Architecture

简体中文

flowchart LR
  A[Caller goal and context] --> B[Program collector]
  B --> C[Private evidence archive]
  B --> D[Context admission and exact rules]
  D --> E[Jev typed decisions]
  E --> F[Program reducer and freshness guard]
  F --> G[Compact agent packet]
  G --> H[Existing authorized executor]

Collection and inference stay inside the program. The main agent sees the selected packet, not a raw-output roundtrip followed by a second classification.

analysis.py validates/plans/reduces. pool.py runs independent requests with up to 30 workers. provider.py owns fixed-origin HTTPX transport, credential loading and bounded retries. command.py captures trusted argv under a byte/time budget. tools.py owns specialized collectors and guards. batch.py preserves existing native question contracts and packs only compatible schemas with isolated state.

There is no result cache. Each live question is evaluated again. Receipts are inspectable evidence, not replayed authority. Unknown failed usage is marked incomplete; returned token usage is counted once, never replicated per record.

Source text can contain prompt injection or incorrect claims. Typed outputs and program checks reduce the execution surface but do not make model judgments true. Commands are caller-authorized, not generated by Jev. Read SECURITY.md for the precise trust boundaries.

Hosted execution

flowchart LR
  S[Surface: browser page or desktop window] -->|observe| T[Indexed controls + visible text]
  T --> Q[space.py: one request, operation + target heads]
  Q --> K[kernel.py: gates, budgets, cycle and stall checks]
  K -->|fresh target| S
  K -->|pause| H[Caller: confirm / value / sign-in]

The execution loop reuses provider.py for every Jev request, so credentials, retries and usage accounting behave exactly as in the filtering commands. See hosted execution for behavior and limits.