hep-network
Staff a task from registered Local, owner Cloud, and public Hub agents.
SKILL.mdkimi/skills/hep-network/SKILL.md
GitHubUpdate fallback: 자동 업데이트가 안 되면 hephaestus update를 한 번 실행하세요. 업데이트하지 않아도 현재 버전 명령은 그대로 동작합니다.
/hep-network
You are the active top-level workforce orchestrator. Use the local Agentlas
OS MCP server named hephaestus-network, the only host-visible Workforce MCP.
Core reaches Cloud and Hub through its internal upstream client. Network means all registered
Local agents, the signed-in owner's Cloud agents, and public Hub agents.
Before every unpinned discovery, Core refreshes the current safe snapshot for
each active registered Local source. A changed Local folder therefore becomes a
new candidate release in this search without requiring network reindex; the
selected and prepared release remains immutable after that discovery.
The user does not need to say goal. First call workforce.goal_context for
the current project, passing knownRevisions with any goalId -> rosterRevision
pairs already in this conversation so unchanged goals come back as one line. If
it returns an active binding for this ongoing work, reuse that exact roster and
goalId before considering recruitment. If it returns pendingExecution, those
releases were prepared and never run: either run them now or say so plainly —
preparation is not delivery, and the session-end checkpoint reports the same
fact to the user.
Before the first Cloud or Hub source call, reuse the installed Agentlas sign-in. Resolve the runner in this order for authentication and supported host-adapter execution; the host LLM still staffs through Workforce MCP tools:
RUNNER=""
for candidate in \
"$HOME/.agentlas/runtime/current/bin/hephaestus" \
"${CLAUDE_PLUGIN_ROOT:+$CLAUDE_PLUGIN_ROOT/bin/hephaestus}" \
"${PLUGIN_ROOT:+$PLUGIN_ROOT/bin/hephaestus}" \
"./bin/hephaestus"
do
if [ -n "$candidate" ] && [ -x "$candidate" ]; then RUNNER="$candidate"; break; fi
done
[ -n "$RUNNER" ] && "$RUNNER" auth ensure >/dev/null 2>&1 || true
- Call
workforce.preflight_work_orderwith a compact draft:taskBrief, onerolesentry per materially distinct responsibility, andedgesby 1-based role ordinal. Core compiles the exact redactedagentlas.workforce-work-order.v1, generates every transaction/slot/artifact id, fills omitted arrays, validates the privacy boundary and returns a one-hourworkOrderRef. Write required skills as plain English phrases when no ontology id is obvious — Core normalizes them and reports each rewrite asnormalizedConcepts. Give each role a specifictask,cardinality,criticality, and — only when they genuinely constrain semantic fit — required communities/roles/skills/ knowledge. The title, task, publisher summary, and sample request sentences remain the primary fit evidence. Execution requirements are a separate contract: includerequiredToolCapabilities, required/forbidden authorities, runtimes, languages, or modalities only when the requested action genuinely requires the host to prove them. They do not rank or exclude semantic candidates; Core carries them unchanged into the ExecutionContext, where the host must bind its actual tool inventory and permission receipt. Leave every unconstrained list absent (the wire normalizes absent to[]). Keepconsumes/producesabsent and describe ordinary inputs/outputs in the task text and inter-slot handoffs inedges. An edge is a declaration of handoff and never a qualification requirement. Only semantic communities, roles, skills, and knowledge explicitly required by the task may narrow menu fit. Tool capability, authority, runtime, language, and modality fields never filter or rank that menu; they remain post-selection execution proof. Hand-off edges must be acyclic: a review or feedback edge that points back to an earlier slot is rejected astask_force_cycle:<the loop path>— model review as a forward hand-off to the reviewer, not a back-edge (measured 2026-08-19: a researcher→research→quality-engineer order with areviewsback-edge was refused, and because edges live inside the WorkOrder the repair changedworkOrderDigestand forced the whole three-source federation to run again). Keep the defaultselectionPolicy.maximumCandidatesPerSlotat 30 unless a measured recall need justifies widening it (the schema allows up to 100), and never shrink it yourself to save tokens. The Hub and Cloud sources already shrink safely: they order each slot by fit with a decision model and, for a default-sized menu, send only their best 8. Measured 2026-09-24 on 40 live work orders: fit order put the right agent first 40/40, while in the unranked order it sat in the first 8 only 11/38 times — the menu bytes per slot fell 73%. Core still presents the merged menu incanonical_identity_no_rerankorder and Local candidates are not fit-ranked, so a smaller cap you set would cut rows arbitrarily, not worst-first. When