graph-engineering

Files (1)

• SKILL.md

SKILL.md (first lines)

> Hermes: first load `skill_view("graph-powers:graph-engineering")` for native calls and host-policy limits. `content/` paths are `file_path` values relative to the common registered-document parent. Source tools/model frontmatter is not host enforcement.

# Graph engineering in Hermes

Graph Powers is canonical upstream; this adapter only translates names and limits Hermes cannot
provide. Read registered documents with `skill_view("graph-powers:<name>")`. In the generated
package every registration shares the `skills/` parent. Read an auxiliary file with
`skill_view("graph-powers:graph-engineering", file_path="content/<source-path>")`; paths never
traverse above that common parent. Auxiliary content is served raw, without template expansion.
Resolve the installed absolute content path before terminal execution; preserve the declared
host-project working directory. Do not maintain another inventory.

| Graph Powers | Hermes |
|---|---|
| Agent invocation | `delegate_task(goal=..., context=...)` with the selected `agent-<slug>` contract |
| Graph Powers skill invocation | `skill_view("graph-powers:<name>")` |
| Read/Grep/Glob/Bash | `read_file` / `search_files` / `terminal` |
| Web tools | `web_search` / `web_extract`, unless a needed MCP is configured |
| Background read-only spawn | set `background=true` explicitly; the native default is false |
| Native hooks | Unsupported: parent carries safety rules and approvals |
| Child clarification | parent `clarify`; child returns the decision point |

Slash names identify contract documents, not installed Hermes slash commands. Native `Workflow`
is unsupported; use the planning method and native delegation. When a document names `hooks/...`,
it identifies an external Claude-only source for inspection during an explicitly authorized
external-client task. Hermes never fetches, copies, installs, or executes those hooks. Source-client
workflow authoring and setup guides remain inspection material and grant no execution permission.
No model-family or tool-list metadata can bypass actual host policy.

## 1. Classify and load only needed context

Use `content/references/shared/020-complexity-routing.md` and `content/references/shared/025-solution-ladder.md`. L1-L2 stay local;
L3 may delegate once; L4-L5 use a disjoint batch; L6+ plans before waves. Read project config when
present, then only matching rules and the nearest authority. Missing config/rule is a stated gap,
not an error or an invented default.

## 2. Choose and dispatch a specialist

Use `content/references/shared/030-agent-assignment-matrix.md` and `content/references/shared/060-skill-domain-matrix.md`. State the agent contract,
why it fits, required skills and deliberate omissions, and the outcome. Read the contract through
`skill_view("graph-powers:agent-<slug>")`; a read-only contract must become an explicit child
`MUST NOT DO`, because Hermes children inherit parent tools.

The `model`, `tools`, `disallowedTools`, `skills` and `memory` frontmatter preserves the source
contract, not native Hermes enforcement. The parent supplies method context, enforces the actual
host policy and includes read-only restrictions in the child prompt. Read-only wording alone
does not remove write-capable tools. Resolve the model through the live host configuration;
Claude model-family labels are not Hermes model IDs. User approvals and host policy prevail.

Use one `delegate_task(tasks=[...])` call for independent tasks. Each child gets the canonical
seven-section prompt and Context Handoff from
`content/references/execution-floor.md §4` and
`content/skills/senior-prompt-engineer/references/agent-handoff-contracts.md`. Declare exact disjoint `Owns:` paths.
Read live `delegate_task` limits; do not hardcode concurrency or models. Children do not delegate,
clarify, use memory, or message. They have their own working directory, so pass absolute paths.

## 3. Implement and Gauntlet translation

`/implement` and `/verify loop` are **NOT SUPPORTED** Hermes slash commands; `/gauntlet` is a
deprecated alias of `/implement`. When requested, read `graph-powers:implement` and
`graph-powers:verify` and follow their entry contract: route by tier, L1-L2 direct and L3 without the
Gauntlet profile. An L4+ objective goes to the canonical planning phases first; reuse valid existing
approval. Dry-run remains read-only and dispatches nobody. For an approved L4+ plan, validate with
the packaged `content/skills/planning/scripts/sdd.py` with `validate ... --profile gauntlet`; invalid input
routes to planning. For a validated non-dry L4+ run, acquire the Gauntlet lease before writers,
use disjoint waves within live limits, run each focused CHECK, then an independent critic. A capped
or blocked lane is not success. Release only after PASS.

## 4. Verify and close

Run or reuse the project-declared checks only while their diff, config, and environment remain
relevant and unchanged. Capture exit codes and actual status; a child report is not evidence.
Undeclared JS/TS gates route to `graph-powers:debugger`. Do not add a verification specialist for
ordinary L1-L2 work.

`VERIFIED` means all required signals passed; `VERIFIED-WITH-NOTES` has only documented minor
issues; `NEEDS-WORK` means a required signal failed or could not run. A regression fix needs a test
that would fail before it. Resolve an independent verification specialist once for final Gauntlet
close. Translate the `graph-powers:verify` loop fallback with native delegation, including its
independent review and bounded correction; report the workflow degradation once. If a required
independent reviewer is unavailable, return `NEEDS-WORK`; self-review cannot close Gauntlet.

## 5. External CLI routing

Use `terminal` with a narrow workdir and explicit safety context. Claude Code is preferred for
implementation; Codex CLI plans detailed changes, reviews work it did not write, or is the fallback
after Claude Code stalls. Inspect the resulting diff and verify it yourself.

## Safety carried into every child or CLI

- No commit, push, protected-branch checkout, merge, rebase, tag, PR, release, deploy, or
  publication without explicit approval for that action and scope. The same approved action remains
  authorized in this session; it does not authorize another action.
- Propose irreversible data work with the exact statement and stop for approval.
- Never expose secrets, weaken production/auth controls, violate tenant isolation, or widen scope.
- An approval denial is a gate, not a route-around.

If an agent name fails, follow `content/references/shared/035-agent-resolution-recovery.md` once. A missing
write-capable specialist blocks the task. Report changed paths, checks with exit codes, and blockers.

From GrupoUS/graph-powers, MIT.

More from GrupoUS/graph-powers