WorkGraph Attention Goals

Status: Implemented architecture record. Initial WorkGraph goals, attention, host controls, turn overlays, tool scoping, and mob lowering shipped starting in v0.6.24. The durable WorkGraph-to-Mob Flow execution bridge shipped starting in v0.8.16. Automatic idle polling and hidden continuation production remain future work; the runtime has typed continuation admission, while ordinary turns resolve and revalidate attention before injecting their overlay.

Summary

Codex-style goals should not become a new durable goal graph in Meerkat. The durable “what must be true” part is already WorkGraph-shaped: objective, status, blockers, claims, evidence, readiness, and terminal state. The missing primitive is attention:

Goal = WorkGraph commitment + runtime attention binding

This document defines the goal feature as an optional meerkat-workgraph capability:

WorkGraphLifecycleMachine continues to own durable work item truth.
A WorkGraph-owned attention binding owns which execution owner should keep attending to which work item, in which stance.
MeerkatMachine continues to own actual runtime admission, wake, turn execution, and context delivery.
meerkat-mob lowers durable AgentIdentity bindings into WorkGraph owner keys. Core Meerkat never needs to name MobId.

The key invariant:

Work item terminality is not attention state.
Attention state is not runtime admission.
Runtime projection is not WorkGraph truth.
Mob routing is not goal ownership.

Motivation

Codex /goal persists a high-level objective and keeps the agent moving by injecting hidden continuation context whenever the thread becomes idle. That is useful, but a direct port is wrong for Meerkat. Meerkat has two important differences:

session-only operation is first-class;
most substantial work can be identity-first, mob-shaped, and role-specialized.

A reviewer in a mob should not internalize “ship the match-3 game” as its own goal. It should see the parent objective as a claim to test and should optimize for finding reasons the claim is false. Likewise, a coder should build toward a child work item, and a planner may coordinate the root. This requires a typed distinction between:

durable commitment truth;
the actor currently expected to attend to it;
the role or stance used when projecting that commitment into a turn.

Existing Authorities

WorkGraph already owns durable commitment semantics. Its public docs say it answers what work exists, what is blocked, what is ready, who claimed it, what evidence supports completion, and what is terminally done, failed, or cancelled. It owns item lifecycle, dependency legality, readiness, claims, terminal state, evidence references, event history, and projections. Runtime already owns session admission and turn execution. Input::Continuation and WakeIfIdle already exist as runtime primitives, and StartTurnRuntimeSemantics is the runtime/session carrier for handling mode, tool overlays, context appends, and turn metadata. Mob already owns multi-agent topology: roster, stable member identity, runtime-binding rotation, wiring, flows, and member lifecycle. Public mob APIs use AgentIdentity; per-runtime details may rotate. Dogma requires a single semantic owner. A side map called “goal bindings” that decides pause, resume, completion, wake, projection mode, or authority would be a hidden lifecycle machine. If the binding has semantic state, that state needs a named authority.

Decision

Do not introduce MobGoal, ThreadGoal, or a generic core GoalLifecycleMachine. Instead, add WorkGraph attention as an optional WorkGraph-owned capability:

struct WorkAttentionBinding {
    binding_id: WorkAttentionBindingId,
    work_ref: WorkItemRef,
    target: WorkAttentionTarget,
    mode: WorkAttentionMode,
    status: WorkAttentionStatus,
    machine_state: WorkAttentionMachineState,
    delegated_authority: AttentionDelegatedAuthority,
    projection_policy: AttentionProjectionPolicy,
    created_at: DateTime<Utc>,
    updated_at: DateTime<Utc>,
}

The binding points at a WorkGraph item. It does not copy the item’s title, status, blockers, evidence, or terminal state.

struct WorkItemRef {
    realm_id: String,
    namespace: WorkNamespace,
    item_id: WorkItemId,
}

The binding target is typed at the surface that owns the identity. The persisted WorkGraph attention store may normalize to WorkOwnerKey, but public callers must use domain handles:

enum WorkAttentionTarget {
    Session { session_id: SessionId },
    LoweredOwner { owner_key: WorkOwnerKey },
}

Session { session_id } is valid in session/runtime surfaces. Lowered owner keys are stored only after the owning feature has validated the source identity. AgentIdentity bindings are created through meerkat-mob, which validates the identity and lowers it to a WorkGraph owner key. Core Meerkat does not need an AgentIdentity type and never names MobId.

