHub during their scan:

github-mcp-server-get_file_contents owner=google repo=skia path=include/core/SkImage.h

1e. Locate binding sources

The C API is reflected in binding/SkiaSharp/SkiaApi.generated.cs as P/Invoke extern methods. The C# wrappers are in binding/SkiaSharp/*.cs. Both are in the worktree.

For the C API headers (our fork), check externals/skia/include/c/ and externals/skia/src/c/. If the submodule isn't checked out, agents can grep SkiaApi.generated.cs for sk_* and gr_* extern function names — this reflects the full C API surface.

Phase 2: Launch Dual-Model Agents

Launch two background agents simultaneously on different models:

task agent_type=general-purpose mode=background model=claude-opus-4.7 name=analyst-opus:
task agent_type=general-purpose mode=background model=gpt-5.4 name=analyst-gpt:

Each agent does the complete job independently:

1. Release notes scan — read Skia RELEASE_NOTES.md, extract features
2. Hidden API scan — fetch upstream C++ headers from google/skia, compare against binding/SkiaSharp/SkiaApi.generated.cs for P/Invoke externs and binding/SkiaSharp/*.cs for wrappers
3. Binding verification — grep the actual code to set bindingStatus
4. Git diff (if diff mode) — analyze API/build/dep changes between refs

For EVERY finding, classify with BOTH lenses:

• Changelog: changeType + importance
• Gap: bindingStatus + impact + priority + effort

Phase 3: Synthesize

When both agents complete, merge their findings:

1. Deduplicate by name/skiaApi — keep richer data from each
2. Resolve conflicts — more cautious bindingStatus wins, higher impact wins
3. Union hidden APIs — combine both agents' header scan discoveries

Phase 4: Generate Outputs

4a. Generate JSON

Save to skia-analyst-report.json in the working directory.

4b. Validate

python3 .agents/skills/skia-analyst/scripts/validate-skia-analyst.py skia-analyst-report.json

4c. Render Markdown

python3 .agents/skills/skia-analyst/scripts/render-skia-analyst.py skia-analyst-report.json skia-analyst-report.md

This produces a GitHub-flavored Markdown file with collapsible details, suitable for pasting into a GitHub issue or sharing as a gist.

Phase 5: Present Results

Show highlights inline:

• Top gaps — transformative and significant items
• Quick wins — partial binding status (C API exists, just needs C# wrapper)
• Breaking changes — if any findings have importance: breaking

Then offer next steps:

1. "Want me to investigate any finding in more detail?"
2. "Should I use the api-add-review skill to start binding a feature?"
3. "Want to upload the markdown to a gist for sharing?"
4. "Should I create a GitHub issue with the gap analysis?"