Agent Skills 0.6.162026-11-14
Agent Skills 0.6.16 lets specs evolve past V1, teaches the docs skill to spot drift, stops the frontend skill from installing toolchains just to verify, and fixes two quality-gate examples that could never fail.
Skills
- Spec evolution (#383, @ayobamiseun). Specs now carry a version and an Out of Scope (Deferred) list, and a change to a shipped feature gets its own small delta spec (SPEC-<change>.md) instead of a rewrite. /spec writes one when SPEC.md already exists, and /build auto picks the delta that isn't Status: implemented and refuses to run a plan generated from a different spec.
- Documentation drift (#514, @vyasdx). When a doc may be stale, documentation-and-adrs compares the last commit to the code with the last commit to the doc, and says "can't tell" instead of "current" when the history is unordered. Change logs are appended to, never rewritten.
- Verify with what the project has (#732, @TwelveNights). frontend-ui-engineering no longer installs a test runner or type checker just to check its work, and follows the WAI-ARIA APG patterns for menus, dialogs, tabs, listboxes and comboboxes. Measured +0.10 across 11 models, with package installs down from 24% to 4% of runs.
- Review findings checked against the code (#736, @CybotTM). code-review-and-quality checks each finding against the callers, guards and tests that reach the change, labels defects already in the base separately, and treats zero findings as a valid result.
- Gates that can fail (#786, @rudycelekli). depcruise --validate src printed usage and exited 0; the constraints skill now runs depcruise src.
- Session secret (#787, @rudycelekli). The hardening example fails at startup when SESSION_SECRET is missing instead of returning 500 on every request.
- OWASP Top 10:2025 (#785, @usaid7015). The checklist and security skill use the 2025 numbering, cited with the year.
- Skills fire on pasted code (#762, #763, #757, @ayobamiseun, part of #620). test-driven-development, code-review-and-quality and debugging-and-error-recovery now fire when the code is pasted inline and the answer looks obvious, each with a stays-quiet case guarding its neighbours.
Evals and tooling
- Behavioral runs stage the skill directory and shared references/ outside the fixture baseline, so a skill can read its checklists (#744, @CybotTM, fixes #584).
- The reference-link validator keeps markup inside code spans in heading slugs (#782, @kevin9327).
- The hook suites also run on macOS's stock bash 3.2, and the staged-secret-scan test runs cross-platform (#778, @ayobamiseun).
Thanks
Thanks to everyone who contributed and reviewed this release: @ayobamiseun, @vyasdx, @TwelveNights, @CybotTM, @rudycelekli, @usaid7015, @kevin9327, @abhisheksharma2411 and @federicobartoli.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.15...0.6.16
Agent Skills 0.6.152026-10-31
Agent Skills 0.6.15 stops the security skill's staged-secret check from printing the secret it finds, fixes two simplify-ignore bugs that could corrupt protected code, and makes the constraints skill's security and accessibility gates actually fail CI.
Security
- The staged-secret check no longer echoes credentials (#718, @santhiprakash, reported by @xmxxc in #717). The old git diff --cached | grep printed every matching line, so a staged key landed in the terminal and in any agent's transcript. It also read as clean when git failed. The check is now git diff --cached -i -G'…' --name-only: it prints only file names, and a git failure exits non-zero by itself. Fixed in all three places it was documented.
Hooks
- simplify-ignore restores re-indented placeholders cleanly (#766, @Quinnod345, #765). If /code-simplify un-indented a placeholder or edited its reason, Stop wrote back a broken comment around the block, a syntax error in any / / language. Blocks marked with -- simplify-ignore-start (SQL, Lua, Haskell) no longer warn that they were deleted, and a missing lock directory no longer stops Stop before every file is restored.
Skills
- Quality gates that fail when they should (#694, @mavericksea-ai, #693). Semgrep and axe exit 0 even with blocking findings, so the constraints skill's examples could pass CI while breaking the written bar. The skill now ships a small report gate that blocks high and critical findings, and it says to run each enforced command against a deliberate violation before calling it enforced.
- Contrast thresholds corrected (#733, @CybotTM): large text is 18pt (24px) or 14pt bold, not 18px.
- One-line gotchas from @CybotTM: compact well before the host's context limit, which varies by host (#739); keep a contradicting caller or guard in the doubt review unit (#754); and an injected exception isn't a crash, and a mocked transaction isn't a rollback (#750).