Attention Modes

Attention mode controls projection and tool authority. It is not WorkGraph item status.

enum WorkAttentionMode {
    Pursue,
    Coordinate,
    Review,
    Falsify,
    Judge,
    Observe,
}

Mode meanings:

Mode	Projection
`Pursue`	Treat this work item as owned work to advance.
`Coordinate`	Manage decomposition, routing, and evidence collection.
`Review`	Inspect whether the work item or child item satisfies its claim.
`Falsify`	Act adversarially; find reasons the claim is not satisfied.
`Judge`	Evaluate evidence under an explicit completion policy.
`Observe`	Provide context only; do not optimize for or against completion.

The parent work item can be the same for several agents, but their attention modes must differ. A reviewer receives the parent objective as a claim to test, not as a motivational target.

Attention Status

Attention status answers whether the runtime should continue projecting this binding. It does not close or block the WorkGraph item.

enum WorkAttentionStatus {
    Active,
    Paused { until: Option<DateTime<Utc>> },
    Superseded,
    Stopped,
}

Terminal WorkGraph item states automatically make attention inactive for continuation. The WorkGraph item remains the authority for completion, cancellation, failure, blockers, and evidence. Pause/resume should be attention-state transitions. They must not be encoded as labels, descriptions, namespace strings, or prompt text. Timed resume should target the binding, not the WorkGraph item: either Paused { until } becomes eligible again when the clock passes, or meerkat-schedule resumes the binding. A WorkGraph item may use its own snoozed_until for item-level readiness, but that fact must not substitute for one actor’s attention pause.

Machine Ownership

The implemented architecture uses a generated authority path and does not add a generic core goal machine. Current ownership shape:

WorkGraphLifecycleMachine
  Owns work item state, readiness, claims, evidence, terminality, and
  completion policy operands.

WorkAttentionLifecycleMachine
  Owns attention binding active/paused/superseded/stopped state and binding
  revision. This is a WorkGraph-owned attention machine, not a generic core
  GoalLifecycleMachine.

workgraph_attention_bundle composition
  Owns the legal WorkGraph-item-close to attention-stop handoff.

trusted WorkGraph/runtime host seam
  Revalidates attention and projection eligibility before turn-overlay or
  continuation admission. The current implementation keeps continuation
  enqueueing as trusted runtime-host authority; no automatic idle-polling
  producer ships yet.

MeerkatMachine
  Owns input admission, WakeIfIdle, active turn state, continuation, and
  delivery of typed context appends.

MobMachine
  Owns AgentIdentity membership and runtime binding resolution.

WorkExecutionLifecycleMachine
  Owns the durable handoff from one WorkGraph item to one exact Mob Flow
  attempt: launch request, uncertainty, terminal observation, evidence
  projection request, and closure feedback. It does not own the Flow run or
  the WorkGraph item's lifecycle.

Attention is WorkGraph-owned commitment metadata. Runtime wake is runtime-owned execution behavior. The generated composition is the seam between them. It must not be a store-only side table or handwritten reducer. The important dogma line is not “never create another generated machine”; it is “do not create a second durable goal authority beside WorkGraph.”

Runtime Handoff

Runtime must treat attention as an observed input, not as its own commitment truth. The shipped trusted-host path performs the revalidation and admission steps below for explicit continuation requests and ordinary turn overlays. The automatic idle trigger in step 1 remains future work. Idle continuation flow:

Runtime becomes idle, or a host explicitly requests continuation for an attention binding.
Runtime or the host-facing runtime surface asks the optional WorkGraph attention service for an eligible active binding for this SessionId.
For each candidate, the service re-reads the WorkGraph item and verifies:
- item still exists;
- item is not terminal;
- attention binding is active;
- target still resolves to the current session or identity binding;
- projection policy permits continuation.
Runtime enqueues Input::Continuation through the normal runtime admission path.
Runtime starts the turn with typed attention context in StartTurnRuntimeSemantics.
Session service applies the typed context at the turn boundary.

