Reasoning | agentantibody docs

Reasoning is optional. The model is an advisory investigator. The service stays authoritative for graph state, retrieval bounds, evidence, and promotion.

From change to finding

A change lands. Each publication that matters to reasoning writes one ordered event to the outbox in the same transaction. Telemetry-only updates do not.
A detector asks whether it is new. It compares the previous graph with a coherent checkpoint. A resource becoming externally addressable, typed semantics disappearing, or a governed-by edge being retracted raises a high-risk signal. A change to an existing typed envelope raises a medium-risk one. For telemetry, a new communication pair and a sensitive destination role raise a novelty score, and only scores that meet the minimum become signals. Windows with no signal never reach a model.
The model gets a small packet. Resource IDs, kinds, typed security semantics, opaque evidence handles, and the signals. No telemetry and no provider-native configuration. A packet holds at most 100 resources, 250 relationships, 250 facets, and 100 telemetry entries.
It investigates with read-only tools. Each result is bounded and pinned to the packet’s analysis token.
The harness validates. See below.
A finding is promoted or dropped.

The six tools

Tool	Returns
`get_security_neighborhood`	The one-hop security neighborhood of the anchor resources
`get_telemetry_chain`	Bounded runtime observations connected to the anchors
`get_governance_posture`	Governance relationships and facets, without inferring policy effect
`find_graph_paths`	Bounded graph paths from the anchors. Follows every relationship except `runs-on` unless the model names the predicates it wants.
`get_incident_timeline`	Bounded, ordered observations across the current incident
`get_behavior_baseline`	Bounded, previously observed communication behavior

Validation

A model answer becomes a finding only if all of these hold.

The model chose candidate_finding.
The category is not new_observed_communication. Novelty without a security consequence is rejected.
It cites a signal from the packet, entities from the packet, fact IDs the graph returned, and evidence IDs attached to those facts. A wrong reference gets one correction turn.
It selects at least one signal of medium or high risk. Only those signals are promoted.
No query it relied on hit its bound. Truncated retrieval blocks promotion.

Remediation suggestions that do not apply to the packet are dropped before the finding is stored. A baseline packet, or one with no signals, is suppressed and never becomes a finding.

Findings

Field	Values
`category`	`new_observed_communication`, `boundary_crossing`, `loss_of_governance`, `new_exposure`, `identity_reuse`, `possible_lateral_path`, `unknown_security_effect`
`severity`	`none`, `low`, `medium`, `high`
`modelUncertainty`	`low`, `medium`, `high`

A finding also lists its summary, affected entities, claims, stated uncertainties, signal IDs, and up to eight remediation suggestions. Each suggestion names one outcome from a closed list, one affected target, and no provider payload, patch, or command. Duplicate outcome and target pairs are rejected.

Finding IDs are deterministic: finding:sha256:<hex>, derived from the promoted signal IDs and the graph digest. A later change that touches an affected entity invalidates older findings. They stay available for audit but cannot launch a remediation.

Retrieval modes

--reasoning-retrieval selects how the investigator gets graph context.

model_selected lets the model choose which tools to call.
deterministic has the service fetch the behavior baseline, the incident timeline, a default graph path, and the governance posture first. The model can still make more calls.

What it does not do

Reasoning does not apply anything. It attaches desired outcomes to a finding, and remediation takes over from there. It also does not reconstruct history: on a restart it analyzes the current state, not what changed while the service was down.