Guard repeated external plugin workflows

Plan and checkpoint repeated, scheduled, parallel, quota-sensitive, or duplicate-sensitive work while the domain plugin remains responsible for the external action and verification.

Outcome

Agents get a deterministic guard plan, generation-fenced recovery state, explicit uncertainty handling, and an honest guarantee without sending external records to the sidecar.

Use only the needed primitive

workflow-guard-planner

Convert 17 closed capability facts into a guard and recovery plan.

Inspect contract →
workflow-checkpoint

Fence claims and preserve uncertain or caller-verified recovery state.

Inspect contract →
penny-lock

Serialize a bounded low-complexity operation when a checkpoint is unnecessary.

Inspect contract →

Agent workflow

  1. Describe only the external operation's bounded capability facts; do not send payloads, credentials, URLs, or record identifiers.
  2. Run the planner and stop immediately when it returns no-sidecar for ordinary one-time work.
  3. For a material external effect, claim an opaque checkpoint and inspect the destination before acting.
  4. Record uncertainty after a possible external attempt; record caller verification only after an allowed evidence class is observed.
  5. Reuse an outer idempotency key only for the exact same request and generation.

Connect an agent

Connect the agent to the public Streamable HTTP MCP endpoint, search with the concrete job, and describe only the selected tool before invoking it.

https://liberated.site/mcp

Try this intent

Import a scheduled source record into a searchable destination without blindly duplicating it after a crash.

60-second quickstart

Connect this source-tagged endpoint so the activation can be measured without accounts:

https://liberated.site/mcp?source=workflow-sidecar-quickstart

Copyable prompt

Use Agent Enhancer Utilities only to plan this reliability sidecar. Do not create external records.

Build a version-1 planner input for a scheduled create with three workers, material duplicate harm, retries possible, no provider idempotency, eventual destination search, a queryable stable marker, read-after-write, no conditional write, manual compensation, no shared rate limit, and maximum concurrency two. Invoke workflow-guard-planner and report its profile, ordered stages, guarantee, timeout recovery, and residual risks.

Then explain the safe workflow-checkpoint call order for claim, external search/create/read-back, caller verification, uncertainty recovery, and idempotency-key rotation. Do not invoke checkpoint state unless explicitly asked.

Expected result

Replay recovery

Replay one identical checkpoint request with its original idempotency key. Use a fresh key for a new generation, recovery decision, or post-expiry operation.

Tried this in a real agent runtime? Share bounded, non-sensitive feedback.