hashgraph-online/awesome-codex-plugins / plugins/alpoxdev/hypercore/skills/hyper-mass
hyper-mass
Verified[Hyper] Use this skill when the user asks for dependency-ordered multi-agent work as a graph: a task DAG, a mass run, "의존성 그래프", or "단계별 병렬", where some tasks must finish before others start and the ordering itself is the point. Plans one phase as explicit nodes with tiers, write scopes, handoff files, numeric caps, and a verification node, then runs it through a native workflow capability or leader-run topological waves with retry, amend, send, and cancel. Do not use for independent parallel tasks with no ordering, executing a written plan file checkbox by checkbox, or producing a research report.
| File | Kind | Size |
|---|---|---|
| SKILL.ko.md | markdown | 20.6 KB |
| SKILL.md | skill md | 17.9 KB |
| assets/evals/hyper-mass-cases.jsonl | other | 18.5 KB |
| references/graph-schema.ko.md | markdown | 18.8 KB |
| references/graph-schema.md | markdown | 17.4 KB |
| rules/graph-planning.ko.md | markdown | 15.9 KB |
| rules/graph-planning.md | markdown | 13.8 KB |
| rules/run-and-recovery.ko.md | markdown | 14.2 KB |
| rules/run-and-recovery.md | markdown | 12.4 KB |
references/graph-schema.md
View on GitHub# Graph Schema
Use this reference while writing the run definition and the journal. `graph.json` is the contract; `run.jsonl` is the record of what happened to it. Both live under `.hyper/mass/<run-key>/`.
## Contents
- Run Directory
- `graph.json`
- Node Object
- Justification Object
- Node Prompt
- Self-Check
- `run.jsonl`
- Resume Matrix
- Worked Example
- Key and Path Rules
## Run Directory
```text
.hyper/mass/<run-key>/
├── graph.json # the definition, written before start
├── run.jsonl # append-only event journal
└── nodes/
├── <node-id>.md # one handoff per node
└── ...
```
The leader writes `graph.json` and `run.jsonl`. Nodes write only their own handoff, and only inside `nodes/`. A node that needs scratch space uses a temporary directory and removes it before finishing.
## `graph.json`
| Field | Type | Required | Meaning |
|---|---|---|---|
| `key` | string | yes | Run identity. Lowercase kebab-case, at most 48 characters. Same key plus same definition is a resume. |
| `name` | string | yes | Human label for reports. |
| `goal` | string | yes | One sentence naming the deliverable this run produces. |
| `success_criteria` | array of strings | yes | Each criterion names its evidence: the command and expected result, or the file and required content. |
| `justification` | object | yes | The six-field record from `rules/graph-planning.md` §4. |
| `nodes` | array of objects | yes | 1-32 node objects. |
| `phase` | string | no | The checkpoint this run covers, when an outer loop owns the goal. |
| `caps` | object | no | User-given overrides for nodes, concurrency, retries, or amends. |
Unknown fields are ignored; the six above are validated by the self-check.
## Node Object
| Field | Type | Required | Meaning |
|---|---|---|---|
| `id` | string | yes | Unique inside the run; lowercase kebab-case; used in the handoff file name. |
| `tier` | `low` \| `medium` \| `high` | yes | Difficulty. `low` is the default; one-line reason for anything higher. |
| `kind` | `writing` \| `visual` \| `git` | no | Specialty kind, chosen by the kind of work, never by difficulty. |
| `dependsOn` | array of strings | yes, may be empty | Node ids that must complete first. Orders work; carries no data. |
| `reads` | array of strings | yes, may be empty | Exact paths or globs the node may read. |
| `writes` | array of strings | yes | Exact paths or globs the node may change, including its handoff. |
| `handoff` | string | yes | `.hyper/mass/<run-key>/nodes/<id>.md`. Must be inside the run directory. |
| `prompt` | string | yes | The complete node prompt with the five markers. |
| `reason` | string | when `tier` is not `low` | One line explaining the escalation. |
| `solutionSpace` | array of strings | no | Short approach names this node may choose between. When present, the node prompt lists it and the node **MUST** pick only from the list; a node that uses an approach outside the list fails the self-check. |
| `resolvedModel` | string | no | The model that actually ran. It is written only from the observed result, never predicted before dispatch. |
`dependsOn` and `reads` serve different purposes: an edge orders two nodes, a read path names what a node consumes. A path in `reads` that no edge produces is either leader-supplied context or a mistake; decide which before starting.
When `solutionSpace` is present, the node prompt lists it under a `SOLUTION SPACE:` line (or inside `TASK`/`SCOPE`) so the worker sees the allowed approaches. The node names the approach it used in its handoff; the leader checks that the named approach belongs to the list before the node counts as complete. Without `solutionSpace` the node is free to choose; with it the choice is bounded and checked.
