TOB

TOB stands for Tool Observation Bus. Nova uses TOB to turn agent tool activity into observed evidence that other areas can compare with agent claims.

TOB records what Nova can observe during real execution: a page was read, a selector was touched, text was entered, navigation was attempted, a tool failed, or a gate stopped a call before it ran.

A completion note is a claim. TOB evidence is the observable trail behind that claim. ETM can reject a task completion when checked units do not have enough observed support, and ALP, LCJ, PKS, OK, CLS, and AAG can use the same evidence according to their own rules.

Users rarely operate TOB directly. Agents and integrators see TOB through fields such as evidenceSummary, verificationResults, taskAwareness.completionAllowed, coverageEvidence, and task URL coverage gates.

In short

• TOB records observed tool evidence, not durable memory.
• It separates agent claims from execution signals.
• Evidence becomes useful only when bound to a task, tab, or scope.
• ETM can block premature completion with reason=evidence_gap.
• Coverage Scan provides server-registered coverageEvidence for exhaustive URL tasks.
• TOB evidence never bypasses user approval or action safety gates.

What TOB observes

TOB keeps evidence-relevant signals from agent tool work. It does not turn every event into trusted memory; it preserves the signals other systems need for review.

1. Tool dispatch TOB can record which tool ran, whether it succeeded, how long it took, which target it used, and whether it produced a logical success or an error.
2. Surface context For browser work, TOB can keep page URL, route, tab, document, selector, input, and before/after context so later checks know what the action touched.
3. Action kind Observed work is classified as read, navigate, interact, write or meta.
4. Signal flags Read, visual exposure, DOM exposure, selector touch, input supplied, navigation committed, mutation attempted, and mutation committed are examples of signals TOB can preserve.
5. Task evidence For ETM, TOB compares checked task units with observed evidence and exposes the result in evidenceSummary.
6. Coverage evidence For exhaustive URL work, nova.coverage_scan returns coverageEvidence from registered scans instead of trusting agent-supplied coverage claims.

When this matters

TOB matters when Nova must prove that work actually happened before accepting a result: large URL reviews, compliance or accessibility checks, task runs with many units, selector repair, learning from real outcomes, or any workflow where a plausible completion note is not enough.

Knowledge model for agents

For an agent, TOB answers four questions: which work was claimed, which signals were observed, how strong is the evidence, and what decision should another system make from that evidence?

Claim

A reported state from the agent, such as a task unit marked checked or a completion request for an instance.

Observation

A tool-execution signal Nova can observe, such as a read, click, input, navigation, failure, blocked call, or structured scan.

Scope

The work boundary that binds observations to a task or target. Without scope, evidence may exist but cannot safely prove a task unit.

Evidence grade

strong means observed support is strong enough for the unit; weak means some signal exists but is incomplete; none means expected evidence is missing after complete ingestion; unknown means Nova cannot judge yet.

Evidence gap

A difference between claimedCheckedUnits and evidence accepted by the active policy. If the gap is too large, completion can return reason=evidence_gap.

Coverage evidence

coverageEvidence is a registered-scan result. It becomes trusted only when server checks pass, including URL consistency and claimed-vs-measured extraction sanity.

Verification report

verificationResults summarizes explicit verification-contract steps submitted through evidenceReport.

How agents use TOB information

TOB is an evidence route, not a user-facing task list. The agent performs work, reports progress, reads evidence state, and only completes when observed evidence supports the claim.

1. Create a scoped task or claim The agent works inside a claimed tab or an ETM instance created with nova.task_instance_create. This gives TOB a boundary for evidence correlation.
2. Use normal browser tools Calls such as nova.perceive, nova.read_text, nova.navigate, nova.click_selector or nova.type_selector create observable dispatch signals.
3. Report what was checked For task instances, the agent marks units through nova.task_instance_progress. Those checked units become the claim set that TOB later tests against observations.
4. Use Coverage Scan for exhaustive URL work When an exhaustive URL task needs block-eligible evidence, the agent uses nova.coverage_scan with a registered scanId instead of claiming coverageEvidence through arbitrary evaluation.
5. Read evidence state before completion nova.task_instance_get can return evidenceSummary, urlCoverage and taskAwareness.completionAllowed so the agent can see whether completion is likely to pass.
6. Complete only when evidence supports it nova.task_instance_complete can return completed=false with verification_failed, evidence_gap or task URL coverage guidance when observed evidence is not enough.

When TOB evidence applies

TOB state describes evidence maturity. It can support or reject a claim, but it never grants permission to perform an unsafe website action.

Failure and guard conditions

These are the hard edges an agent or integration should expect when using TOB-facing fields.

Agent interpretation example

The example shows TOB as a completion contract: the agent reads evidence, sees that observed support is incomplete, and gathers more proof before retrying completion.

Task snapshot evidence

{
  "tool": "nova.task_instance_get",
  "arguments": {
    "instanceId": "task_42"
  }
}

Relevant response fields

{
  "structuredContent": {
    "evidenceSummary": {
      "claimedCheckedUnits": 12,
      "observedCheckedUnits": 9,
      "strong": 7,
      "weak": 2,
      "none": 2,
      "unknown": 1,
      "ingestionComplete": true
    },
    "taskAwareness": {
      "completionAllowed": false
    }
  }
}

Agent interpretation

{
  "claim": "12 units marked checked",
  "observedSupport": "9 units have TOB evidence",
  "risk": "completion may be rejected as evidence_gap",
  "nextStep": "inspect the none/unknown units before calling task_instance_complete again"
}

MCP contract

This is the deterministic layer under the explanation. TOB does not publish one single tool; agents read its contract through task-instance, verification, and coverage fields.

Execution rule: No TOB field may authorize a website action by itself; TOB evidence can support completion, learning, or state review only after the current task and safety gates agree.