- git-workflow-and-versioning fires on a pasted diff (#758), and security-and-hardening names session cookies in its description (#759), both @ayobamiseun as part of #620.
- The shipping error boundary types its children prop (#683, @rudycelekli).
Evals
- --eval <id> runs a single behavioral case (#742, @CybotTM).
- The runner names any positive prompt that passes top_k without ranking first (#760), and the plugin-eval graders no longer expect command names that #714 removed, with the re-run recorded (#756), both @ayobamiseun.
Thanks
Thanks to everyone who contributed and reviewed this release: @santhiprakash, @xmxxc, @Quinnod345, @mavericksea-ai, @CybotTM, @ayobamiseun, @rudycelekli and @federicobartoli.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.14...0.6.15
Agent Skills 0.6.142026-10-29
Agent Skills 0.6.14 is mostly about code samples. More than twenty snippets across the skills either failed as written or taught the wrong thing, and they're fixed. It also closes two supply-chain and injection risks in the samples, stops Claude Code from starting /ship or /build on its own, and brings the root manifest in line with Agent Plugins 1.0.0.
Security fixes in the samples
- The Lighthouse CI example ran the unscoped lhci package (#721), which isn't Lighthouse CI. It now invokes @lhci/cli.
- The rollback workflow passed its input straight into run: (#691), a script injection in any repo that copied it. It goes through env now.
- The constraints skill scans the working tree for secrets, not only commits (#729), and the hardening reference returns 404 before an ownership check reads a missing task (#726).
Samples that now work as written
- Accessibility: labels are tied to their controls (#719), the loading region has a role so its label is exposed (#725), and the axe example runs the CLI that actually ships a binary (#720).
- Testing and debugging: the Playwright example finds list items by their text (#688), the Jest example isolates a file in a way Jest 30 still accepts (#712), and the coverage setup installs Vitest's coverage provider, without which CI fails (#713).
- Frontend: the task list passes its toggle and delete handlers down (#706), and the empty state declares its action prop (#692).
- API, CI and constraints: the idempotency guard returns the skill's own error shape (#710), setup-node is described as caching the package-manager cache, not node_modules (#711), a ratchet reads its direction from its own column (#708), and --fix reaches the linter through npm run (#690).
- Everything else: the deprecation adapter type-checks (#698), the attribution-build onINP import is aliased so it doesn't redeclare (#699), Lighthouse points at http://localhost (#727), the ADR example comment describes its own code (#707), the source-driven example quotes a sentence the React 19 post actually contains (#728), and the idea-refine helper runs from the installed skill (#724).
Skills and commands
- The lifecycle commands are user-invoked only (#714, @monraza). Claude Code now lists commands next to skills, so without this the model could start /ship, /build and the rest by itself. The skills underneath still route as before.
- code-simplification fires on pasted code (#696, @ayobamiseun), with the same one-clause change that worked for review and test-driven-development in #620, plus a case that keeps it quiet on "make it fast".
Hosts
- The root plugin.json is an Agent Plugins 1.0.0 manifest (#463). It validates against the published schema, and Antigravity and Copilot CLI load the same 25 skills as before. The Copilot guide explains what the new $schema line changes.
- The Antigravity guide tells you to write your own project AGENTS.md instead of copying this repo's (#747), Gemini MCP setup points at settings.json (#705), the Windsurf rules file gets real newlines between skills (#704), and the Cursor guide links the rules and skills pages (#723).
Evals and validators
- Routing understands clipped words (#668, @Yi-111-a). "docs", "auth", "config", "repo" and "deps" now match their long forms, and "specification" no longer stems into "specific". Rank-1 routing is 94 of 95.
- Behavioral runs that stop early are rejected, not graded (#743, @CybotTM). A run that hit max turns with a tool call left hanging used to score 4 of 4.
- The artifact-path check stops reading test-plan.md as plan.md (#702), and SDD-CACHE uses "reuse" throughout (#735, @DeadY2J).
Thanks
Thanks to everyone who contributed and reviewed this release: @rudycelekli (twenty-four of these fixes), @ayobamiseun, @Yi-111-a, @CybotTM, @monraza, @DeadY2J, @federicobartoli and @nucliweb.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.13...0.6.14
Agent Skills 0.6.132026-10-13
Agent Skills 0.6.13 fixes code samples in three skills that failed as written, two simplify-ignore bugs that could lose protected code or bring back a reverted change, and gets test-driven-development to fire when the code is already in the prompt. It also sets a written bar for which hosts the README lists.