If any check fails, the handoff fails closed. No stale snapshot should wake an agent.

Projection Contract

Runtime should not inject raw WorkGraph fields as hidden instructions. Descriptions, labels, and evidence summaries are user-authored or agent-authored data, not policy. The attention service should emit a typed projection:

struct AttentionContextProjection {
    binding_id: WorkAttentionBindingId,
    work_ref: WorkItemRef,
    mode: WorkAttentionMode,
    binding_revision: u64,
    item_revision: u64,
    parent_refs: Vec<WorkItemRef>,
    parent_context: Vec<AttentionProjectionParentContext>,
    evidence_refs: Vec<WorkEvidenceRef>,
    authority: ProjectedAttentionAuthority,
    text: AttentionProjectionText,
}

Projection requirements:

source truth is named;
rebuild trigger is item/binding revision;
staleness policy is explicit;
role stance is typed;
max rendered shape is bounded;
parent objective text is framed as data, not instruction priority.

Example falsifier projection:

Assigned stance: falsify

Work item under test:
  Ship a match-3 game.

Your task:
  Find evidence that this work item is not yet satisfied. Do not optimize for
  release. Report blockers, missing tests, incorrect behavior, and unsafe
  assumptions. You may add evidence to the review item. You may not close the
  parent item unless explicitly granted judge authority.

Completion Authority

Completion is WorkGraph terminal state plus policy. It is not “the model said complete” unless policy grants that model authority to close the item. This proposal requires explicit completion policy on the WorkGraph item:

enum WorkCompletionPolicy {
    SelfAttest,
    HostConfirmed,
    PrincipalConfirmed,
    Supervisor { owner_key: WorkOwnerKey },
    ReviewerQuorum { threshold: u16 },
}

The attention binding may carry delegated authority, but delegation is not the canonical completion policy:

enum AttentionDelegatedAuthority {
    AddEvidence,
    CloseOwnReviewItem,
    RequestClosure,
    CloseIfPolicyAllows,
}

For adversarial roles, the normal authority should be evidence contribution, not parent closure. A reviewer may close its own review child item and attach blocking or non-blocking evidence. The parent item closes only when its policy allows it. WorkGraph host surfaces expose observability; CLI and trusted in-process hosts add narrow goal and attention controls. Trusted principal-gated goal policies belong to in-process host authority. Public REST/RPC clients can inspect goal status and attention bindings, but cannot smuggle principal authority through JSON.

Session-Only Shape

Session-only Meerkat remains first-class. Creating a session goal:

Host creates a WorkGraph item in a realm namespace such as session/{session_id}.
Host creates an active attention binding targeting SessionId.
Runtime observes the binding and continues the session while eligible.
The session agent can add evidence, block the item, or close it only if the authority policy permits.

Without WorkGraph enabled, this feature is unavailable unless a separate explicit session-local objective contract is designed. There must be no silent fallback to hidden local goal state, because that would create two goal systems.

Mob Shape

There is no mob-bound goal lifecycle. A mob objective is a WorkGraph root item, often scoped by namespace or linked through a typed external reference. Mobs route attention; they do not own work terminality. Example:

Root WorkGraph item:
  Ship a match-3 game.

Planner attention:
  target = AgentIdentity(planner)
  work = root item
  mode = Coordinate

Coder attention:
  target = AgentIdentity(coder)
  work = child implementation item
  mode = Pursue

Reviewer attention:
  target = AgentIdentity(reviewer)
  work = child implementation item plus parent context
  mode = Falsify

Current meerkat-mob support lowers AgentIdentity targets into durable owner-key attention targets. For ordinary turns, meerkat/src/surface.rs resolves those keys against the session’s mob_id and agent_identity labels, selects the newest matching session generation deterministically, and revalidates the current attention projection before composing the turn overlay. Automatic idle continuation for a lowered owner remains future work. Mob docs should avoid saying the persisted mob owns “work” in the WorkGraph sense. Mob owns roster, wiring, member lifecycle, flows, and routing. WorkGraph owns shared durable work.