## Justification Object
```json
{
"ceiling_removed": "",
"topology": "",
"state": "",
"bounds": "",
"approval": "",
"rollback": ""
}
```
Validation rules:
- Every field is a non-empty string.
- `approval` is either the approval answer for each irreversible node or the literal `not required`.
- `rollback` names the concrete undo: branch, commit range, paths to revert, or "no files changed".
- `bounds` repeats the caps in force, including user overrides.
## Node Prompt
```text
TASK: <one imperative sentence>
DELIVERABLE: <artifact: files, report shape, evidence>
SCOPE: <exact read paths; exact write paths; what is out of scope>
VERIFY: <literal command and expected result>
STOP WHEN: <one observable condition>
```
A prompt missing any marker fails the self-check. A prompt that names a handoff but not the path, or a path outside the run directory, fails it too.
## Self-Check
Run before `run.created`:
1. `key` matches `^[a-z0-9]+(-[a-z0-9]+)*$`, is at most 48 characters, and contains no `..`, `/`, or absolute path.
2. `goal` is one sentence; `success_criteria` is non-empty and every entry names evidence.
3. `justification` has six non-empty fields.
4. Node count is within the cap; at most one node has `tier: high`.
5. Every `id` is unique and matches `^[a-z0-9]+(-[a-z0-9]+)*$`.
6. Every `dependsOn` id exists in the graph.
7. The graph is acyclic. A cycle is rejected before start; it is never repaired during a run.
8. Every wave has at least one runnable node.
9. No two nodes that can run in parallel share a `writes` path. Overlap means serialize or merge.
10. Every `handoff` equals `.hyper/mass/<key>/nodes/<id>.md` and is also listed in `writes`.
11. Every `prompt` carries TASK, DELIVERABLE, SCOPE, VERIFY, and STOP WHEN.
12. At least one terminal verification or review node depends on every producer.
13. The graph changes code: the terminal node's VERIFY names the real test or build command.
14. Every `solutionSpace` entry is a non-empty string, every approach a node reports using belongs to its own list, and every `resolvedModel` value came from an observed result rather than a prediction.
## `run.jsonl`
One JSON object per line, append-only. Common fields: `ts` (ISO 8601 UTC), `event`, `run`.
| Event | Additional fields |
|---|---|
| `run.created` | `mode`, `nodes`, `definition_digest`, `justification` |
| `run.resumed` | `from_event`, `reused_nodes` |
| `node.started` | `node`, `attempt` |
| `node.completed` | `node`, `handoff`, `evidence`, `resolvedModel` (the model observed in the result, run journal only); `verifier` on verification nodes (`self (no subagent capability)` in sequential mode) |
| `node.failed` | `node`, `attempt`, `error` |
| `node.retried` | `node`, `attempt`, `reason` |
| `node.sent` | `node`, `boundary` |
| `node.cancelled` | `node`, `reason` |
| `run.amended` | `changed`, `reason` |
| `run.settled` | `status`, `reason`, `criteria`, `landed` |
Example:
```json
{"ts":"2026-11-03T09:00:00Z","event":"run.created","run":"checkout-refactor-2026-11-03","mode":"waves","nodes":6,"justification":{"ceiling_removed":"serialized latency","topology":"5 components","state":"handoffs and worktree","bounds":"6 nodes, 4 concurrent","approval":"not required","rollback":"revert branch"}}
{"ts":"2026-11-03T09:00:05Z","event":"node.started","run":"checkout-refactor-2026-11-03","node":"audit","attempt":1}
{"ts":"2026-11-03T09:04:10Z","event":"node.completed","run":"checkout-refactor-2026-11-03","node":"audit","handoff":".hyper/mass/checkout-refactor-2026-11-03/nodes/audit.md","evidence":"13 call sites listed"}
{"ts":"2026-11-03T09:04:11Z","event":"node.started","run":"checkout-refactor-2026-11-03","node":"api","attempt":1}
{"ts":"2026-11-03T09:30:00Z","event":"node.failed","run":"checkout-refactor-2026-11-03","node":"api","attempt":1,"error":"handoff missing; node exited before writing"}
{"ts":"2026-11-03T09:31:00Z","event":"node.retried","run":"checkout-refactor-2026-11-03","node":"api","attempt":2,"reason":"missing handoff, rerun with write-first instruction"}
{"ts":"2026-11-03T10:20:00Z","event":"run.settled","run":"checkout-refactor-2026-11-03","status":"completed","reason":"all criteria pass","criteria":["checkout tests exit 0","e2e log attached"],"landed":["9f2c1ab","41d7e02"]}
```
Rules:
- `node.completed` is written after the leader opens the handoff, not when the node replies. The leader sets `resolvedModel` from the observed result (the runtime's report of which model ran) at that moment; it is never written before dispatch and never predicted from the `tier`.
