Session publication evidence
harn session publication-evidence SESSION --project-root PATH reads the
existing canonical session store and emits one JSON object with schema
harn.session_publication_evidence.v1. This diagnostic command does not start
an agent, create a missing store, or emit transcript prose. Missing sessions,
incomplete event coverage, and a store changing during the read return an error.
Coverage and origin#
stored_event_count, last_event_id, and chain_root_hash describe the complete
stable event read. assistant_message_count counts every assistant row, and
records retains each row and its source_event_id. missing_source_count
includes all origins.
Each record's origin is one of:
| Origin | Meaning |
|---|---|
harness_bookkeeping | The transcript owner's boolean harn_bookkeeping_turn: true marks a harness-authored message, such as a withdrawn-answer placeholder. |
unmarked_assistant | The assistant row has no true bookkeeping flag. This does not prove that a model was called. |
bookkeeping_count counts the explicitly marked harness rows. Bookkeeping rows
have no model call role, call stage, or publication disposition. Their absent
model metadata does not contribute to missing model metadata counts.
Model metadata#
For unmarked assistants, call_role, call_stage, stage-metadata presence and
validity, and publication disposition come from the existing transcript owners.
Missing or unrecognized metadata remains counted explicitly. A string or other
malformed bookkeeping flag does not exempt a row from those diagnostics.
pending_count, published_count, and withheld_count count only measured
publication decisions on unmarked assistants. Missing metadata never implies
publication, withholding, or an empty successful model-dispatch census.