Repository navigation
Architecture
언어: 🇺🇸 English | 🇰🇷 한국어
The user supplies the objective, constraints and permission. The model responding in the current conversation selects work and judges meaning. Core is an instruction and code package. The native host executes requested tools and registered hooks and returns results to the relevant conversation.
| Handoff | Input and actor's action | Return and next condition |
|---|---|---|
| Model → host → Core | The model supplies the current session/input and authorized work selection. Core validates the route bindings. | The stored route or a named error returns to the model. The host runs only permitted tools and returns actual observations. |
| Model → host → Core → model | The model identifies the complete draft, claims under review and original sources. Core returns exact review inputs and references. | The model judges each claim against its original sources, then makes a separate whole-response judgment using those judgments, the complete draft and the user's objective and constraints. |
| Model → host → Core → model | The model submits both judgments. Core validates their bindings and stores them together. | Only supported reviews return an answer for submission. An unsupported judgment or missing prerequisite returns a decision input for affected correction or unresolved reporting. |
| Host → Core → host | At Stop, the host adapter passes the answer and the current input to Core. Core compares the completion claim with the recorded review and the remaining necessary work. | Supported with no remaining work: the host stops with no added output (one exception: when the session first waits for an input whose origin Core could not confirm, the Stop can carry one intake warning, once per session and cause). Otherwise the host shows the progress and the recorded necessary work, then stops. |
At Stop, Core compares and reports, then the host stops. The host has already displayed the answer, so Core's notice follows it.
flowchart TB
T["Host, Claude Code or Codex: the model finishes an answer and Stop fires"] -->|Host-specific Stop input| AI["Host adapter: convert to the common input"]
AI -->|Common input: answer and current input| CG{"Core: is the completion claim supported by the<br/>recorded review, and does necessary work remain?"}
CG -->|Common result: supported, no remaining work| AO["Host adapter: no added output,<br/>except one warning for a first input wait"]
CG -->|Common result: unsupported or work remains,<br/>with progress and recorded necessary work| AN["Host adapter: convert to a notice"]
AO -->|Stop allowed| E["Host: stop and return control to the user"]
AN -->|Notice shown, stop allowed| E
An explanation creates no new completion proof. Recording, host delivery and business execution are separate results.
A supported answer can accurately describe unfinished work; partial evidence does not establish whole-task completion. A mismatch's location does not prove its cause, and the model selects investigation or correction only as needed for the objective and required dependencies. A recording error does not justify replaying unchanged business work.
The source architecture owns the runtime diagram, exact interfaces and implementation references. The owning review procedure defines calls, evidence reuse, recording recovery and delivery. Design principles, verification and evidence, and focus and recovery explain the model's reasoning.
Core binds native tool activity to the latest human input, including a new question inside an ongoing host turn. Earlier tool calls do not make that question a tool-using request; unresolved actor or intake binding stays unavailable. Registered source-backed explanations use their own complete-response review. Automatic retained-claim matching applies only when original retained claims cover the entire answer, and explicit evidence reports remain validated. Review lookup permits only blank separator lines between a completion control marker and its first field; stored original proof and source digests remain unchanged.
The installed Core guard loads the ledger and related modules from the selected Core runtime deployment. The source checkout is an installation input, not a fallback for missing installed modules.