Agent Skills' spec workflow now stops and waits for your approval before planning
Addy Osmani's Agent Skills changed its spec-driven-development skill on October 3rd. Once the spec is saved, the skill now ends its turn and waits for your approval before any planning starts.
On Codex, one model ran spec, plan, tasks and code in a single turn
“I need to perform a manual review after the spec document is generated”
A Codex user running a DeepSeek model asked for a spec to review. Instead the skill ran straight through planning, tasks and implementation in one turn. It wrote code before the user could review the spec.
The pause lived in the /spec command, which Codex never loads
The instruction, confirm with the user before proceeding, lived only in the /spec command for Claude Code. Codex loads just the skill, which said to validate each phase but not who validates, or that the agent should wait.
The skill now tells the agent to stop its turn after the spec
Now the skill itself carries the gate. The agent summarizes the spec, lists open questions, and asks for approval. Then it stops its turn. Planning starts only after you approve, in a later turn.
Claude Opus and Haiku already stopped, so the fix is unmeasured there
The run-through was not reproduced with Claude models.
The author tested Claude Opus and Haiku, and both already stopped after the spec, with or without the change. So the run-through wasn't reproduced on Claude, and the fix is unmeasured there.
The stop covers the spec phase only
The change covers the spec phase and nothing past it. The author says the capability map that comes before it, and the /build command, have similar gaps, and left both out of this fix.
The launch checklist no longer tells you to run a Prisma command that doesn't exist
The shipping-and-launch skill's rollback plan told you to run a Prisma migrate rollback command, and Prisma has no such command. The template now asks for your project's verified rollback command or runbook, whatever migration tool you use.
The idempotency example now parses as JavaScript
In the api-and-interface-design skill, the idempotency example ended its catch block with a bare throw, which is a syntax error in JavaScript. It now rethrows the caught error.
performance-optimization now loads at 267 lines instead of 496
Worked fixes moved to references/optimization-patterns.md
The performance-optimization skill got shorter. It went from almost 500 lines to about 270, roughly a quarter fewer tokens each time it loads. The worked code fixes moved into a reference file the agent opens when it needs one.
Code review now checks the tests by breaking one condition
The code-review-and-quality skill asked whether the tests would catch a regression, without saying how to find out. Now it tells the reviewer to invert one condition the change adds and run the suite. If the suite stays green, that's a missing test.
Setup guides for Oh My Pi and Dojo Workspace shipped, then were withdrawn
New rule in CONTRIBUTING: the README lists only hosts a maintainer has run.
The release shipped setup guides for Oh My Pi and Dojo Workspace. Both were withdrawn on the main branch that day, because the maintainers hadn't run either host. The contributor guide now says the readme lists only hosts a maintainer has run, and the hosts already there will be checked against that.
Close
The spec stop is written for models that skip an implied gate. On Claude Opus and Haiku it was already the behavior.
Agent Skills' spec workflow now stops and waits for your approval before planning
The spec stop is written for models that skip an implied gate. Claude Opus and Haiku already stopped. Agent Skills' spec workflow now stops and waits for your approval before planning. The notes are on the project's GitHub releases page.
