data_ingestor¶
mode: subagent — always the first execution step, after the plan is
approved.
What it does¶
Standardises the raw catalogue and writes manifest.json by calling
ingest_to_session — one call handles
inspect → column resolution → standardisation → per-bin stats → ancillary
file discovery. For a 3×2pt task it dispatches once per probe (see
Multi-format ingestion).
Registry decisions¶
Owns the entire dataset registry decision surface — bootstrap is registry-agnostic. Two file-backed human gates can open per catalogue:
registry-<basename>— opens BEFORE any standardisation when the catalogue'sDATA_DESCRIPTION.mdhash matches one or more registered datasets.options=[<candidate_dataset_id_1>, ..., "register-new", "ignore"].registration-<basename>— opens AFTER standardising a previously unseen dataset.options=["register", "ignore"].
ingest_to_session returns the gate_id string and an
orchestrator_relay string alongside needs_input; the agent prints
the relay verbatim so @cosmotron closes the gate via
authorise_gate(session_dir, gate_id, decision) — the ONLY close path,
carrying opencode's permission: ask on the primary agent. No free-text
dataset_decision=/user_reply= kwargs remain on bootstrap_session
or the ingest surface; the decision lives entirely in the gate file and
the runtime confirmation dialogue.
Remote ingest. Data on a cluster (site:path data_dir) or a session that
pins ingest remote makes ingest_to_session run as a Slurm job — the agent
polls the async receipt to completion, unchanged otherwise. See
Remote execution.