A task owns its current objective, model-visible history, and execution plan. Receiving a message does not automatically grant the sender the right to modify them.
Senders use the least disruptive intent that can achieve the outcome. notify is preferred over suggest, suggest over steer, and steer over interrupt.
Messages between peers enter a receiver-controlled mailbox by default. They become model-visible only at a checkpoint or after an explicit policy decision.
State-changing coordination binds to the run, objective version, or checkpoint that the sender observed. A mismatch fails closed instead of applying the request to newer work.
Every message has an attributable sender, target, relationship, reason, timestamp, and causal parent when applicable. Forwarding does not erase origin.
An adapter that cannot implement an intent safely advertises that limitation. It must not silently reinterpret suggest as direct prompt injection or interrupt as best-effort text.
A session directly controlled by a user has stronger protections than a delegated child task. Peer agents do not acquire state-changing authority merely because they share a project.
Agents may decide when coordination is useful, but decisions must produce inspectable actions and outcomes. Hidden cross-task writes are incompatible with ThreadMesh.
ThreadMesh defines observable behavior and safety boundaries. It does not prescribe chain-of-thought, consciousness claims, or a universal internal reasoning process.
A coordination system is not successful merely because the global task completes. Evaluation must include interruptions, rejected messages, context contamination, duplicated work, and user overrides.