hw-verification-bar

Defines adversarial verification checks that branches must survive before merging.

Updated Aug 8, 2026
One-click install
npx skills add https://github.com/headwater-ai/headwater --skill hw-verification-bar-headwater-ai
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: hw-verification-bar
Source: https://github.com/headwater-ai/headwater/tree/main/.claude/skills/hw-verification-bar
Command: npx skills add https://github.com/headwater-ai/headwater --skill hw-verification-bar-headwater-ai

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? A green test suite, linter, and CLI run catch almost nothing the build agent has not already caught. This Skill supplies the adversarial checks and review questions a verifier must run to find defects that a passing suite cannot, so branches are not merged on the strength of exit code 0 alone. ## Core Features & Use Cases - Adversarial check catalog: Twelve named attacks such as "Make the new thing fail", "Test a gate in three directions", and "Re-derive one number by hand", each targeting a specific class of defect a green suite misses. - Review questions: Rulings on injectivity of cache keys, Option types at parse boundaries, empty arms, and checks that cannot run, grounded in recorded decision documents. - Verifier self-suspicion protocol: Recorded failure shapes (wrong denominator, stale locks, git-dependent suites over archives) so the verifier suspects its own check before contradicting the build note. - Use Case: Before merging a branch that adds a new lint rule, the verifier injects defects by hand, confirms the rule fires with a message naming the offender, predicts the finding count before running, and re-derives one reported number with an independent script. ## Quick Start Invoke the hw-verification-bar skill before verifying a branch as hw-verify and run the three mandatory checks: make the new thing fail, re-derive one number by hand, and suspect your own check first.

Frequently Asked Questions about hw-verification-bar

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I verify a branch beyond running the test suite?

Run adversarial checks that a green suite cannot catch: break the corpus by hand to confirm new rules fire, inject defects into new material, and re-derive one claimed number with your own script. A passing suite only proves the build agent's own checks passed.

What is the verification bar in the Headwater workflow?

It is the set of adversarial checks and review questions a branch must survive before merging, invoked before verifying any branch as hw-verify. Three checks run on every branch: make the new thing fail, re-derive one number by hand, and suspect your own check first.

How do I test that a new lint rule actually works?

Break the input by hand and confirm the rule fires with a message naming the offender, since a rule that fires on no input is indistinguishable from one that does not work. Also check the rule gained instances, not only existence, and predict the finding count before running.

Why does a passing test suite not prove a fix works?

A suite can stay green because the regression test is narrower than its name, the check cannot run at all, or the defect lives one component downstream. Revert the fix and watch a named test go red with a regression the builder did not use to confirm the fix is held.

When should I get a second opinion from a different model?

Spend it on a ruling about to merge, a claim you cannot cheaply re-derive where being wrong costs a milestone, or the largest change of a run, never on routine merges. Ask it to refute rather than review and verify its checkable claims before acting.