Skill fixes
- Three samples that did not work as written (@rudycelekli). The error boundary in shipping-and-launch threw on every render and could never show its fallback (#637). The error state in frontend-ui-engineering called a refetch that was never defined (#638). The zod title check in the hardening reference accepted a title made only of spaces (#639).
- test-driven-development fires on pasted code (#678, @federicobartoli). "Add unit tests for this function" and a pasted bug with no mention of tests went from 0 of 6 to 6 of 6 runs with one added clause, with no false fires across 24 adjacent prompts. Measured with the model pinned, as proposed in #620.
- documentation-and-adrs gains a prose review step (#661, @manavmishra): mark the details that must survive an edit, and never replace an unsupported claim with an invented example or a stronger guarantee.
- /review and /test are documented as what they are (#682, reported by @sundeer in #681). They run their skill in the main session; they do not spawn the code-reviewer or test-engineer personas, whatever the docs said. The review commands now ask for the same severity labels the skill uses (Critical, Required, Optional, Nit).
- idea-refine follows the standard skill anatomy and says where interview-me and spec-driven-development take over (#624, @zhujunxiongfeiyi, with @nucliweb). JavaScript blocks in performance-optimization are labelled as JavaScript (#677, @ayobamiseun).
Hooks
- simplify-ignore no longer loses a block whose placeholder was removed (#649), and keeps a change you reverted with git instead of restoring it at Stop, re-hiding the file afterwards (#651). Both @Sunil56224972, both with tests that fail on 0.6.12.
Hosts
- A written bar for host guides (#641). The README install section lists only hosts a maintainer has run. Every other host gets one line in docs/other-hosts.md. The Oh My Pi and Dojo Workspace guides that shipped in 0.6.12 were withdrawn under it, and both are now listed there alongside Pi and fx (#659 @H3xKatana, #655 @ayobamiseun, #642 @un33k).
- The Cursor guide says .cursor/ lives in your project, not this repo (#663), and the Copilot guide says up front that the lifecycle slash commands never appear in Copilot (#662), both @ayobamiseun.
- In Claude Code, use /agent-skills:review where a built-in command shares the name (#682).
Evals and validators
- Grading rejects blank evidence with its own message (#666, @Aliciawque), duplicate eval IDs can no longer overwrite a result (#632), a failed fixture setup cleans up after itself (#633), fixture commits work with commit signing on (#636), and a quoted skill name is read the same way by every script (#675), all four @rudycelekli.
- Heading anchors match GitHub's numbering for repeated headings (#634) and command descriptions are read with the shared frontmatter parser (#676), both @rudycelekli. The linter accepts frontmatter comments and no longer lets a key nested under metadata overwrite a top-level one (#658, @ALANDVO). Personas are linted for name, description and their Composition section (#630, @ayobamiseun).
- The developer docs list the same checks CI runs (#644, @ayobamiseun).
Thanks
Thanks to everyone who contributed and reviewed this release: @rudycelekli, @federicobartoli, @nucliweb, @Sunil56224972, @ayobamiseun, @manavmishra, @zhujunxiongfeiyi, @Aliciawque, @ALANDVO, @H3xKatana, @un33k, and @sundeer.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.12...0.6.13
Agent Skills 0.6.122026-10-03
Agent Skills 0.6.12 is mostly about skills saying the right thing. It fixes a spec workflow that ran past its own approval gate, removes a command that does not exist and a code sample that does not parse, and slims performance-optimization the same way 0.6.11 slimmed security-and-hardening.
Skill fixes
- spec-driven-development stops after the spec (#629, @nucliweb, fixes #622). On hosts that load only the skill, some models ran Specify, Plan, Tasks and Implement in one turn, because the explicit "confirm before proceeding" step lived only in the /spec command. The skill now ends the turn after saving the spec and waits for approval, the same fix #497 made for interview-me.
- No more invented rollback command (#628, @jsynowiec). The shipping-and-launch checklist told you to run npx prisma migrate rollback, which Prisma does not have. It now asks for your project's verified rollback command or runbook, whatever migration tool you use.
- The idempotency example parses (#627, @jsynowiec). A bare throw; in api-and-interface-design is a syntax error in JavaScript; it now rethrows the caught error.
