build: upgrade skillgym to 0.6.0

[thymikee]

Summary

Upgrade SkillGym to v0.6.0 and wire the local benchmark harness into the new suite controls.

• tag fixture smoke and skill-guidance cases for v0.6 tag filtering
• switch the planning suite to schedule: parallel so case/runner pairs can use the SkillGym available-machine parallelism cap
• use assert.commands.includes / assert.commands.notIncludes with structured commandMatcher(...) planned command checks for both agent-device ... and bare command lines
• use assert.output.includes for positive final-output assertions
• document v0.6 reporters, tags, and parallelism guidance in the SkillGym README and AGENTS.md

Touched files: 5. Scope stayed within SkillGym tooling/docs/tests.

Validation

• pnpm format
• pnpm check:tooling
• pnpm typecheck
• pnpm check:quick
• pnpm exec skillgym run ./test/skillgym/suites/agent-device-smoke-suite.ts --config ./test/skillgym/skillgym.config.ts --case open-and-snapshot --tag fixture-smoke --runner codex-mini --reporter json
• pnpm exec skillgym run ./test/skillgym/suites/agent-device-smoke-suite.ts --config ./test/skillgym/skillgym.config.ts --case text-replace-uses-fill --tag skill-guidance --runner codex-mini
• pnpm exec skillgym run ./test/skillgym/suites/agent-device-smoke-suite.ts --config ./test/skillgym/skillgym.config.ts --case open-and-snapshot --runner codex-mini --tag fixture-smoke (confirmed default Parallel 12)
• pnpm exec skillgym run ./test/skillgym/suites/agent-device-smoke-suite.ts --config ./test/skillgym/skillgym.config.ts --tag fixture-smoke --max-parallel 6 (48 runs passed in 3m43s, confirmed out-of-order parallel completion)
• pnpm exec skillgym run ./test/skillgym/suites/agent-device-smoke-suite.ts --config ./test/skillgym/skillgym.config.ts --tag fixture-smoke (uncapped default reported Parallel 12; 48 runs passed in 2m27s)
• pnpm test:skillgym ran full matrix under previous isolated-by-runner config in 31m48s with 66/70 cases passed and 136/140 runs passed; the 4 failed case/runner pairs all passed when rerun once
• git diff --check
• git diff --check -- AGENTS.md

Known gap: SkillGym does not currently have a built-in retry-failed option; filed callstackincubator/skillgym#17. Uncapped parallel runs currently emit Node MaxListenersExceededWarning at Parallel 12; noted on the upstream issue.

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://callstackincubator.github.io/agent-device/pr-preview/pr-473/
Built to branch gh-pages at 2026-04-29 18:28 UTC. Preview will be ready when the GitHub Pages deployment is complete.