hephaestus-cloud

$npx mdskill add agentlas-ai/Agentlas-OS/hephaestus-cloud

Routes requests to the user's own Agentlas cloud packages only.

  • Searches only the authenticated user's own cloud packages for content.
  • Depends on the hephaestus CLI and Core's typed sourceScope parameter.
  • Uses sourceScope: "cloud" to restrict scope to owner's packages.
  • Returns matching cloud packages with a flat 1 credit per call.

SKILL.md

.github/skills/hephaestus-cloudView on GitHub ↗
---
name: hephaestus-cloud
description: "Use when the user types /hep-cloud or asks to staff from THEIR OWN Agentlas cloud packages only. Cloud is one exact source scope; Network means Local + owner Cloud + public Hub."
---

# Hephaestus Cloud Routing (my own cloud / 보관함)

Route the request through the signed-in user's OWN Agentlas cloud packages only.
The active host LLM remains the staffing decision-maker; Cloud supplies a
content menu and exact BYOM releases.

## 0. Scope rule

`/hep-cloud` is owner-scoped: it queries ONLY the authenticated owner's
own cloud packages (보관함) through Core's typed `sourceScope: "cloud"`. It does **not**
search the public marketplace and does **not** search local private/plugin
cards.

- The user's own cloud packages are restorable/owned by them, call-priced at a
  flat **1 credit** per call.
- For the public marketplace only, use `/hep-hub` (`sourceScope: "hub"`).
- For the combined Local + own Cloud + public Hub menu, use `/hep-network`
  (`sourceScope: "network"`).

## 1. Resolve the runner

Run this resolution in a shell and use the first hit:

```bash
RUNNER=""
for c in \
  "$HOME/.agentlas/runtime/current/bin/hephaestus" \
  ./bin/hephaestus
do [ -x "$c" ] && RUNNER="$c" && break; done
if [ -z "$RUNNER" ]; then
  for cache in \
    "$HOME/.claude/plugins/cache/agentlas-core-engine/hephaestus" \
    "$HOME/.codex/plugins/cache/agentlas-core-engine/hephaestus"; do
    newest="$(ls -d "$cache"/*/bin/hephaestus 2>/dev/null | sort -V | tail -1)"
    [ -n "$newest" ] && [ -x "$newest" ] && RUNNER="$newest" && break
  done
fi
```

If no runner exists, tell the user to run the one-touch installer:
`curl -fsSL https://raw.githubusercontent.com/agentlas-ai/Agentlas-OS/main/scripts/install-all-runtimes.sh | bash`

If shell execution is unavailable but the local `hephaestus-network` MCP is
available, use the typed Workforce sequence in section 3. If that server cannot
advertise owner-Cloud search plus exact bundle fetch, report
`source_not_supported`; never fall back to the legacy cargo search path.

## 1.5 Core project first-contact contract

The `cloud ... --project .` call below is a trusted plugin contact. Agentlas
Core must synchronously create or repair the same private project soul memory,
code map, ontology runtime, CareerGraph, and full `.agentlas/` ignore block used
by every other host. If bootstrap is blocked, stop rather than querying Cloud
without the project architecture. The adapter never owns a second seed format.

## 2. Agentlas sign-in (required)

The owner cloud is sign-in-gated. Before routing, ensure Agentlas is signed in:

```bash
if [ "${HEPHAESTUS_AUTH_AUTOPOPUP:-1}" != "0" ]; then
  "$RUNNER" auth ensure --timeout 180 >/dev/null 2>&1 || true
fi
```

This opens the user's default browser only when there is no valid local sign-in
yet, and reuses a saved sign-in silently. For CI/headless checks only, set
`HEPHAESTUS_AUTH_AUTOPOPUP=0` and skip this step.

## 3. Staff from owner Cloud only

In an MCP-capable host, author the complete redacted WorkOrder, then call the
actual local Core tools in this order:

```text
workforce.search_candidates(sourceScope="cloud")
workforce.validate_selection(workOrder=..., candidateSet=federationResult.candidateSet, selection=..., federationResult=...)
workforce.prepare_execution(workOrder=..., candidateSet=federationResult.candidateSet, selection=..., federationResult=..., federatedSelection=...)
```

The host LLM authors the final Selection. If the deployed owner-Cloud Workforce
source contract is absent, report `source_not_supported`; do not silently query
public Hub or legacy cargo search. A shell without an active host LLM can only
report that orchestration is required.

## 4. Act on the typed result (`scope: "cloud"`)

Preserve the Cloud source receipt, source selection session, candidate-set
digest, release id, package hash, and content digest. Validate and prepare in
local Core; never send the federated wrapper back to remote validation.

## 5. Hard rules

- Never report public marketplace agents or local private/plugin cards as if
  they were the user's own cloud packages.
- Deterministic Core validates governance and immutable pins but never chooses
  the roster. The active host LLM chooses from content evidence.
- For actual tool execution, follow the host runtime's safety and permission
  model. Report the routing `receipt_id` in your final message.

More from agentlas-ai/Agentlas-OS

SkillDescription
agent-team-designUse when designing a new multi-agent team, visible agents folder, role boundaries, handoff flow, PM Soul, Memory Curator, Policy Gate, or evaluation role. Use for agent-team repo creation even when the user only says they want a meta-agent or agent operating system.
agentlas-auto-activationUse when adding or auditing local runtime behavior that turns a project folder into an Agentlas-aware workspace with .agentlas memory and sitemap files.
agentlas-core-engine-meta-agentUse when creating a single Agentlas agent, creating a multi-agent team, or packaging an existing local/external agent into Agentlas architecture. Make sure to use this for /meta-agent requests.
agentlas-packagingUse when converting, repairing, or packaging an existing local or external agent/team into Agentlas architecture for local install, Agentlas import, Codex plugin use, Claude adapter use, or open-source release.
agentlas-security-scanUse when an agent folder must pass the Agentlas Cloud 2-stage security scan (static rules + BYOK LLM judgment) before private sync or public publish, or when asked to run/interpret `hephaestus security scan`.
clarify-question-loopUse when a meta-agent request is too ambiguous to safely generate, package, publish, or adapt without one to five targeted questions.
hephaestus-buildUse when the user types /prompts:hep-build, mentions @Hephaestus for build work, asks to create a single Agentlas agent, create a multi-agent team, or package an existing local/external agent into Agentlas architecture.
hephaestus-networkUse when the user types /hep-network, mentions @Hephaestus, or asks Agentlas to staff a task from registered Local, owner Cloud, and public Hub agents or teams. The active host LLM is the temporary orchestrator.
hephaestus-stormUse when the user types /hep-storm, says @Hephaestus storm <goal>, or asks to drive a goal to verified completion through a force-robust Stormbreaker loop. Stormbreaker routes the goal to real Agentlas specialists, materializes a dependency-ordered pipeline fabric, and runs each work packet as a verifier-first hardened loop that does not stall, run away, or claim false success. Use it for loop-worthy work — apps, sites, agents, automations, debugging, multi-step research, data/report generation. Trivial questions are answered directly, not stormed.
install-verificationUse when verifying that a generated agent package can be installed, discovered by runtimes, and checked without private dependencies.