addyosmani/agent-skills 0.6.10

Agent Skills 0.6.10

on GitHub

Agent Skills 0.6.10 is a reliability and platform release. It changes one default behavior worth reading first, fixes the Windows SessionStart failures, closes a data-loss bug in the simplify-ignore hook, tightens the eval and validator tooling, and lands a wave of cross-host portability docs.

Behavior change (read this first)

• The plugin no longer auto-injects the using-agent-skills meta-skill on SessionStart (#574, @santhiprakash, fixes #569). Claude Code and Codex already discover and activate skills from their descriptions, so the old hook was stacking a second router on top of the native one, the pattern our own getting-started.md warns against, and it cost roughly 3.5k tokens per session for no gain on those hosts. The standalone hooks/session-start.sh stays in the repo as an opt-in helper for hosts without native skill routing. If you were relying on the meta-skill being force-loaded every session, wire that script into a SessionStart hook yourself.

Reliability and platform

• Windows SessionStart failures fixed (#488, #475). They were caused by the hook above being run through PowerShell; removing the plugin-registered hook removes the failure entirely.
• simplify-ignore no longer loses work on Stop (#578, @mavericksea-ai). It now expands the current file contents instead of blindly restoring the backup, with a .recovered fallback for wholesale rewrites.
• SessionStart payload test fixed and wired into CI (#510, @santhiprakash). It now asserts the correct hookSpecificOutput envelope, replacing an assertion that was stale on main, and runs on every PR.

Evals and validators

• skill-lint handles every CommonMark fence form (#568, @ayobamiseun, fixes #437). A real line-by-line scanner replaces the single regex, so tilde fences, indented fences, and longer closers are all recognized; its test is now wired into CI.
• The eval grader is stricter (#576, @mavericksea-ai). Each grader result is bound to a declared expectation by id, the canonical expectation text overrides the grader's paraphrase, and pass_rate is recomputed rather than trusted. Paired with rejecting null grader expectations without crashing (#563, @RogueTex).
• Reference-link check exempts fenced examples (#517, @abhisheksharma2411) so an example link inside a code block is no longer treated as a link to resolve, while a real broken link outside a fence still fails.
• Reference-link assertion is portable on Windows (#573, @dajiaohuang).
• Routing description gaps closed and the rank-1 floor raised to 95 (#434, @ayobamiseun).

Docs and portability

• Don't stack a second router on native skill hosts (#557, @kaluli123123), the guidance the behavior change above now matches.
• Authoring rule against model-specific workarounds (#545, @ayobamiseun, from #534): write the procedure, not the workaround around one model's failure.
• Cross-host portability and clarity: map the shared core to native adapters (#560), define restartable task boundaries (#559), clarify external artifact ownership (#554), setup for existing projects (#553), the stable browser skill name (#552), and link runnable workflow tutorials (#556), all @kaluli123123. README skill count corrected to 25 (#570, @abhisheksharma2411).
• Context Budget Management added to context-engineering (#447, @HMAKT99), plus SLOs and error budgets and runbook guidance across observability and shipping (#415, #422, @HMAKT99).
• Incomplete-plan guard mirrored across tools (#531, @ayobamiseun, from #518).

Thanks

Thanks to everyone who contributed and reviewed this release: @santhiprakash, @mavericksea-ai, @ayobamiseun, @kaluli123123, @abhisheksharma2411, @HMAKT99, @dajiaohuang, @RogueTex, @nucliweb, and @federicobartoli.

Full Changelog: 0.6.9...0.6.10