local_agent_launch_plan_v1

local_agent_launch_plan_v1 is a public-safe dry-run contract for previewing how LoopX could assign local peer agents to current todos before any host starts a worker, daemon, server, or external process.

It answers one narrow question: "Given registered peers, quota, todo claims, task policy, and current gates, what launch preview should the operator see?" It is not a launcher, scheduler, task lease store, or permission grant.

Boundary

The source of truth remains:

registry coordination.agent_model=peer_v1 and registered_agents;
quota should-run and its interaction_contract;
todo projection, claims/leases, capabilities, gates, and run history;
task/repository workspace and review policy.

The plan is mode=dry_run. It must not start a process, allocate a shell, open a daemon connection, call a remote agent service, claim a todo, or write LoopX state.

Shape

{
  "schema_version": "local_agent_launch_plan_v1",
  "mode": "dry_run",
  "goal_id": "loopx-meta",
  "agent_model": "peer_v1",
  "generated_at": "2026-07-10T04:00:00Z",
  "configured_agents": [],
  "task_assignments": [],
  "launch_preview": [],
  "status_projection": {},
  "evidence_projection": {},
  "future_gates": [],
  "truth_contract": {
    "source_of_truth": [
      "registry",
      "quota_should_run",
      "todo_projection",
      "run_history"
    ],
    "plan_is_authoritative": false,
    "plan_is_executable": false,
    "write_api": false,
    "launch_command_allowed": false,
    "recompute_rule": "Recompute before each preview."
  }
}

There is no goal-level leader id. A host may show which peer is selected for a task bundle, but that coordinator is deterministic and temporary; it gains no durable authority over other identities.

Configured Agents

configured_agents[] is the discovered peer list after identity and capability checks. Each item includes:

agent_id;
agent_model=peer_v1;
optional advisory profile_role and scope_summary;
source, normally registry.coordination.registered_agents;
can_receive_work and compact blocked_by reasons.

Do not copy raw automation prompts, private chat history, local paths, or connector payloads into the preview.

Task Assignments

task_assignments[] maps current todos to peers for display. Required fields:

agent_id and todo_id;
assignment_kind: claimed, unclaimed_candidate, monitor_only, or blocked;
responsibility: one compact public-safe sentence;
claim_policy: how the todo must be claimed or transferred;
optional action_kind, excluded_agents, and unblocks_todo_id for task semantics, executor separation, and dependency lineage.

Assignments are advisory preview rows. A real claim or transfer still uses the todo lifecycle. A profile cannot supply an implicit reviewer.

Launch Preview

Each preview describes what the host would show before launch and remains non-executable:

{
  "preview_id": "preview_peer_delivery",
  "agent_id": "codex-product",
  "todo_id": "todo_public_slice",
  "next_step_label": "Build the public dry-run fixture slice.",
  "workspace_policy": "selected_task_requires_isolation",
  "host_execution": {
    "will_start_process": false,
    "tool_call_allowed": false,
    "shell_command": null,
    "daemon_required": false,
    "external_service_call": false
  }
}

The preview may name a todo and task workspace policy, but it must not include a runnable command, process id, credential, auth token, private path, or remote URL. Real launch support requires a separate host execution contract.

Status And Evidence Projection

status_projection includes waiting_on, next_action, user_action_required, agent_can_continue, first_agent_todo, gate_state, quota_state, and launch_state.

evidence_projection carries compact source and validation refs plus explicit false markers for raw logs, transcripts, credentials, and private paths. Evidence refs are join keys, not embedded payloads.

Future Gates

These remain future-gated:

server_daemon_launch;
external_agent_execution;
credentialed_host_actions;
state_write_from_preview.

Each gate names future_gated or blocked_without_authority and the required contract before it can be enabled.

Acceptance Checks

A fixture or implementation is acceptable when:

schema_version=local_agent_launch_plan_v1, agent_model=peer_v1, and mode=dry_run;
configured agents are unique peers with no leader/parent role;
every task assignment references a configured peer and existing todo;
review remains action_kind=review; any executor exclusion is explicit;
every preview keeps all host execution booleans false and command null;
status and evidence projections remain compact and public-safe;
all real launch, credential, and preview-write capabilities remain gated.

FilesExpand file tree

local-agent-launch-plan-v1.md

Latest commit

History