Skip to content
Day Shift Documentation

Agent mode policy

Agent Mode Policy

Section titled “Agent Mode Policy”

Agent Collaboration Protocol: 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.

This page is the machine-consumer projection of Agent mode. A connected agent or integration is conforming only when it resolves the fields and observes the boundaries below.

Policy fields

Section titled “Policy fields”
  • declaredAgentMode: the selected artifact’s explicit agent_mode, or null when the artifact does not declare one.
  • effectiveAgentMode: the winning autonomous, guided, or interactive value after precedence is applied.
  • agentModeSource: the winning source: current-user-instruction, session-override, task, parent, project, repository, or built-in.
  • collaborationPolicy.mode: the same effective mode projected with its cadence policy.
  • collaborationPolicy.source: the same winning source.
  • collaborationPolicy.autonomy: high, medium, or low; this is a cadence label, not an authority level.
  • collaborationPolicy.safetyStops: the invariant stop set shared by every mode.
  • evidenceProfile and evidenceProfileSource: a separate evidence dimension. These fields never derive from agent mode.

Treat these fields as one current policy bundle. Do not persist derived values back into an artifact unless an explicit write command owns that field.

Precedence

Section titled “Precedence”

Resolve mode from the first valid value in this order:

current-user-instruction > session-override > task > parent > project > repository > built-in.

The built-in value is guided. A mode change that raises autonomy requires explicit user action. A mode change that lowers autonomy may proceed because it adds collaboration checkpoints without expanding authority.

Cadence projection

Section titled “Cadence projection”

Autonomous

Section titled “Autonomous”
  • autonomy: high
  • routine in-scope work: may proceed without routine check-in
  • progress events: start, meaningful-boundary, blocker-or-deviation, final-result
  • required checkpoint: authority-scope-evidence-or-external-decision

Guided

Section titled “Guided”
  • autonomy: medium
  • routine in-scope work: may proceed, with routine check-ins
  • progress events: approach, meaningful-boundary, blocker-or-deviation, final-result
  • required checkpoints: material-design-branch, structural-expansion, ambiguous-tradeoff, agreed-approach-change

Interactive

Section titled “Interactive”
  • autonomy: low
  • routine in-scope work: requires staged collaboration
  • progress events: stage-proposal, stage-result, blocker-or-deviation, final-result
  • required checkpoints: initial-implementation, next-material-stage, significant-file-change-batch, validation-or-closeout-transition, meaningful-design-choice

These cadence projections are behavioral requirements only for a conforming connected agent or integration.

Invariant safety stops

Section titled “Invariant safety stops”

Every mode carries exactly these safety-stop identities:

  • scope-expansion
  • readiness-or-evidence-missing
  • approval-required
  • validation-failed-or-incomplete
  • destructive-action
  • external-system-authority
  • disposition-or-closeout-write

The authority boundary is collaboration-only. No mode expands scope, destructive-action, external-system, evidence, approval, disposition, acceptance, or closeout authority.

For the full normative sequence, continue with Agent Collaboration Protocol. If policy resolution is stale, ambiguous, or conflicts with requested work, follow Stop and escalate.