- performance-optimization is slimmer (#580, @Syamsuddin, with fixes from @nucliweb's review). Step 3's code patterns moved into a skill-local references/optimization-patterns.md, taking SKILL.md from 496 to 267 lines. The condensed summary keeps the guidance that matters without opening the reference: when an index will not help and what to do instead (partial, trigram or expression indexes), that a bad rows= estimate means ANALYZE rather than a new index, and stating a cache's staleness window.
- code-review-and-quality answers its own regression question (#618, @CybotTM): invert one condition the change adds, run the suite, and report a mutation that stays green as a missing test.
- The ADR template's Status line now starts at Proposed, matching the lifecycle below it (#615, @mavericksea-ai).
Host guides (withdrawn on main)
This release shipped setup guides for Oh My Pi (#617) and Dojo Workspace (#616). Both were withdrawn on main in #641: they were merged ahead of community host guides that had waited longer, and no maintainer had run either host. CONTRIBUTING now has a written bar for host guides. Oh My Pi is listed as one line in docs/other-hosts.md.
Tooling
- Anchors in reference links are checked (#626, @nucliweb, fixes #625). Renaming a heading in a reference file used to break every #anchor link to it while CI stayed green. The validator now resolves anchors with GitHub's slug rules.
- The floor-guard reference no longer mistakes an added ++ line or a removed -- line for a file header, and it runs from the top of the work tree so untracked files outside a subfolder are not missed (#614, @mavericksea-ai).
- Frontmatter values that start with a YAML indicator are rejected (#603, @abhisheksharma2411), and so are top-level frontmatter keys outside the spec, with the per-agent configuration guide now linked from CONTRIBUTING, the README and skill-anatomy (#608, @ayobamiseun).
- Every plugin eval grader counts the slash-command path as well as the skill (#613, @federicobartoli).
Thanks
Thanks to everyone who contributed and reviewed this release: @nucliweb, @jsynowiec, @andrebrait, @un33k, @mavericksea-ai, @federicobartoli, @CybotTM, @ayobamiseun, @abhisheksharma2411, and @Syamsuddin.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.11...0.6.12
Agent Skills 0.6.112026-09-26
Agent Skills 0.6.11 is a correctness and tooling release. A few fixes change what you actually get from the pack, so they come first; the rest tightens the evals, validators, and CI that keep the catalog honest.
Fixes worth knowing about
- The floor-guard reference now does what it promised (#600, @mavericksea-ai, fixes #599). The reference guard in constraint-driven-development silently passed on untracked suppressions, deleted test files, and loosened maximum budgets, 7 of the 14 cases it claimed to catch. It now catches all of them, respects .gitignore, and has a test suite in CI. If you adapted it into your own project, pull the new version.
- security-and-hardening is slimmer and keeps its rules (#579, @Syamsuddin; #589, @nucliweb). Code patterns moved into a skill-local references/hardening-patterns.md, cutting SKILL.md from 524 to 216 lines. The condensed prose had dropped a few rules, and #589 restores them: sameSite as the CSRF defense, npm audit signatures / pnpm audit signatures, retention with a working deletion path, and validating every tool argument an LLM passes.
- simplify-ignore no longer corrupts restored blocks on bash 5.2 (#501, @ayobamiseun). With patsub_replacement on by default, & and backslashes in protected code were mangled on restore. Fixed, with sdd-cache and simplify-ignore regression suites now running in CI.
- code-review-and-quality fires on pasted diffs and pull requests (#612, @federicobartoli), a finding from the plugin eval pilot below.
Skills
- interview-me stops its turn once intent is confirmed instead of starting downstream work (#497, @Varun-Chinthoju).
- frontend-ui-engineering adds reference-led finish checks: references as evidence, never templates (#496, @samuelbushi).
Evals
- claude plugin eval pilot (#596, #611, @federicobartoli): loads the whole plugin and measures whether Claude Code's own router picks a skill and what it adds to the reply. On demand only, scoped to evals/plugin/ through the manifest.
- Graders are stricter: stale grading from a crashed run can no longer pass as current, and runs record their identity (#587, @mavericksea-ai); the ADR eval grades status and date instead of an expectation the skill never states (#598, @mavericksea-ai); negative trigger cases now name their owner, so they can't pass vacuously (#493, @Zlobin-K).
Validators and CI
- Validators run on macOS and Windows as well as Linux (#590, @ayobamiseun).
