{{Framework name}}

1. Snapshot

Identify the exact source boundary so another reader can reconstruct every citation and count.

The SHA is required because a date cannot reconstruct a moving repository. Run measurements against that checkout and say exactly what each command proves; do not use an adjacent metric as evidence.

2. The core loop

Trace one agent turn from input through model decision, action, observation, and termination, naming the pinned files, functions, and executable tests that establish the path.

3. State and persistence

Describe the in-memory state, persisted state, storage boundary, and resume semantics, including what survives a process crash and what is replayed rather than restored.

4. Tools and the outside world

Follow a tool from declaration through schema/validation, invocation, isolation or sandboxing, result handling, and failure return to the agent.

5. Composition

Explain whether the pinned implementation composes agents through handoffs, sub-agents, graphs, or another boundary, and identify what data and control cross that boundary.

6. The human in the loop

Trace interruption, approval, rejection, editing, and steering mechanisms, or state the bounded evidence for their absence.

7. The abstraction bargain

Mark as Opinion the judgement about what the framework makes easy, what it makes hard, and what it deliberately leaves to applications; keep the implementation evidence beside each judgement.

8. What rungs takes

Map findings to the canonical catalogue without redefining patterns or changing the catalogue in a per-framework extraction.