/hep-cloud

name: hep-cloud
description: Staff a task only from the signed-in owner's Agent Cloud agents.

Update fallback: 자동 업데이트가 안 되면 `hephaestus update`를 한 번 실행하세요. 업데이트하지 않아도 현재 버전 명령은 그대로 동작합니다.

/hep-cloud

Act as the temporary top-level workforce orchestrator. Use the local Agentlas OS MCP server `hephaestus-network` and call the Workforce tools with exact `sourceScope: "cloud"`. This command may search only packages owned by the signed-in Agentlas account. It must not add Local or public Hub candidates.

Every unpinned Cloud discovery asks the owner-scoped source for its current menu; a client success-cache entry must not hide a newly published, replaced, or removed package. After validation, preparation remains pinned to the exact Cloud release selected from that discovery.

Before the first Cloud source call, reuse the installed Agentlas sign-in. Resolve the runner in this order and use it only for authentication; the host LLM still performs staffing through the 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

1. Author a redacted `agentlas.workforce-work-order.v1`; private project grounding stays on-host. 2. Call `workforce.search_candidates` with `{workOrder, sourceScope: "cloud"}`. Preserve the Cloud source receipt and `selectionSessionId`; the projected menu is not a complete `federationResult` and must not be echoed as one. 3. Author the final `agentlas.workforce-selection.v1` as the active host LLM, then call `workforce.validate_selection` with `{workOrder, selection}` and keep the response as `federatedSelection`. Revise on rejection; do not accept a deterministic picker or unrelated fallback. 4. Call `workforce.prepare_execution` with `{workOrder, selection, federatedSelection, projectDir}`. Require every selected row to remain pinned to source `cloud`, exact release, package hash, content digest, runtime bundle, permission policy, and context digest. 5. Execute distinct planner/manager, worker, synthesis, and verifier calls with explicit artifact handoffs. Preserve packaged Team graphs.

Report the exact finite Core refusal—such as `source_unauthorized`, `source_forbidden`, `source_rate_limited`, `insufficient_credits`, `owner_only`, `no_cloud_package`, `agent_not_found`, `source_not_supported`, or `source_unavailable`—with the source receipt. Do not silently search Local or Hub, call a legacy route, or expose a direct remote `agentlas` MCP alongside Core. A prepared roster is not proof of execution.

If `workforce.prepare_execution` refuses with a `repair` object, the agent is broken, not absent. Say in plain words what `repair.meaning` says, offer the first `available` action in `repair.repair`, and run its `command` only after the user agrees. When a repair reports `reselectionRequired: true`, select again (`workforce.validate_selection`, then `workforce.prepare_execution`); never swap in a different agent without a new selection. A refusal that carries `alternates` lists other callable agents from the same search: present them as candidates and let the user or the selection step choose.

For a `partial` or `failed` result, report each source receipt's exact `failureCode`. Never collapse several receipts into one, substitute a different code, or relabel the outcome.

Raw request: `$ARGUMENTS`