- Links inside a skill's own references/ files are now checked (#588, @nucliweb, fixes #585), on shared CommonMark fence parsing (#586, @ayobamiseun).
- Frontmatter must be valid YAML for both skills and commands (#594, #595), negated trigger clauses are all stripped rather than only the first (#605), and the skill-anatomy layout rules plus a 500-line SKILL.md budget are enforced (#571), all @abhisheksharma2411.
- The artifact-path guard covers every command surface, including the Gemini and root TOML copies (#607, @ayobamiseun).
- Manifest versions are validated against the root plugin.json rather than the latest git tag (#498, @aoright).
Thanks
Thanks to everyone who contributed and reviewed this release: @nucliweb, @federicobartoli, @mavericksea-ai, @ayobamiseun, @abhisheksharma2411, @Syamsuddin, @aoright, @Zlobin-K, @samuelbushi, and @Varun-Chinthoju.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.10...0.6.11
Agent Skills 0.6.102026-09-18
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: https://github.com/addyosmani/agent-skills/compare/0.6.9...0.6.10
0.6.92026-09-05
Agent Skills 0.6.9 is a docs, hardening, and contributor-process release. No new skills this time; it sharpens the security and observability guidance in existing skills, adds setup and workflow docs, and lands a governance mechanism for the catalog's own development.
Highlights
- Advanced per-agent configuration guide (#316, @nucliweb). The canonical reference for keeping SKILL.md frontmatter portable and pushing vendor-specific runtime controls (model, tools, turn limits, thinking level) into metadata or a per-agent adapter, so the same skill file works across Claude Code, Cursor, Gemini, Antigravity, and other spec-conformant clients.
- Rejected skill-change ledger (#544, from #535). An append-only record of skill and description changes that were proposed and rejected, together with the eval score that rejected them, so the same idea does not get re-litigated in a new PR. Checking it is now a pre-flight step in CONTRIBUTING, and the ledger row lands on the default branch so closing the rejected PR cannot erase the record.
Hardening (existing skills)
- security-and-hardening: destructive path operations (delete, move, overwrite) now require symlink resolution, an allowlisted root, a minimum depth, and an ownership check read before the call (#547, @notbucki). And rate limiting must count in a shared store across instances, since a per-instance counter silently stops working the moment you run more than one instance (#536, @CahidArda).
- observability-and-instrumentation: when several entry points write to one log stream, stamp the entry point next to the correlation ID at the point the run starts. A correlation ID says which run; it does not say which code path started it (#546, @notbucki).
Docs
- Copilot: the standalone copilot CLI (plugin install) and Copilot inside VS Code are separate setups with different invocation models, now documented as such (#550, @DEOWL-kan, addressing #542).
- Lifecycle: the spec, plan, and task artifacts are the handoff between sessions. Run the whole lifecycle in one session for a small task; use a fresh session per phase for anything non-trivial (#549, @DEOWL-kan, from #513).
- Antigravity: the command-wrapper discoverability limitation is documented (#532, @ayobamiseun, from #445).
- Routing: database-migration and dependency-audit vocabulary added to the relevant skill descriptions (#537, @notbucki).
Thanks
Thanks to everyone who contributed this release: @nucliweb, @DEOWL-kan, @notbucki, @CahidArda, @ayobamiseun, @mvanhorn, and @federicobartoli.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.8...0.6.9
0.6.82026-08-28
Agent Skills 0.6.8 adds a new skill for setting and holding a project's quality bar, deepens the performance guidance with real backend and database substance, and clears a round of fixes.
First, a thank you. This repository just passed 90,000 stars. That is entirely down to the people who use these skills, file sharp issues, send careful PRs, and field-test them on real projects. Thank you, genuinely. This release is a good snapshot of that: a new skill, a large contribution from a collaborator, and a handful of fixes that came straight from your reports.
Highlights
- New skill: constraint-driven-development (#472). Establishes a project's quality bar as a written contract (CONSTRAINTS.md), interviews you on which dimensions matter, supplies sane defaults when you have no number in mind, maps every dimension to a de facto tool with install and run commands, and watches the diff for a weakened bar: new suppressions, skipped tests, stripped assertions, unimplemented stubs, thresholds edited down. It ships a reference floor-guard so agents enforce the floor the same way across runs and stacks instead of reinventing one each time. Thanks to @federicobartoli for a field-test-driven review that shaped it.
