Validates skill outputs against frozen expectations in a bounded PASS/FAILED retry loop.
npx skills add https://github.com/hanh-nd/agent-kit --skill validate-hanh-nd
Please help me install this Agent Skill. Skill: validate Source: https://github.com/hanh-nd/agent-kit/tree/main/plugins/agent-kit/.gemini/skills/validate Command: npx skills add https://github.com/hanh-nd/agent-kit --skill validate-hanh-nd
with /validate, running the producer end-to-end and validating its artifact with an isolated validator subagent. - Frozen Expectation Contract: Hashes the expectation source before the loop starts so acceptance criteria cannot drift mid-run; any change halts the run. - Bounded Retry with Verbatim Feedback: Feeds the validator's FAILED report back to the producer unchanged, up to a configurable budget (default 3 attempts), then reports PASS, FAILED, or PARTIAL. - Use Case: Run /code @plans/plan.md with /validate --budget=2 to have the code skill implement a plan, then automatically verify the implementation against the plan and retry once if blockers are found. ## Quick Start Append "with /validate" to any skill command, such as asking the agent to run /plan ticket YR-1234 with /validate, and it will produce the artifact and verify it against the ticket before reporting a final verdict.High-intent search queries and answers about installing and using this skill.
Append "with /validate" to any skill command, such as "/code @plans/plan.md with /validate". The validate skill runs the producer end-to-end, then spawns an isolated validator subagent that returns PASS or FAILED with blockers, looping until pass or budget exhaustion.
Use standalone Mode B: run "/validate @artifact --against @expectation-or-text". This skips the producer run and emits a single verdict, which is useful when the artifact came from a previous session or an external tool that cannot be re-invoked.
Yes, via the opt-in --isolate flag, but only for allowlisted non-interactive producers like code and delegate. Interactive producers must run in main context because subagents cannot handle user turn-taking; requesting isolation for them halts the run.
When attempts are exhausted (default budget of 3, overridable with --budget=N), the run ends with Status PARTIAL. The final report includes the artifact's last state plus the final validator's FAILED report verbatim so you can decide next steps.
The validator runs as a subagent with fresh context and no exposure to the producer's reasoning trace, because producers tend to rationalize away their own flaws. Verdicts are returned verbatim and PASS only exists when the validator literally returns PASS.
It requires a resolvable expectation (design brief, ticket, or --against source) and halts if none exists. It also halts if the expectation file changes mid-loop, if the producer halts before producing an artifact, or if the validator subagent fails twice.
Ben2pcCoordinate structured brainstorming sessions to produce design briefs.
logo-solutionsSimplify source code while preserving exact behavior and passing tests.
norrbackaDelegate independent debugging or implementation problems into parallel agent tasks.
OpenLAIRScore research ideas across five dimensions with three reviewer personas and Area Chair synthesis.