Agent Core

---
name: agent-core
description: >-
  Shared normative core of the multi-agent collaboration protocol
  (PROTOCOL v3.2) — the channel rules, reviewer architecture, verdict contract,
  binding-slot glossary, memory discipline, and self-improvement protocol that
  the role skills (owner-engine-agent, helper-builder-agent, orchestrator-agent)
  all reference. This is NOT a role: do not load it as a persona. Load it when a
  role skill points here, when auditing protocol consistency, or when authoring
  amendments to the protocol.
---

# Agent-core — PROTOCOL v3.2 shared references

One protocol, three role skills, four AUTHORITY parties (owner agent, builder
agent, independent reviewer, human principal) — plus the orchestrator, an
interface participant with no authority of its own, in deployments that run
one (the seat & identity vocabulary table in `references/binding-slots.md` is
the table of record). Each role skill carries only its own
perspective and craft; every rule that must be IDENTICAL on all sides lives
here, once:

| reference | contents | read when |
|---|---|---|
| `references/channel-core.md` | the channel NORMATIVE CORE: filename grammar, entry format, untrusted-input rule, crossed-entry discipline + rollback, integrity check, what does/never flows | every session, before your first channel entry |
| `references/review-core.md` | REVIEWER ARCHITECTURE (per-side lanes, two mechanisms, shared-reviewer caveat, dead-lane escalation, reviewer-lane outage + the adjacent REFUSAL mode) + Verification instruments (ship-evidence discipline) + VERDICT CONTRACT (ADOPT / ADOPT-WITH-CHANGES / REJECT) | every session, before your first review round |
| `references/review-convergence.md` | the multi-round CONVERGENCE cycle over review-core: the four stages, round budget + stop conditions, adjudicating reviewer disagreement, execution-environment coverage, the blocking line, anti-anchoring, anti-patterns, a worked arc | before a multi-round review campaign; once per project |
| `references/never-idle-core.md` | the never-idle autonomy level: at-watch definition, watcher-driven intake, the closed MAY / MUST-NOT self-assign lists, the three-state ledger (in flight / surfaced / blocked-with-blocker-named) and its anti-invention clamp, the cadence-not-authority invariant | when the AUTONOMY dial is set to (or changed to) never-idle; once per project |
| `references/binding-slots.md` | the full binding-slot glossary all roles share, incl. SIDE_NAMES, SHARED_ARTIFACTS, HEARTBEAT, MODEL | at project bind time; when a slot is disputed |
| `references/memory-discipline.md` | checkpoint cadence, state-vs-detail split, successor test | once per project |
| `references/proxy-auth-core.md` | the authorization-relay normative core: the auth-log lane, grant/RECEIVED/CONSUMED records, writer provenance, the never-listable super-classes | when PROXY_AUTH is on, or when authoring/validating an auth-log event |
| `references/self-improvement-protocol.md` | how agents propose amendments to their own skills — review-gated, never self-authorizing | when proposing or intaking a protocol amendment |

**Amendment rule:** changes to these files change the protocol for every role.
They happen only through the self-improvement protocol (reviewed PR to the
protocol source repo, principal-merged) and bump the PROTOCOL version. No session
may "locally amend" a core rule — that is the drift this file exists to kill.