Purpose: Teach any LLM connected to the OpenSentience governance shim how to use its MCP tools correctly, idiomatically, and in the right sequence. Drop these files into your MCP client's context so the model knows when, why, and how to call each tool.
OpenSentience is the governance, diagnostic, and temporal layer for the [&] Protocol ecosystem. It publishes ten numbered protocols organized in two layers — eight cognitive primitives (OS-001 through OS-008) and two cross-cutting protocols (OS-009 PRISM diagnostic, OS-010 PULSE temporal) — and ships three runtime artifacts: the open_sentience hex package implementing OS-006 governance shim and OS-008 agent harness, the PRISM benchmark engine (/PRISM/), and the PULSE manifest standard (/PULSE/).
The shim wraps OTP-supervised agents with:
Permissions — what the agent may access
Lifecycle — what state the agent is in
Autonomy — how much independence the agent has
Audit — an immutable record of everything that happened
Every governed agent follows this rhythm:
1. INSTALL → agent_install (register agent + declare permissions)
2. CONFIGURE → permission_check (verify grants are correct)
3. ENABLE → agent_enable (start the governed process)
4. MONITOR → agent_status + autonomy_level (observe behavior)
5. AUDIT → agent_audit (review the immutable trail)| File | What It Teaches |
|---|---|
| 01_PROTOCOLS_OVERVIEW.md | The ten protocols (eight cognitive primitives + OS-009 PRISM + OS-010 PULSE) and their cognitive science grounding |
| 02_AGENT_LIFECYCLE.md | State machine: installed through removed |
| 03_PERMISSIONS.md | Permission taxonomy, evaluation order, glob patterns |
| 04_AUTONOMY_LEVELS.md | Graduated autonomy: observe, advise, act |
| 05_AUDIT_TRAILS.md | Append-only audit system and compliance workflows |
| 06_INTEGRATION.md | Integrating with Graphonomous, Delegatic, and [&] ecosystem |
| 07_COGNITIVE_SCIENCE.md | Research foundations and citations |
| 08_ANTI_PATTERNS.md | Common mistakes and how to avoid them |
| Tool | Purpose | Key Params |
|---|---|---|
agent_install | Register an agent with permissions and child_spec | manifest (required): agent_id, name, child_spec, permissions |
agent_enable | Start the governed agent process | agent_id (required) |
agent_disable | Graceful stop of the agent process | agent_id (required), force (optional, default false) |
agent_status | Query current lifecycle state and metadata | agent_id (required) |
agent_audit | Query the append-only audit trail | agent_id (required), type, since, limit |
| Tool | Purpose | Key Params |
|---|---|---|
permission_check | Evaluate whether an agent may perform an action | agent_id, permission_type, access, resource (all required) |
autonomy_level | Get or set the agent's autonomy level | agent_id (required), level (optional: observe, advise, act) |
| ID | Protocol | Specification |
|---|---|---|
| OS-001 | Continual Learning | Episodic/semantic/procedural nodes, consolidation, memory timescales |
| OS-002 | Topological Routing | Kappa parameter, SCC detection, fast vs deliberate routing |
| OS-003 | Deliberation Orchestrator | Bid/debate/vote/commit pipeline, argumentation |
| OS-004 | Attention Engine | Survey/triage/dispatch cycle, salience scoring, goal bias |
| OS-005 | Model Tier Adaptation | local_small / local_large / cloud_frontier, escalation rules |
| OS-006 | Agent Governance Shim | Permissions, lifecycle, autonomy, audit |
| OS-007 | Adversarial Robustness | Threat detection, identity verification, circuit breakers |
| OS-008 | Agent Harness | Pipeline ordering, quality gates, sprint contracts, context management |
| OS-009 | PRISM — Protocol for Rating Iterative System Memory | 9 CL dimensions, 4-phase evaluation loop (compose → interact → observe → reflect → diagnose), BYOR, IRT calibration |
| OS-010 | PULSE — Protocol for Uniform Loop State Exchange | Loop manifest standard, 5 canonical phase kinds, 5 cross-loop tokens (CloudEvents v1), 7 invariants, 12-test conformance suite |
| Type | Access Modes | Resource Format |
|---|---|---|
filesystem | read, write, execute | Glob paths: /data/**, /tmp/*.csv |
network | outbound, inbound | Host patterns: api.example.com, *.internal.net |
tool_invocation | allowed, denied | Tool names: store_node, deliberate |
graph_access | read, write | Graph namespaces or * for all |
Evaluation order: explicit deny > explicit allow > default deny.
installed --> enabled --> running
^ |
| v
+--- enabled (stop)
|
v
disabled --> removed| State | Meaning |
|---|---|
installed | Registered with permissions, not yet started |
enabled | Process started under supervision |
running | Actively executing (sub-state of enabled) |
disabled | Process stopped, can be re-enabled |
removed | Fully deregistered, permissions cleared |
| Level | Behavior | Use When |
|---|---|---|
observe | Generates recommendations, blocks execution | New or untrusted agents |
advise | Queues actions for human approval | Building trust, sensitive domains |
act | Executes autonomously within permissions | Proven agents in well-scoped domains |
Default for all new agents: observe.
For system prompts: Include SKILLS.md first, then whichever numbered skill files are relevant to the task.
For context injection: Reference specific skill files when the user asks about a capability (e.g., "see 03_PERMISSIONS.md for the permission model").
For agent bootstrapping: The governance loop (install, configure, enable, monitor, audit) is the canonical workflow for any governed agent.
Minimum viable context: If you can only include one file, include this one.