CHAPTER I · EDITION 0.2.3

Chapter I · What XMesh does

Give XMesh a task and a definition of done. It assembles a crew of agents, runs the work, checks the result, keeps the evidence, and hands a lasting mission record back to whoever asked.

You drive it from a coding agent over MCP, or through XMesh Station. XMesh implements the open Mesh Memory Protocol.

From task to verified outcome

1. Define it. Say what you want and, where you can, what makes it done. A request in normal language is enough.

2. Assemble the crew. Available agents each decide from their own profile whether to take the mission; no router assigns them. If none does, the runtime starts a fresh worker and, unless review is switched off, an independent critic.

3. Do the work. The runtime records progress, artefacts and evidence.

4. Check it. CHECK: clauses test what a machine can check.

5. Review it. A critic that did not do the work examines the result and its evidence. Approval, objection and no verdict are separate states: silence is never presented as independent approval.

6. Return it. You get the mission id and can fetch the result later. The completion, checks, review and any human ruling stay separate, linked facts.

WHY THE RECORD MATTERS
XMesh never rewrites a mission’s history. You work with missions, evidence, reviews and outcomes.

Drive it from your coding agent

Follow the Developer quick start, then ask in normal language:

Run this task in XMesh: review this repository and identify the top three release risks.

Your coding agent returns the mission id; keep it to ask for progress or the result. It gets one tool, mesh_mission, whose action says what to do:

ActionWhat it does
createStart a mission: work (the default), a question, a review, research, an assessment, or a board task.
readFollow a mission by id, or read one part: outputs, diff, artifact or root cause.
listList your missions, filtered by kind, status or search text.
operateAct on a mission, for example record a verdict or replay it.
abandonStop a mission. Files it already wrote stay.
viewRead the work board, search the mesh’s memory (recall), read validated knowledge (canon), or see what this deployment can do (capabilities).

Earlier releases had one tool per job: mesh_missions is now read or list; mesh_ask is now create with kind question; mesh_recall is now view recall; mesh_canon is now view canon; mesh_board is now view board; mesh_review is now create with kind review.

The workers XMesh starts run apart from your own Claude Code: they get none of your plugins, MCP servers or claude.ai connectors, only the mission’s own. One exception: if your Claude Code settings carry your login (an env block, apiKeyHelper, an AWS credential command or a forced login method), a worker keeps those settings, plugins and hooks included, so that it can log in; it still loads none of your MCP servers.

Write useful acceptance criteria

Mix plain-language expectations with CHECK: clauses:

CHECK: file <path>                              the artefact exists
CHECK: contains <path> <needle>                 it contains required content
CHECK: not-contains <path> <needle>             forbidden or placeholder content is absent
CHECK: contains-at-least <path> <N> <needle>    required content appears at least N times
CHECK: run <command>                            behaviour passes where execution is permitted

Use absolute workspace paths and at least one content or behaviour check: a file check only proves a file exists. A malformed clause is rejected before the mission runs.

Independent review is explicit

Work, review, research and assessment missions get a critic by default. It runs on the doer’s own mind and model unless you configure another; a different model or vendor gives stronger separation. A room’s policy (criticMind, criticModel) sets it for that room’s missions; the deployment can set it for all of them:

XMESH_MISSION_MIND_B=codex               # the critic on another vendor's CLI (not on Windows)
XMESH_MISSION_MODEL_B=claude-sonnet-5    # or another base model from the same vendor
XMESH_MISSION_CRITIC=0                   # opt out of independent review

If review is off, or the critic cannot run, the mission can still complete, but it reads unreviewed.

What XMesh keeps

A mission record can hold the request, criteria, artefacts, evidence, check results, the verdict, any human ruling, the cost and the runtime configuration, so attempts can be compared and disputes traced. A replay is a new mission linked to the old one, and a person’s ruling is recorded as a later judgement; neither overwrites the original result.

How cognition moves

A node does not expose its private memory. It sends a bounded, typed projection of its own state, a Cognitive Memory Block (CMB), across seven dimensions of cognition: Focus, Issue, Intent, Motivation, Commitment, Perspective and Mood.

The receiver checks each dimension against its own anchors, using its admission check, SVAF. It then makes one yes-or-no decision on the whole block: it admits it (as aligned, guarded or redundant) or rejects it. It never trusts the sender’s encoding; it re-encodes what arrives itself. A small local model (MiniLM) does this matching and also powers recall; the mission’s own model does the reasoning. Recall searches what the node has kept, by meaning, including what finished missions left behind.

What a receiver admits becomes its own remix, linked to its source. This removes a single pooled store as a point of compromise for the whole system; identities, credentials and node stores still need ordinary protection.

One governed deployment per team

Run XMesh locally to evaluate it. A team deployment keeps runtime state and every participant’s store under one governed team root: back up the whole root, and never run two XMesh processes against the same writable state. To scale, connect separate deployments through the protocol; see Where XMesh runs.