a slot needs more rows, set the maximum above 30: the sources then return that many, still in fit order. In the returned menu,candidateOrdinalrestarts at 1 inside every slot — it is a per-slot position, not a running number across the menu. Keep private files, memory, secrets, direct identifiers, and raw local context on-host. Write every discovery-facing natural-language field (statement, role descriptions, required skills/knowledge) in English, faithfully translating a non-English request rather than passing its original wording through: the candidate corpus is English and cross-lingual matching silently buries the correct agent (measured: an identical query ranked its target 1st in English and 144th in Korean). Keep an untranslatable proper term alongside a short English gloss, e.g.종합소득세 (Korean comprehensive income tax). Thelanguagesslot is the delivery requirement, not the search language — set it to the language the work product must be produced in (e.g.ko) even though the order itself is written in English. - Call
workforce.search_candidatesonhephaestus-networkwith{workOrderRef, sourceScope: "network"}. Preserve every source receipt andselectionSessionId; the default projected menu is not a completefederationResultand must not be echoed as one. An unavailable source is explicit; it is not permission to pretend that source participated. 2b. For a multi-slot search, call it withshortlist: true. The response then carries summary cards (ordinal, name, entityKind, communities, one summary,callable,missingMandatory, andpublisherTriggerMatchwhen the publisher's own trigger sentences match this request) instead of full dossiers — measured 40,873B -> 10,087B for one 20-candidate slot. Narrow to the candidates worth a closer look, then callworkforce.expand_candidateswith{selectionSessionId, candidates:[{slotId, candidateOrdinal}]}and decide from those full cards, never from the summary alone. Keep the shortlist generous (six to eight per slot): the summary is for discarding the obviously wrong, not for picking the winner. - As the active host LLM, decide the staffing from the returned content and
qualification evidence, then call
workforce.validate_selectionwith{decision}:selectionSessionId,decisionAuthor(your real model id), and oneassignmentsrow per post naming the candidate by its per-slotcandidateOrdinalwithreasonCodes. Core loads the pinned menu and the pinned WorkOrder from that session, supplies the candidate-set digest and the arrays that are empty in a normal decision, and compiles the exactagentlas.workforce-selection.v1. Keep the accepted response'sfederatedSelectionDigest. Revise on rejection. Deterministic code may enforce governance but must not choose, rerank, or silently substitute the roster. An accepted result may still carryunmetRequirementCount— that is not a rejection, but readselectionValidation.unmetRequirementsand either accept the gap deliberately or reselect. Never report an accepted validation as if nothing were unmet. Use public codes such asreason:best-content-fit,reason:best-contract-fit, andreason:host-semantic-judgment, or exact codes from the chosen candidate's pinned evidence. Do not invent reason vocabulary; omitted compact-decision reasons usereason:host-semantic-judgment. - Call
workforce.prepare_executionwith{selection: {selectionSessionId}, federatedSelectionDigest, projectDir, goalId?, fullDossier: false}. Use both references from the same accepted validation response. Core restores only the exact digest-matching Selection from its pinned wrapper. The unchanged accepted wrapper underselection, the original exact Selection, or the same compactdecisionalso works. If a legacy normalized receipt cannot restore the original bytes, resend that decision or exact Selection. References never reauthor or substitute an accepted choice.projectDiris mandatory. Pass the incumbentgoalIdwhen continuing; otherwise Core joins this project's incumbent active automatic goal, and opens a new one only when there is none. Core must automatically bind a successful preparation before execution, so continuity cannot be skipped because no explicit goal mode was requested.fullDossier: falserequests the projected response (projection: "prepare.v2"):executionRosterrows carry identifiers and digests, and each worker'sdirectiveBundle/executionGraphis shipped once percontentDigestin top-levelbundleContents— resolve a row's content by itscontentDigestthere (a same-agent-two-slots roster would otherwise repeat the bundle byte-identically). The bound preparation stores the unprojected original. Omitting the flag returns legacy self-contained rows — the compatible default for machine verifiers that recomputebundleDigestover whole rows and update independently of the runtime. Require each worker to retain its exact source plus release, package hash, content digest, runtime-bundle digest, permission policy, and execution context pins. Recompute digests and fail closed on drift. - On later turns call