- A node whose `solutionSpace` is present names the approach it used in its handoff. The leader checks it against the list before `node.completed`; a match outside the list is a failed self-check, not a completion.
- `run.settled` is written once per run; a resumed run appends a new settlement only after a new dispatch wave.
- Timestamps are UTC; the report may render local time, but the journal does not.
## Resume Matrix
| Journal state | Handoff file | Resume as |
|---|---|---|
| `node.completed` | exists and non-empty | completed, do not re-run |
| `node.completed` | missing or empty | failed; attempt count unchanged |
| `node.started` only | any | pending; dispatch again in the ready set |
| `node.failed` | any | failed; retry only if attempts remain |
| `node.retried` after `node.failed` | any | failed with a higher attempt count |
| `node.cancelled` | any | cancelled; retry only when the goal is re-opened |
| no event | any | pending |
A resumed run journals `run.resumed` before dispatching anything.
## Worked Example
A five-component refactor, one phase:
```json
{
"key": "checkout-refactor-2026-11-03",
"name": "Checkout refactor",
"goal": "Move checkout onto the new payment client without changing user-visible behavior.",
"success_criteria": [
"`npm test -- checkout` exits 0 and its output is captured in nodes/verify.md",
"`npm run e2e -- checkout` exits 0 and its log is captured in nodes/verify.md"
],
"justification": {
"ceiling_removed": "serialized latency: three independent surfaces plus one integration proof",
"topology": "C1 audit, C2 API adapter, C3 UI copy, C4 docs, C5 end-to-end proof",
"state": "handoff files and the working tree only; no shared in-memory state",
"bounds": "6 nodes, at most 4 concurrent, 2 retries per node, 3 amends",
"approval": "not required: no deploy, spend, or external send in this phase",
"rollback": "revert the branch commits listed in nodes/verify.md; no data migration"
},
"nodes": [
{
"id": "audit",
"tier": "low",
"dependsOn": [],
"reads": ["src/checkout/**", "src/payments/**"],
"writes": [".hyper/mass/checkout-refactor-2026-11-03/nodes/audit.md"],
"handoff": ".hyper/mass/checkout-refactor-2026-11-03/nodes/audit.md",
"prompt": "TASK: List every checkout call site that uses the old payment client.\nDELIVERABLE: `.hyper/mass/checkout-refactor-2026-11-03/nodes/audit.md` containing a table with file, line, current call, and replacement call.\nSCOPE: read src/checkout/** and src/payments/** only. Write only `.hyper/mass/checkout-refactor-2026-11-03/nodes/audit.md`.\nVERIFY: every row names an existing file and line; run `grep -rn oldPayment src` and paste the count.\nSTOP WHEN: every call site in those directories is listed or marked read-only."
},
{
"id": "api",
"tier": "low",
"dependsOn": ["audit"],
"reads": ["src/payments/**", ".hyper/mass/checkout-refactor-2026-11-03/nodes/audit.md"],
"writes": ["src/payments/**", ".hyper/mass/checkout-refactor-2026-11-03/nodes/api.md"],
"handoff": ".hyper/mass/checkout-refactor-2026-11-03/nodes/api.md",
"prompt": "TASK: Replace the old payment client with the new adapter at every call site listed in nodes/audit.md.\nDELIVERABLE: changed files plus a handoff listing each replacement and the command output.\nSCOPE: write only src/payments/** and nodes/api.md. Do not touch src/checkout/ui/** or tests/**.\nVERIFY: run `npm test -- payments` and paste the exit code and the final summary line.\nSTOP WHEN: the test exits 0 and every audit row is replaced."
},
{
"id": "ui",
"tier": "medium",
"kind": "visual",
"reason": "copy and state changes across four checkout screens",
"solutionSpace": ["copy-update", "state-map", "a11y-pass"],
"dependsOn": ["audit"],
"reads": ["src/checkout/ui/**", ".hyper/mass/checkout-refactor-2026-11-03/nodes/audit.md"],
"writes": ["src/checkout/ui/**", ".hyper/mass/checkout-refactor-2026-11-03/nodes/ui.md"],
"handoff": ".hyper/mass/checkout-refactor-2026-11-03/nodes/ui.md",
"prompt": "TASK: Update the checkout screens to the new payment states named in nodes/audit.md.\nDELIVERABLE: changed components plus a handoff listing each screen and its state mapping.\nSCOPE: write only src/checkout/ui/**. Do not touch src/payments/** or tests/**.\nVERIFY: run `npm run build` and paste the exit code; confirm each mapped state appears in the built bundle.\nSTOP WHEN: the build exits 0 and every listed screen has a mapped state."