- performance-optimization, deepened (#521, @nucliweb). Real backend and database depth: reading the query plan before reaching for an index, when an index will not help, connection-pool exhaustion, and caching strategies (layer choice, key design keyed on the viewer, one invalidation strategy, stampede protection). Decision material in the skill, patterns and checklist in references/performance-checklist.md.
Skills and content
- performance-optimization now documents its Guard step (it was declared in the workflow but never written up), and a new validator rule enforces that any workflow advertising numbered steps documents each one (#525, @CBOSSX).
- planning-and-task-breakdown no longer silently overwrites an incomplete plan: if tasks/plan.md still has unchecked tasks for different work, it stops and asks rather than destroying state that may be mid-build in another session (#526, @ayobamiseun, from #518).
Fixes and tooling
- The code-reviewer persona now uses the same severity labels as the code-review skill (Critical / Required / Optional / Nit), so a finding is labeled the same whichever entry point runs the review.
- Validator: the section-exemption lookup uses Object.hasOwn, so a skill directory named constructor can no longer slip past the required-section checks via the prototype chain (#505, @abhisheksharma2411).
- OpenCode setup docs corrected for repo-scoped skill installation (#516, @alexmarco, closes #515).
- git-workflow-and-versioning picked up pull-request vocabulary so it routes for PR-phrased asks (#508, @superandy8288).
Thanks
Thanks to everyone who contributed this release: @nucliweb, @ayobamiseun, @CBOSSX, @abhisheksharma2411, @alexmarco, @superandy8288, and @federicobartoli. And thank you to the wider community for getting us past 90,000 stars.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.7...0.6.8
0.6.72026-08-14
Agent Skills 0.6.7 is a fixes-and-content release: a live Codex hook fix, honest plugin-manifest versions, a new provider, real skill-content additions, and a round of CI guards.
Highlights
- Codex SessionStart hook fixed (#474, @ayobamiseun). The session-start hook now emits the standard SessionStart envelope on every path, so Codex CLI (and Claude Code) stop rejecting it. Resolves the hook failure in #465.
- Plugin manifests now tell the truth about versions (#451, @Liyuk, closes #440). Every manifest is pinned to the release version, with a CI check that keeps them in lockstep with the tag, so update detection works and the versions can't silently drift again.
- Command Code is a native provider (#450, @vipulgupta2048). Command Code reads the pack's skills/ directly via cmd skills, so it's real support, not just a listing. See docs/commandcode-setup.md.
- Shared reference links fixed and guarded (#469, @coolTheWorld, closes #468). The skill links to the shared references/ checklists were wrong even in a full install; they now resolve, with a CI validator to keep them that way.
Skills and content
- spec-driven-development: a Phase 0 capability map for large, multi-capability requests, propose a module map and dependency order before speccing, while leaving the simple single-spec path untouched (#476, @ayobamiseun, from #431).
- planning-and-task-breakdown: the task-list target is now pluggable, so teams on an external tracker (beads, Jira, Linear) aren't forced into tasks/todo.md (#487, @ayobamiseun, closes #482).
- api-and-interface-design: an idempotency-key implementation guide, derive the key from the intent, not the attempt (#479, @abhisheksharma).
- security-and-hardening: a Data Privacy & Compliance section, classify data as you add it (#470, @HMAKT99).
- source-driven-development: retrieval safety, treat fetched docs as untrusted data and never hardcode outbound endpoints from examples (#53, @KoolP).
Tooling and fixes
- Codex install flow completed and plugin metadata aligned (#452, @dev-minggyu).
- Antigravity install path corrected for current agy releases (#481, @yunaremaia, from #445).
- Eval hardening: reject incomplete grader results (#473, @NgoQuocViet200), a more discriminating TDD case (#471, @federicobartoli), throwaway-workspace cleanup (#458) and --behavioral skill-name validation (#456, @Sunil56224972).
- New CI guards: artifact-path validator for spec/plan/todo (#464, @nucliweb) and Node 24 runtimes (#459, @ENNEADLABS).
Thanks
Thanks to everyone who contributed this release: @ayobamiseun, @Liyuk, @coolTheWorld, @vipulgupta2048, @HMAKT99, @abhisheksharma, @federicobartoli, @NgoQuocViet200, @yunaremaia, @dev-minggyu, @KoolP, @Sunil56224972, @nucliweb, and @ENNEADLABS.
Full Changelog: https://github.com/addyosmani/agent-skills/compare/0.6.6...0.6.7