Hermes adapter

Hermes is an autonomous agent platform that loads skills via the AgentSkills standard and maintains episodic memory across sessions. DWP integrates with both surfaces: the skill pack loads from .agents/skills/ the same way as any other AgentSkills-compatible pack, and state.json’s outcome records feed directly into Hermes’s memory index.

Support level

Full (Documented) — Hermes reads AGENTS.md and loads the skill via AgentSkills-standard loading. All DWP sub-skills are available.

Installation

Install the skill pack into the workspace’s .agents/skills/ directory, or use the platform’s skill manager if available.

Invocation

Plain text:

create a plan for <goal>
resume the open plan

Episodic memory synergy

Hermes indexes agent memory to make past work retrievable across future sessions. In v6 and v7 plans, DWP’s deterministic verification receipts (outcomes.py receipt) and outcome records feed directly into Hermes’s memory index: each completed task carries what was tried, failed, and worked in a structured JSON schema. Completed plans become retrievable episodic memory — an agent can recall how a problem was solved.

Ephemeral sessions append events to journal.ndjson and project state.json via ledger.py project rather than manually rewriting state files. See Plan state.

Cron-driven continuation

On platforms with cron scheduling, continuation follows the same pattern as other autonomous platforms: each scheduled turn wakes, runs the DWP Resume Protocol, executes at most one task, updates state.json, and yields. The plan, not the session, is the unit of continuity.

The machine-readable state layer (manifest.json + state.json) is REQUIRED for unattended execution. See Agent protocol for the full stop-condition and escalation protocol.

Notes

Hermes and OpenClaw are both autonomous agent platforms in the supported-agents table in Agent protocol. Both run under the unattended execution profile and both benefit from the agent workspace archetype’s state.json-as-git-log equivalence.