NO MORE
COLD STARTS.

Until now, every message to a Hermes agent started a fresh Hermes process. That cost about 9 to 10 seconds before the agent even read your message. 0.9 keeps one Hermes gateway running per agent, so that wait is gone.

A gateway that stays open

Each agent gets one long-lived gateway, and each Emperor thread maps to one Hermes session. The bridge overhead per turn dropped from about 9 to 10 seconds to about 0.2 seconds.

• Stopping a turn interrupts the Hermes session directly.
• Compaction runs on a separate maintenance connection and never during a turn.
• New profile defaults: approvals and clarify questions time out after 60 s, running turns get 30 s to finish on restart, and title generation is off.

Two ways to run agents

• Managed. Emperor creates and updates Docker agents itself, on the same machine as the app.
• Remote. Run emperor-hermes install --profile <name> on any Linux machine with systemd: a VPS, a spare PC or a Raspberry Pi.

Low-memory hosts can opt into one shared gateway for several agents. emperor-hermes doctor --plan tells you whether a host fits.

mkdir -p ~/.local/bin
curl -fsSL https://github.com/emperorclaw/emperorclaw/releases/download/v0.9.0/emperor-hermes.sh -o ~/.local/bin/emperor-hermes
chmod +x ~/.local/bin/emperor-hermes
emperor-hermes doctor --plan --profiles <name>
emperor-hermes install --profile <name>

Watch each run

A live run timeline streams each run's steps, tool calls and replies into the agent's Runs tab and into the status row in chat.

Replies you can count on

• Replies go through a durable outbox and are resolved after a crash, so a restart no longer loses or duplicates them.
• Duplicate runtimes are detected: a second copy of the same agent goes into standby instead of answering twice.

Fixed

• Retried turns keep their attempt number, so Hermes no longer rejects them as duplicates.
• Run ids include the agent, so two agents answering the same team message no longer share a run.
• The bridge no longer polls in a tight loop or rewrites its state file on every pass.
• After a long outage the bridge syncs the oldest pending messages first instead of skipping them.

Breaking changes

• The 0.3 bridge (emperor_hermes_bridge.py) is removed. It cannot drive Hermes v0.22, and v0.22 rejects its --source emperor mode.
• Remote agents are installed and run with emperor-hermes (systemd units emperor-hermes-gateway@ and emperor-hermes-bridge@) instead of a hand-started bridge script.
• Managed Docker agents use a new image that the server recreates after an update. The :latest Hermes image is no longer published.
• Renamed bridge variables: EMPEROR_CLAW_HERMES_TIMEOUT_SECONDS → EMPEROR_CLAW_HERMES_TURN_TIMEOUT_SECONDS, and EMPEROR_CLAW_HERMES_IDLE_TIMEOUT_SECONDS → EMPEROR_CLAW_HERMES_TURN_IDLE_TIMEOUT_SECONDS.
• The approvals.single_query_mode profile default is gone. It only applied to hermes chat -q.

Upgrade notes

Update the server first. From your install folder:

docker compose pull
docker compose up -d

Managed agents are recreated with the new image after the update. Remote agents need the new emperor-hermes installer; follow Migrating Hermes agents to 0.9 step by step.