Flow Execution Seam

Flow execution uses an explicit durable association rather than a prompt-level WorkGraph id. The controlling host commits WorkExecutionBinding before launching the Flow. The binding fixes the Mob and Flow identities, effective run-config digest, deterministic run id, activation parameters, idempotency key, and retry predecessor. WorkExecutionLifecycleMachine emits the next cross-system obligation. The composition shell realizes that obligation and feeds the typed outcome back:

FlowLaunchRequested
  -> Mob external-delivery admission
  -> deterministic pending MobRun persistence
  -> durable realization claim
  -> Flow running, completed, failed, canceled, or launch uncertain

EvidenceProjectionRequested
  -> idempotent WorkGraph evidence write

WorkClosureRequested
  -> WorkGraphLifecycleMachine closure admission
  -> WorkClosed or EvidenceProjected

No Flow status is copied into WorkGraph as a second authority. Reconciliation reads the canonical Mob run and advances only through generated execution transitions. A completed Flow never bypasses WorkGraph completion policy. The pending run precedes the realization claim. This ordering makes the cross-store crash boundary recoverable without pretending to have a distributed transaction: before pending-run persistence, begun proves target absence; after it, the exact run is the durable target fact. A realizing record with no run fails closed as quarantine. The public bridge cannot resolve that state from caller testimony. Recovery requires the deterministic run to become observable or a future sealed host authority to prove a safe resolution.

Tool Authority

Current WorkGraph exposure is coarse: enabling WorkGraph gives a member create, claim, update, block, close, link, and evidence tools. That is too broad for adversarial roles. Attention projection should narrow tool authority by mode and work item:

Mode	Typical authority
`Pursue`	Claim/update assigned item; add evidence; close if policy allows.
`Coordinate`	Create/link child items; assign attention; collect evidence.
`Review`	Read target/ancestors; add evidence; close review child item.
`Falsify`	Read target/ancestors; add blocking evidence; cannot close parent.
`Judge`	Evaluate evidence; close target only under explicit policy.
`Observe`	Read-only context.

Attention-aware filtering should be the final capability-narrowing layer:

base tool registry
  -> ToolCategoryOverride / session config
  -> mob profile capabilities
  -> attention projection scope
  -> per-call WorkGraph service authorization

ToolCategoryOverride remains coarse-grained capability exposure. Mob profiles remain role defaults. The final authorization check must know the attention binding id, mode, work item scope, and requested operation before a WorkGraph mutation reaches WorkGraphService.

Public Surfaces

Suggested user-facing vocabulary:

/goal <objective>: create WorkGraph item plus session attention binding.
/goal pause: pause attention binding, not WorkGraph item.
/goal resume: reactivate attention binding if item is still non-terminal.
/goal clear: stop the attention binding; item remains unless user explicitly cancels it. Supersession is reserved for future binding replacement flows.
/goal status: show attention status plus referenced WorkGraph item status.

Current trusted host surface:

workgraph/goal/create: transactionally create a WorkGraph item and its initial attention binding.
workgraph/goal/status: return referenced item status plus attention status.
workgraph/goal/confirm: attach host/principal confirmation evidence.
workgraph/goal/request_close: request policy-gated item closure.
workgraph/attention/list: list active attention bindings.
workgraph/attention/pause: pause a binding without mutating the item.
workgraph/attention/resume: resume a paused binding if the item remains non-terminal.
trusted runtime-host attention continuation: enqueue a hidden continuation for a session-bound binding.

