Current status: v0.9 preview. This repository is ready for serious controlled live testing, not yet a v1.0 stability release.
Agent Mesh is a consolidated monorepo for bounded multi-agent workflows around runtime-adapter agent teams.
It merges the previously separate policy/plugin package, Discord gateway sidecar prototype, and runtime wrapper integration into one testable tree.
@openclaw-agent-mesh/core— runtime-agnostic policy, task-turn, run-scope, phase, and replay-control primitives.@openclaw-agent-mesh/openclaw-plugin— OpenClaw plugin adapter exposing planning tools backed by the core package.@openclaw-agent-mesh/gateway— controlled Discord gateway sidecar with policy decisions, proactivity, model selection, and host binding facades.@openclaw-agent-mesh/runtime-wrapper— OpenClaw runtime wrapper/plugin integration for sidecar rollout and dry-run safety.@openclaw-agent-mesh/tmux-bridge— agnostic tmux bridge for CLI-to-CLI agent intercommunication, surfaced in the gateway as thetmux-transportadapter (a peer of the Discord adapter). Seedocs/transports.md.
This repo is an OpenPack package for AgentWheel. openpack.json is the
canonical package manifest; the older agentwheel.json manifest is not used.
Install the shared runtime-neutral instructions and tmux skills:
npm i -g agentwheel
agentwheel registry update
agentwheel add nestdev-mesh --adapter codex
agentwheel install --dry-run
agentwheel installOpenClaw runtimes that need the optional mesh plugin should select it explicitly instead of installing it everywhere:
agentwheel install nestdev-mesh \
--adapter openclaw \
--select plugins/openclaw-agent-mesh \
--dry-run
# after reviewing the plan:
agentwheel install nestdev-mesh \
--adapter openclaw \
--select plugins/openclaw-agent-meshThe package provides the claude-tmux and codex-tmux skills from skills/.
With AgentWheel asset-includes, each installed skill also receives the bridge
scripts and agent configs it needs under its own bin/ and agents/
directories. The canonical script source remains packages/tmux-bridge/bin in
this repo; no generated script copies are committed.
Repository mode remains supported by setting AGENT_MESH_ROOT and using
$AGENT_MESH_ROOT/packages/tmux-bridge/bin directly.
- Core code must stay runtime-agnostic: no real Discord ids, host paths, bot names, secrets, or private deployment assumptions.
- Runtime integrations must load participants, targets, tasks, and allowlists from configuration.
- Bot and worker output is untrusted input.
- Real external side effects remain gated by host policy and explicit approval.
- Every bounded live run should carry an explicit run or orchestration id.
- Keep generated output, runtime state, session transcripts, and private raw logs out of the repo.
See:
docs/consolidation-inventory.mddocs/agent-beta-runtime-source-notes.md
npm ci
npm run verifynpm run verify runs build, tests, local Mesh harness smoke, and readiness privacy checks.
Useful Mesh commands:
npm run mesh:smoke
npm run mesh:decentralized
npm run mesh:readiness
npm run mesh:harness -- --participant agent-beta --state-file /tmp/agent-mesh-state.json --prettyPreferred live Discord shape:
<@bot>
ccm:v1 id=live-peer-20260603-a from=agent-beta turn=nestdev final=1 seen=runtime,agent-beta hop=3
body...
The legacy multi-line cc-mesh-* headers are still accepted for compatibility. In the compact form, turn is also treated as the single recipient unless an explicit to= field is present.
See docs/mesh-v0.9-live-testing.md before any live Discord test.