Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use Registry Broker to discover and summon specialist agents for focused subtasks from inside Codex.
.claude/skills/hashgraph-online-registry-broker-orchestrator/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | -34% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 33% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 63% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 44% | 0% |
This is the Codex-specific wrapper skill.
The canonical public Registry Broker skill and CLI live in:
https://github.com/hashgraph-online/registry-broker-skills@hol-org/registryUse this plugin when a task would benefit from a specialist broker agent inside Codex instead of only local reasoning.
registryBroker.delegate to see where specialist help would actually add leverage.delegate-now means summon-ready, review-shortlist means inspect candidates first, and handle-locally means keep the work local unless the user has a known target.
registryBroker.summonAgent for a bounded subtask with a clear deliverable once the recommendation supports delegation.mode: "best-match" when one strong answer is enough.mode: "fallback" when you want the top ranked candidate first and a backup if the first message fails.mode: "parallel" only when comparing multiple approaches is useful.dryRun: true when you want to preview the exact outbound dispatch before opening a broker session.message when the target agent expects a very direct prompt or protocol-specific phrasing.Use these when the delegated subtask needs a stronger contract than a single sentence:
deliverable for the exact artifact you want backconstraints for hard limits the delegate must respectmustInclude for required sections or factsacceptanceCriteria for what makes the response usableregistryBroker.findAgentsreview-shortlist and you want the next action to stay obvious.dryRun, treat the returned dispatch plan as the last check before sending.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-08 | fail→pass | 9,766 | 7,273 | -26% | 1 | 1 | 0% | 1,658 | 1,097 | -34% | 0 | 0 | — |
case-01 | fail→fail | 17,564 | 18,199 | +4% | 1 | 1 | 0% | 1,815 | 1,037 | -43% | 0 | 0 | — |
case-02 | fail→fail | 26,951 | 64,245 | +138% | 1 | 1 | 0% | 4,457 | 2,272 | -49% | 0 | 0 | — |
case-07 | fail→pass | 14,226 | 12,210 | -14% | 1 | 1 | 0% | 1,418 | 1,885 | +33% | 0 | 0 | — |
case-03 | fail→fail | 16,650 | 12,844 | -23% | 1 | 1 | 0% | 2,610 | 1,176 | -55% | 0 | 0 | — |
case-04 | pass→fail | 11,482 | 20,432 | +78% | 1 | 1 | 0% | 882 | 1,194 | +35% | 0 | 0 | — |
case-05 | fail→fail | 9,544 | 22,450 | +135% | 1 | 1 | 0% | 1,344 | 1,203 | -10% | 0 | 0 | — |
case-06 | fail→pass | 5,595 | 3,822 | -32% | 1 | 1 | 0% | 826 | 1,346 | +63% | 0 | 0 | — |
case-09 | fail→pass | 12,384 | 3,695 | -70% | 1 | 1 | 0% | 1,120 | 1,281 | +14% | 0 | 0 | — |
case-10 | fail→pass | 18,259 | 2,645 | -86% | 1 | 1 | 0% | 788 | 1,132 | +44% | 0 | 0 | — |
case-11 | fail→fail | 12,154 | 3,752 | -69% | 1 | 1 | 0% | 1,141 | 1,272 | +11% | 0 | 0 | — |
case-12 | fail→pass | 10,915 | 3,653 | -67% | 1 | 1 | 0% | 838 | 1,237 | +48% | 0 | 0 | — |
case-13 | pass→fail | 4,330 | 8,631 | +99% | 1 | 1 | 0% | 662 | 1,284 | +94% | 0 | 0 | — |
case-14 | pass→pass | 6,917 | 8,871 | +28% | 1 | 1 | 0% | 1,159 | 1,212 | +5% | 0 | 0 | — |
case-15 | pass→pass | 7,564 | 9,345 | +24% | 1 | 1 | 0% | 1,210 | 1,267 | +5% | 0 | 0 | — |
case-16 | pass→pass | 18,960 | 10,990 | -42% | 1 | 1 | 0% | 2,724 | 1,254 | -54% | 0 | 0 | — |
case-17 | pass→pass | 15,031 | 7,045 | -53% | 1 | 1 | 0% | 1,445 | 995 | -31% | 0 | 0 | — |
case-18 | fail→fail | 13,624 | 2,947 | -78% | 1 | 1 | 0% | 1,159 | 1,205 | +4% | 0 | 0 | — |
case-19 | fail→pass | 11,904 | 13,249 | +11% | 1 | 1 | 0% | 1,562 | 1,777 | +14% | 0 | 0 | — |
case-20 | pass→pass | 10,458 | 2,362 | -77% | 1 | 1 | 0% | 1,647 | 1,069 | -35% | 0 | 0 | — |
case-21 | fail→pass | 7,135 | 8,021 | +12% | 1 | 1 | 0% | 961 | 1,176 | +22% | 0 | 0 | — |
case-22 | fail→fail | 12,123 | 9,638 | -20% | 1 | 1 | 0% | 1,014 | 1,547 | +53% | 0 | 0 | — |
case-23 | pass→fail | 20,471 | 3,513 | -83% | 1 | 1 | 0% | 1,959 | 1,323 | -32% | 0 | 0 | — |
case-24 | pass→pass | 8,896 | 9,983 | +12% | 1 | 1 | 0% | 1,506 | 1,518 | +1% | 0 | 0 | — |
case-25 | fail→pass | 17,588 | 9,557 | -46% | 1 | 1 | 0% | 2,410 | 2,043 | -15% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 25 cases were attempted, and 19 counted toward the lift figure. The other 6 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +24 percentage points is the difference between those two pass rates over the 19 comparable cases. 5 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.