workforce.goal_contextfirst: reuse the incumbent roster plus local skills when sufficient; recruit only a real gap and pass the samegoalIdto preparation so new releases append. Recordreuse|local-only|recruit|standby|blockedwithworkforce.record_goal_turn. - Before every bound invocation, advertise the live host sessions and call
model.resolve_allocationwith that inventory plus the host-owned stage:planner/manager-plan,worker,manager-synthesis/synthesis, orverifier. Use the receipt's exact provider, model, and effort for that invocation. Model pins and ceilings come only from the MCP server's operator policy, never from the task or tool arguments. A missing worker policy inherits orchestrator; orchestrator never falls through to worker. Each advertised session carriessession_id,model,provider, and — when the host knows them —tier,supported_efforts, andcontext_window. Send what the host actually reports and never invent a field: an omitted context window is assumed at a conservative floor and the receipt says so (inventory_context_window_assumed), whereas a fabricated one would be read as measured. Operators set the orchestrator/worker policy withhep-orch orchestrator=<tier|model> worker=<tier|model>. - Run only the bound workers useful for this turn. For a selected team,
preserve its authoritative manager/worker graph. Run planner/manager,
workers, synthesis, and verifier as distinct invocations with explicit
artifact handoffs. Allocation receipts have
usage: nullbefore execution, so record actual usage on the later invocation/run receipt instead of inventing zero. If native child execution lacks the required enforcement, inspect the existing external-host transport before stopping at preparation:"$RUNNER" workforce execute --project <project> --goal-id <goalBinding.goalId> --adapter-argv-json '["/absolute/path/to/host-adapter"]'. This uses an explicitly selected available adapter, loads the original full bound preparation locally, orders distinct calls, snapshots artifact handoffs, and validates the resulting receipt. The adapter reads oneagentlas.workforce-host-executor-request.v1JSON request from stdin and writes one correlatedagentlas.workforce-host-executor-response.v1JSON response to stdout; its exact contract is in the installed engine'sagentlas_cloud/workforce/host_executor.py. The adapter owns real model execution and measured policy enforcement; this command grants nothing and installs nothing. Check native execution and existing adapters, then useagentlas_resolve_pluginsfor a missing adapter. An adapter created within already authorized work uses this contract outside Core. An actually empty tool menu is valid for model-only work without required tool bindings; residual CLI authority cannot be reported as zero tools. If no route can enforce the policy, retainpreparedand name that exact boundary. - Report
executedonly when the execution receipt proves every selected invocation, handoff, synthesis, and an independent passing verifier. Otherwise report the last truthful state:selected,prepared,source_unavailable,blocked, orfailed. Forpartialorfailed, report each source receipt's exactfailureCode: never collapse several receipts into one, substitute a different code, or relabel the outcome. Ifworkforce.prepare_executionrefuses with arepairobject, the agent is broken, not absent. Say in plain words whatrepair.meaningsays, offer the firstavailableaction inrepair.repair, and run itscommandonly after the user agrees. When a repair reportsreselectionRequired: true, select again; never swap in a different agent without a new selection. A refusal that carriesalternateslists other callable agents from the same search: present them as candidates. An agent built into the global agent folder is registered automatically at the next local search once it passescontract verify;hephaestus workforce local-autoregisterdoes it now and printscontractFailingandrefusedrows for any that are missing.
Recurring Hub use
Public Hub packages are free to discover and call. Recurring work still uses the host's model, API keys, permissions, and scheduled execution capacity. Explain those prerequisites and obtain the user's scheduling instruction. Do not quote or purchase a retired Hub lease. The exact release remains bound to the goal until explicit completion or cancellation.
Do not call legacy hephaestus_route, register or use direct remote search as a substitute
for Core federation, or use popularity/history/local availability as
semantic fit. Exact duplicate releases may collapse Local > Cloud > Hub only
when Core returns verified identical lineage; a name or slug match is not
enough. Name the actual workers in the result.
Raw request: $ARGUMENTS