These are WorkGraph-feature host APIs. Public REST and JSON-RPC entrypoints expose goal/attention observability only; mutating goal/attention calls are trusted in-process host/runtime authority and are not part of the public catalog. They are unavailable when WorkGraph is not compiled or enabled. The CLI mirrors the narrow session-first subset under rkat workgraph: goal-create, goal-status, goal-confirm, goal-close, attention-list, attention-pause, and attention-resume. The runtime has a typed trusted-host continuation seam, but no automatic idle-polling producer or REST-backed CLI command ships today. CLI confirmation does not accept raw principal authority. The CLI intentionally does not expose the broad WorkGraph mutation surface as the goal UX. Core host goal creation targets a session handle. Mob/agent targets are not accepted through this core host contract; meerkat-mob must validate AgentIdentity and lower it to a WorkGraph owner binding through its own feature surface. This keeps core Meerkat from freezing a public WorkOwnerKey::Agent or WorkOwnerKey::Mob bypass.

Failure Modes

The design must fail closed when:

WorkGraph is disabled;
WorkGraph item is missing;
item revision changed in a way that invalidates the projection;
item is terminal;
attention binding is paused/stopped/superseded;
target identity no longer resolves;
runtime is not idle;
queued user or peer work has priority;
projection cannot be built safely;
authority policy would allow self-review of a parent item.

Non-Goals

Do not add MobGoal.
Do not add a core GoalLifecycleMachine duplicating WorkGraph.
Do not make WorkGraph readiness equivalent to runtime wakeup.
Do not use namespace strings as semantic proof of session, mob, or agent ownership.
Do not inject a parent objective as every member’s own goal.
Do not let reviewer/adversarial profiles depend only on prompt text for their stance.
Do not provide a local fallback goal table when WorkGraph is disabled.

Resolved Recommendations

Model attention as WorkGraph-owned state with WorkAttentionLifecycleMachine and a trusted runtime-host continuation seam. Do not add a generic core goal machine.
Model pause/resume on the attention binding. Timed pauses may use Paused { until } or schedule-driven resume, but should not mutate item readiness.
Expose a narrow host/principal mutation surface for goal creation, pause/resume/stop/reassign, status, confirmation evidence, and policy-gated closure. Do not expose a broad WorkGraph editor as the first goal surface.
Put completion policy on the WorkGraph item. Let attention bindings carry delegated authority only.
Apply attention-aware tool filtering after session config, category overrides, and mob profile tooling, with a final per-call WorkGraph authorization check.

Acceptance Bar

The implementation remains conformant only if:

WorkGraph remains the only owner of durable work truth.
Runtime remains the only owner of turn admission and continuation.
Attention binding lifecycle is WorkGraph-owned and generated, currently via WorkAttentionLifecycleMachine.
The WorkGraph/runtime handoff has one named seam that revalidates attention eligibility and enters runtime through normal continuation admission.
Mob remains optional and feature-owned.
Session-only operation remains first-class.
Adversarial roles receive typed falsification/review projections, not the parent objective as their own motivational goal.

Reference

Architecture

Historical Implementation Records

Operations

WorkGraph Attention Goals

WorkGraph Attention Goals

Summary

Motivation

Existing Authorities

Decision

Attention Modes

Attention Status

Machine Ownership

Runtime Handoff

Projection Contract

Completion Authority

Session-Only Shape

Mob Shape

Flow Execution Seam

Tool Authority

Public Surfaces

Failure Modes

Non-Goals

Resolved Recommendations

Acceptance Bar

​WorkGraph Attention Goals

​Summary

​Motivation

​Existing Authorities

​Decision

​Attention Modes

​Attention Status

​Machine Ownership

​Runtime Handoff

​Projection Contract

​Completion Authority

​Session-Only Shape

​Mob Shape

​Flow Execution Seam

​Tool Authority

​Public Surfaces

​Failure Modes

​Non-Goals

​Resolved Recommendations

​Acceptance Bar

WorkGraph Attention Goals

Summary

Motivation

Existing Authorities

Decision

Attention Modes

Attention Status

Machine Ownership

Runtime Handoff

Projection Contract

Completion Authority

Session-Only Shape

Mob Shape

Flow Execution Seam

Tool Authority

Public Surfaces

Failure Modes

Non-Goals

Resolved Recommendations

Acceptance Bar