All modes and statuses

“Mode” is often used conversationally for several different things. Engine keeps them separate: goal behavior, cognitive decisions, invocation duration, execution status, control layer, risk, privacy, policy, evidence, learning, routines, and a plugin kill switch are distinct axes.

Status: action lifecycle values remain v2 contracts; generic autonomy uses engine.plugin/v3. Values below are current unless marked legacy or roadmap.

1. Goal mode: GoalModeV2

Value	Meaning	End behavior
achieve	Reach the desired effect once	Observed true -> completed
maintain	Keep the desired effect true over time	Observed true -> monitoring; drift -> active again

Both exist now. maintain is not a realtime mode: detection latency depends on events, polling, and provider freshness.

2. Persistent goal status

In the current implementation, this is a set of stored strings, not a separate public SDK enum.

Status	Meaning
active	The goal needs another pass
completed	An ACHIEVE goal has independently been established as true
monitoring	A MAINTAIN goal is true and remains under observation
waiting	Waiting for change, approval, a task poll, or another resume condition
uncertain	Required evidence or the oracle is UNKNOWN; do not treat it as false or success
degraded	An isolated goal/provider/brain route failed or exhausted its budget/circuit
abandoned	A stop condition or explicit abandonment ended the goal

A COMPLETE brain decision does not directly set completed; the Heart does so only after condition/oracle evaluation.

3. Cognitive decision: DecisionKindV2

Value	Intent	Authority/effect
query_world	More or different observation is needed	No mutation; the current Heart waits/re-observes
consult_specialist	Invoke a named specialist	Specialist advice remains a proposal
propose_effect	Return a semantic ProposedActionV1	Still no execution rights
wait	Do nothing now and wait for relevant change	Durable waiting
complete	The brain thinks the work is complete	Advisory only; oracle/conditions decide
abandon	Advise leaving the goal	Heart may record status abandoned

The v1 terms CONSULT_BRAIN and USE_TOOL are legacy. V2 deliberately has no physical USE_TOOL fast path: proposals always pass through controller, policy, authorization, executor, and oracle.

4. Invocation mode: InvocationModeV2

Value	Semantics	Implementation status
immediate	Dispatch returns a terminal receipt or a result that can be reconciled immediately	Implemented and Fake/simulation-tested; used by Homey
task	Dispatch returns an external handle; Heart polls, cancels at deadline, and reconstructs after restart	Implemented and Fake/simulation-tested in the reference warehouse
stream	Long-running stream with cursor/reconnect semantics	Contract/store scaffolding; no end-to-end reference proof yet

Invocation mode describes the duration of a capability, not goal duration, risk, or autonomy.

5. Execution status: ExecutionStateV2

Status	Terminal?	Meaning
requested	No	The request has been created/recorded
accepted	No	An external executor accepted it; a handle is required for task recovery
running	No	Execution is in progress; a handle is required
succeeded	Yes	Executor reports successful execution; the effect still requires independent verification
partial	Yes	Only part was executed or achieved
failed	Yes	Executor reports failure
cancelled	Yes	Task was cancelled
unknown	Yes	The lifecycle cannot reliably determine execution/acknowledgement

succeeded therefore means “the execution receipt succeeded,” not automatically “the goal effect was achieved.” The latter appears in EffectDeltaV1 and fresh goal evaluation.

6. Control layer: ControlLayer

Value	Use
query	Observe/query; an opaque capability must be query-only
semantic	High-level domain effect translated by a controller into target parameters
actuator	Capability is closer to an actuator but remains outside hard-realtime control

A control layer grants no permission and does not determine the risk class by itself.

7. Risk class: RiskClass

Value	Interpretation
read_only	No intended world mutation
low	Bounded low-risk action within enrolled limits
medium	Higher impact; policy/mandate must carry it explicitly
high	Current policy requires approval unless the mandate explicitly allows it

This is a software taxonomy, not certification. A misclassified capability does not become physically safe because its manifest says low.

8. Privacy class: PrivacyClass

Value	Meaning
public	Public information
local	Local operational data
sensitive	Explicit privacy permission required
camera	Camera/image data; explicit permission required

A remote model does not automatically receive all observations. Context projection and plugin needs remain separate boundaries.

9. Policy outcome: PolicyOutcome

Value	Consequence
ALLOW	Policy may create an exactly bound authorization
DENY	Final refusal for this request/state binding
REQUIRE_APPROVAL	An external authorized approval is required; a brain cannot provide it
DEFER	Not yet decidable/active, for example because a mandate is not yet valid

Only ALLOW reaches dispatch.

10. Evidence grade: EvidenceGrade

Value	Meaning
OBSERVED	Emitted directly by an identified sensor/tool/executor within its coverage
DERIVED	Deterministically derived from named observations
INFERRED	Model or statistical conclusion
UNKNOWN	Insufficient evidence
CONFLICTING	Sources disagree
STALE	Evidence is too old for this decision

Confidence and provenance remain separate from the grade. High confidence does not turn INFERRED into OBSERVED.

11. Preference promotion: PreferencePromotionMode

Value	Meaning
explicit_only	Only an explicit owner correction may change the preference
shadow_low_risk	Inferred evidence may enter a candidate/shadow route after fixed gates

This changes preference state, not model weights or authority scope.

12. Preference learning status: LearningStatus

Status	Meaning
candidate	A candidate has been identified
shadow	Evaluation period without direct authority expansion
promoted	A new versioned GoalSpec preference is active
rejected	Gates or outcomes did not pass
rolled_back	The exact old value was restored in a new version

In practice, the current generic route starts at shadow once the evidence gates pass; candidate remains a public contract status.

13. Routine candidate status: RoutineCandidateStatus

Status	Meaning
candidate	Pattern candidate
shadow	Counterfactual test; dispatch count must remain zero
ready_for_approval	Real shadow opportunities and the agreement gate passed; owner approval is required
promoted	Routine, goal, and mandate were activated atomically
rejected	Pattern/conflict/shadow failed
rolled_back	Promotion was exactly reversed

14. Active routine status: RoutineStatus

Status	Meaning
shadow	Routine exists inertly during evaluation
ready_for_approval	Ready for explicit activation
active	Guard and authority permit evaluation/goal execution
dormant	Guard false, cooldown active, recurrence already handled, or override active
guard_uncertain	Guard contains UNKNOWN, STALE, or CONFLICTING; fail-closed
conflicted	An opposing routine blocks or wins on priority
suspended	Authority, rate, manifest, or another hard gate fails
rejected	Routine was rejected
rolled_back	Routine, linked goal, and mandate were rolled back

15. Global autonomy mode: AutonomyModeV1

Mode	Meaning
observe	Run enrolled strategies as durable shadow; dispatch count is exactly zero
supervised	Persist proposals for owner approval; approval reobserves and reevaluates before dispatch
delegated	Permit enabled exact enrollments to instantiate templates and execute at most low-risk work
paused	Continue observation, learning, and in-flight recovery; start no strategy, brain, or dispatch work

Mode is global and revisioned by an epoch. It grants no target authority by itself. An enabled AutonomyEnrollmentV2 separately freezes exact strategy, templates, targets, entities, capabilities, context/privacy, cognition route, limits, budget, expiry, privileges, and fingerprints.

yolo is only a CLI alias: enable selects delegated, disable selects paused, and status shows generic autonomy status. It creates no enrollment. Legacy AutonomyProfileV1 rows remain audit data and grant no v3 authority. See Generic plugin autonomy.

16. Autonomy decision and cognition route

An AutonomyDecisionV1 is one of NOOP, DEFER, PROPOSE_EFFECT, PROPOSE_GOAL_CANDIDATE, REQUEST_EXECUTIVE, or REQUEST_SPECIALIST. deterministic makes zero brain calls; executive and specialist admit only their declared destination; hybrid runs the strategy first and admits at most one requested cognition call. None of these decisions is authority.

17. Homey transport mode

The Homey plugin also has an operational configuration value:

Mode	Behavior
observe	Read-only; recommended starting point
act	The mutation path may open, but only together with ENGINE_HOMEY_ARMED=1, allowlists, fresh state, mandate, policy, and request-bound authorization

act is therefore only a transport kill switch. It bypasses no Engine policy. Do not confuse observe/act with ACHIEVE/MAINTAIN or IMMEDIATE/TASK/STREAM.

18. CLI preview versus activation

engine setup is preview-only by default. --activate writes the goal and mandate. This is a mutation choice in the CLI, not a persistent runtime mode.

Legacy v1

V1 contains corresponding goal, invocation, and receipt values, plus Affordance (query, action, event). New product features use v2. The v2 registry projects v1 plugins as observe-only, so a v1 action affordance grants no v2 mutation rights.

See Architecture for the sequence in which these axes come together and How Engine learns for the learning gates behind the statuses.