Exclude Self-Generated Feedback Events

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

A learning observer’s reads/writes enter its own queue, generating more observers. Teaching regular interactive events should remain recorded while explicitly automated child events are filtered before metadata or new analysis creation.

Mechanism

Propagate an inspectable marker on owned automated runs and check it at capture entry. Automated events skip recursive queue/lazy initialization; permitted interactive events remain. Cover producer paths without fuzzy name/path filters swallowing valid activity.

Bad example

Capture every observer action and launch more observers; filter everything named observer, including human actions.

Good example

Mark the teaching observer child explicitly; the source instance uses ECC_SKIP_OBSERVE=1. Check before project metadata/new work and skip self-feedback while preserving authorized interactive runs. Test propagation/placement across entrypoints. Do not infer actor from broad names or treat the cooperative marker as authentication.

Why the change matters

Filtering after scheduling does not prevent work already created. Entry filtering cuts self-excitation while retaining valid interaction evidence.

Observable expectation

Automated teaching runs create neither new learning events nor lazy metadata; permitted interactive versions of the same tool operation remain observed. Tests expose missing markers/entrypoints rather than one quiet run proving coverage.

Inspect scoped filtering without hidden legitimate activity.

Limits

Markers can be mistaken/forged and are not isolation. Host identity/authorization remain separate. Frozen variables/entrypoint lists are not universal runtime mechanisms; no learning Hook is enabled here.