},
{
"id": "docs",
"tier": "low",
"kind": "writing",
"dependsOn": ["audit"],
"reads": ["docs/checkout.md", ".hyper/mass/checkout-refactor-2026-11-03/nodes/audit.md"],
"writes": ["docs/checkout.md", ".hyper/mass/checkout-refactor-2026-11-03/nodes/docs.md"],
"handoff": ".hyper/mass/checkout-refactor-2026-11-03/nodes/docs.md",
"prompt": "TASK: Update docs/checkout.md to describe the new payment client and its states.\nDELIVERABLE: the edited page plus a handoff listing each changed section.\nSCOPE: write only docs/checkout.md and nodes/docs.md. Do not change code.\nVERIFY: every code identifier in the page exists in src/; paste the check command and its output.\nSTOP WHEN: every stale identifier is replaced and the check passes."
},
{
"id": "e2e",
"tier": "medium",
"reason": "integration surface across API and UI",
"dependsOn": ["api", "ui"],
"reads": ["tests/e2e/**", "src/payments/**", "src/checkout/ui/**"],
"writes": ["tests/e2e/**", ".hyper/mass/checkout-refactor-2026-11-03/nodes/e2e.md"],
"handoff": ".hyper/mass/checkout-refactor-2026-11-03/nodes/e2e.md",
"prompt": "TASK: Add and run the end-to-end checkout test for the new payment client.\nDELIVERABLE: the test file plus a handoff with the command, exit code, and log path.\nSCOPE: write only tests/e2e/** and nodes/e2e.md. Do not change product code; report a product defect instead.\nVERIFY: run `npm run e2e -- checkout` and paste the exit code and the last 20 log lines.\nSTOP WHEN: the suite exits 0 or the failing assertion is written to the handoff."
},
{
"id": "verify",
"tier": "medium",
"reason": "independent judgment over every producer",
"dependsOn": ["audit", "api", "ui", "docs", "e2e"],
"reads": [".hyper/mass/checkout-refactor-2026-11-03/nodes/**"],
"writes": [".hyper/mass/checkout-refactor-2026-11-03/nodes/verify.md"],
"handoff": ".hyper/mass/checkout-refactor-2026-11-03/nodes/verify.md",
"prompt": "TASK: Verify the checkout refactor against the run's success criteria.\nDELIVERABLE: nodes/verify.md with each criterion, the literal command, exit code, and captured output.\nSCOPE: read every handoff and the repository; write only nodes/verify.md. Do not fix product code.\nVERIFY: `npm test -- checkout` and `npm run e2e -- checkout` both exit 0; every producer handoff exists and is non-empty.\nSTOP WHEN: every criterion has a captured result, or a failing criterion names the producer and the exact failure."
}
]
}
```
The audit fans out to three disjoint surfaces, `e2e` joins the two code producers, and `verify` depends on every producer, `audit` included because its handoff feeds all three surfaces. No second `high` node exists, and no two parallel nodes share a write path.
## Key and Path Rules
- `key` and `id` use lowercase ASCII kebab-case: `^[a-z0-9]+(-[a-z0-9]+)*$`. Reject `..`, `/`, and absolute paths.
- Paths in `reads`, `writes`, and `handoff` are repository-relative or absolute, exactly as the node will see them. Globs are allowed in `reads` and `writes`, never in `handoff`.
- The handoff lives inside the run directory so a resumed run can find it without guessing.
- One run, one directory. Never reuse a directory for a different key or a changed definition; that is what `amend` is for.
## Sources
> Checked 2026-11-09.
| Claim | Source |
|---|---|
| The run directory shape, node fields, dependency rules, wave execution, journal events, and recovery semantics, adapted from the `mass-ulw` skill and its planning reference | <https://github.com/code-yeongyu/oh-my-openagent/tree/0d80470e4201925727e88afbafcdf24d7801a588> |
| Cycle bounds, missing is not empty, a verifier separate from the producer, and fan-out caps | `instructions/graph-engineering/GRAPH_ENGINEERING.md` |
| Checkpoint resume tested rather than assumed, and the transition record for a dynamic path | `instructions/graph-engineering/GRAPH_ENGINEERING.md` |
| Evidence file or it did not happen, and `SKIP` is not a pass | `instructions/harness-engineering/HARNESS_ENGINEERING.md` |
| The family contract's `.hyper/mass/<run-key>/` paths and the six justification fields | family contract, 2026-11-03 |
| Node `solutionSpace` bounds, approach-outside-list failure, `resolvedModel` from observed results, cluster digest with a token bound and handoff citations, and the final reducer reading digests only | `dbc-hbin/omp-ulw @ 74df40f` (2026-11-09), re-authored; the source repository has no license file and nothing was copied |
### Evidence grade
`LOCAL` - the schemas are this package's implementation of the family contract. Field names match the contract; `run.resumed`, `node.sent`, `node.cancelled`, `phase`, `caps`, and `reason` are additive and documented here.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 469k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |