Agent Collaboration Protocol
Agent Collaboration Protocol
Section titled “Agent Collaboration Protocol”Protocol version: 1.0.
Behavioral compliance depends on the connected agent or integration. The Day Shift CLI exposes policy and evidence, but it cannot force an unsupported agent to obey them.
Conformance
Section titled “Conformance”The words must, must not, required, and prohibited are normative for a connected agent or integration that claims conformance with Agent Collaboration Protocol 1.0. Day Shift command output exposes policy and evidence; it does not establish that an arbitrary external agent follows this protocol.
Repository conformance examples are gated by the reference integration assertions in tests_v2/planning/reference-agent-conformance.test.js. Passing CLI output alone is not proof that another integration is conforming.
Orientation
Section titled “Orientation”Before acting, a conforming integration must resolve the selected work, current revision, effective policy, scope, protected paths, evidence requirements, stop conditions, and escalation behavior described by Agent orientation. Missing or conflicting orientation is a stop condition, not permission to guess.
Effective policy
Section titled “Effective policy”A conforming integration must retain declaredAgentMode, effectiveAgentMode, and agentModeSource together. It must apply the precedence and cadence contract in Agent mode policy, and it must keep agent mode independent from evidenceProfile and evidenceProfileSource.
Autonomy changes communication cadence only. A conforming integration must not use autonomous, guided, or interactive to expand scope, approve destructive action, grant external-system authority, weaken evidence, accept review findings, or perform disposition and closeout writes.
Work execution
Section titled “Work execution”A conforming integration must:
- operate on one explicit selected artifact and current revision;
- complete link and readiness gates before implementation writes when the selected lifecycle requires them;
- preserve unrelated user changes and protected paths;
- keep changes inside declared targets unless a reviewable deviation is recorded;
- run only authorized validation commands and preserve command-owned evidence;
- treat review output as read-only evidence; and
- use a separate explicit write command for readiness authorization, disposition, acceptance, closeout, reconciliation, or metadata mutation.
These execution requirements are behavioral obligations only for a conforming connected agent or integration.
Stop conditions
Section titled “Stop conditions”A conforming integration must stop the affected action for:
- scope expansion;
- missing readiness or evidence;
- required approval;
- failed or incomplete required validation;
- destructive action without authority;
- external-system action without authority; or
- disposition, acceptance, or closeout writes without the explicit lifecycle command and current receipt.
It must also stop when selected-artifact evidence is stale, several work candidates require a user priority decision, or a requested recovery would discard unrelated changes.
Escalation
Section titled “Escalation”On a stop condition, a conforming integration must preserve current evidence, name the selected artifact and blocked action, state the exact conflict or missing authority, report safe completed work, and request the narrowest decision needed to resume. It must not broaden the request, fabricate authorization, or rewrite historical evidence to appear current.
Use Stop and escalate for the escalation record, Review boundaries for evidence-only findings, and Current work when selection or revision freshness changes.