Agentlas Cloud Runtime Contract

June 26, 2026 ยท View on GitHub

Hephaestus now emits and repairs agentlas.json before an agent package is allowed into Agentlas Cloud call flows.

V1 Scope

  • Generate or repair agentlas.json.
  • Run local risk screening before private sync or public publish.
  • Compile a runtime bundle instead of sending the whole ZIP to an LLM.
  • Gate agentlas.read_agent_file through allowRead and denyRead.
  • Keep private sync and public clean-copy behavior separate.
  • Attach a workspace-scoped personalization overlay only after authentication. The overlay may include promoted memory summaries, promoted playbook cards, plugin locks, and retrieval receipt ids; it must not include raw prompts, transcripts, secrets, credential values, or private local files.

CLI

bin/hephaestus wizard ./some-agent --name instagram-operator
bin/hephaestus security scan ./some-agent --strict
bin/hephaestus runtime bundle ./some-agent
bin/hephaestus runtime read-agent-file ./some-agent AGENTS.md
bin/hephaestus field-test

Lazy File Read

runtime bundle sends the manifest, bounded file index, package hash, and risk summary first. Full file contents are fetched only through runtime read-agent-file, and only when agentlas.json allows the requested path.

The security scan is risk screening, not a safety guarantee. It reports file paths, risk types, and redaction status without printing secret values.

Workspace Personalization

Agentlas Web treats a called Cloud/Hub agent as:

effective_agent = immutable_base_bundle + workspace_personal_overlay

The durable identity is agentBindingId = workspaceId + sourceScope + baseAgentId. packageHash is stored as compatibility metadata, not as the binding id, so a workspace can keep one personal agent identity while filtering hints by package version.

The runtime bundle may include:

  • personalization.binding: binding id, base agent id, source scope, slug, and current package hash.
  • personalization.memoryItems: promoted summaries only.
  • personalization.playbookCards: promoted recipes only.
  • personalization.pluginLocks: plugin slug/version/permission snapshots.
  • personalization.retrievalReceiptId: audit id for what was included.

Candidate writes use separate MCP tools:

  • agentlas.record_agent_memory
  • agentlas.record_agent_playbook
  • agentlas.propose_agent_evolution

These tools save candidate records for curator or explicit user promotion. They do not mutate the public Hub package and do not self-apply rule or skill changes.

Non-Goals

  • Cloud server-side model execution.
  • User key or OAuth token storage.
  • Vault snapshot upload.
  • Web-based agent editing.
  • Enterprise policy implementation.