Tool Output Semantics

Tool results distinguish an observation, an authenticated review, and an executed action. Human labels help readers; the documented machine fields govern consumers. Compatibility labels do not replace those fields.

Host and hook responsibilities

The native host validates tool arguments and enforces permission. A reported argument error determines the correction to an authorized call. Governance hooks inspect work authority and evidence; they do not infer tool capabilities from serialized text.

Hook status distinguishes a matching installation, configuration drift, disabled hooks, missing components and an unsupported platform. Drift requires examining the affected installed binding; it does not mean that every hook is absent. The platform compatibility contract owns installation decisions.

Counts retain their declared unit. A unique finding count differs from physical occurrences, and an absent count is unknown.

Review, output and continuation

The responding model judges each claim against original evidence, then judges the complete answer against the user’s objective. Core authenticates those judgments and source bindings. A review receipt returns status and references; it does not repeat the answer body. The model submits its reviewed draft through native output. An unfavorable review or missing prerequisite remains unresolved under the review procedure.

flowchart LR
  M["Responding model: claims and whole-answer judgments"] --> C["Core: authenticate bindings"]
  C -->|Status and references| M
  M --> H["Native host: display answer"]
  H -->|Stop observation| A["Autopilot: apply eligible evidence"]
  A -->|Authorized task request| H

An accepted answer can accurately report incomplete work. Authentication, publication and criterion application are separate results. Original evidence keeps its original verification time during unchanged reuse. A recording failure does not authorize repeating the business operation.

Native Stop observes the response and permits the native conversation to continue. A diagnostic assessment does not establish answer visibility or create missing judgments. Autopilot selects continuation from an admitted plan; a returned request proves neither delivery nor execution. See the architecture and session gate owner map for responsibilities.