Problem

There is no structured way for AI coding agents (Claude Code, GitHub Copilot, Codex) to know which expertise or workflow to apply for different Proton development tasks (build, SQL, C++ review, CI, etc.).

Solution

Add a single canonical skill library under .github/skills/ and wire each agent platform to it via lightweight entry-points — no duplication.

.github/skills/          ← canonical skill library (10 skills)
     ↑                        ↑                    ↑
.claude/skills (symlink)   .github/copilot-    .codex/skills (symlink)
.claude/CLAUDE.md          instructions.md     .codex/config.toml
(Claude Code)              (Copilot)           (Codex)

Skill library — .github/skills/

Agent entry-points

• Claude Code: .claude/CLAUDE.md (project instructions) + .claude/skills → ../.github/skills
• GitHub Copilot: .github/copilot-instructions.md
• Codex: .codex/config.toml + .codex/skills → ../.github/skills
• AGENTS.md at repo root for other agent frameworks

Additional fixes (found during local macOS development)

• tests/ported-clickhouse-test.py: use ThreadPoolExecutor on macOS to avoid fork() EAGAIN with coverage builds; fix HTTP connection leak; guard setpgid/killpg
• src/CPython/tests/CPythonTest.h: auto-discover Python 3.10, skip suite gracefully
• src/Bootstrap/ServerDescriptor.cpp: unify port config to node.*.port pattern
• base/base/coverage.cpp: suppress -Wreserved-identifier
• local_coverage.sh, start-local-proton.sh: local developer helper scripts

Acceptance Criteria

• Claude Code, Copilot, and Codex agents route tasks to the correct skill
• tests/ported-clickhouse-test.py runs parallel tests on macOS without fork errors
• CPython unit tests skip gracefully when Python 3.10 is absent