Command line

schemacast casts text; schemacast-bench runs the benchmark. Model settings come from the environment.

schemacast sheet

Print what a model is shown for SCHEMA. Makes no model call.

Option	Value	Meaning
--schema	FILE	GraFlo schema or manifest (YAML or JSON). (required)
--json-schema	flag	Print the JSON Schema of the reply instead of the type sheet.

schemacast select

Show the slice of SCHEMA each text would be cast against. Makes no model call. Prints one JSON line per text: the mode, the seeds and how each lane ranked them, the retrieval metrics and the slice's vertex and edge counts.

Option	Value	Meaning
--schema	FILE	GraFlo schema or manifest (YAML or JSON). (required)
--text	TEXT	A single text to select for.
--episodes	FILE	Episodes, one JSON object per line.
--sheet	flag	Print each slice's type sheet.

schemacast cast

Cast EPISODES into SCHEMA and save the result under OUT. Each text becomes a mention graph; what its keys identify is the graph, the rest is staged. With --db, keys are looked up there too; a manifest's connector-owned types are linked, never created. Episodes that produce nothing are quarantined and the run goes on; a misconfigured or unreachable model, or a cache miss under --replay, stops it. Exit status: 0 when every episode was cast or failed on its own content (too long, refused, truncated, unparseable), which a rerun reproduces; 75 when some failed at the provider, which a rerun retries; 1 when the run stopped or every episode failed.

Option	Value	Meaning
--schema	FILE	GraFlo schema or manifest (YAML or JSON). (required)
--episodes	FILE	Episodes, one JSON object per line. (required)
--out	DIRECTORY	Directory for the graph, groundings, quarantine and manifest. (required)
--db	FILE	graflo connection file (YAML); the graph is written there too.
--replay	flag	Serve every reply from the cache; a miss stops the run. Makes no request.
--recreate-schema	flag	Drop and redefine the schema on the --db target first.
--concurrency	INTEGER	Episodes cast at once. (default 8)

schemacast pick

Link staged mentions to their candidates: one model call per episode. Each staged mention gets one of its candidates or none, checked without a model. A pick links; it writes nothing. Exit status: 0 when every call answered; 75 when some failed at the provider, which a rerun retries; 1 when the run stopped or every call failed.

Option	Value	Meaning
--input	FILE	One JSON object per line: an episode, its staged mentions with their candidates, and the entities its metadata names. (required)
--out	DIRECTORY	Directory for picks.jsonl, episodes.jsonl and manifest.json. (required)
--replay	flag	Serve every reply from the cache; a miss stops the run. Makes no request.
--concurrency	INTEGER	Episodes picked at once. (default 8)

schemacast serve

Serve POST /cast and POST /pick over HTTP (needs the server extra). Every request needs Authorization: Bearer $SCHEMACAST_SERVE_API_KEY; without a key the service refuses to cast unless SCHEMACAST_SERVE_ALLOW_ANONYMOUS is set. Model settings come from SCHEMACAST_LLM_*.

Option	Value	Meaning
--host	TEXT	Address to bind; default SCHEMACAST_SERVE_HOST or 127.0.0.1.
--port	INTEGER	Port; default SCHEMACAST_SERVE_PORT or 8997.

schemacast-bench

Cast and score the corpus; write JSON and Markdown reports.

Option	Value	Meaning
--model	TEXT	Model to benchmark, repeatable; the endpoint is SCHEMACAST_LLM_BASE_URL. Without it only the free oracle row runs. (repeatable)
--oracle / --no-oracle	flag	Add the oracle row: the expected graph through the whole pipeline. Free. (default True)
--split	dev | test | all	dev is tuned against; only test is publishable. (default all)
--family	TEXT	Only these families. (repeatable)
--case	TEXT	Only these cases, e.g. units/lab. (repeatable)
--corpus	DIRECTORY	A corpus of your own: DIR/cases (with SPLITS.yaml) and DIR/schemas, laid out like the bundled one. Never publishable.
--replay	flag	Serve every reply from the cache; a miss stops the run.
--out	DIRECTORY	Where the JSON and Markdown reports go. (default bench-results)
--concurrency	INTEGER	(default 8)

schemacast-bench score

Score saved casts against the corpus, with no model. Each CASTS file is a POST /cast response body, or {"episodes": [...], "cast": <body>}. A corpus episode no file holds counts as missed.

Option	Value	Meaning
CASTS	argument
--corpus	DIRECTORY	A corpus of your own: DIR/cases (with SPLITS.yaml) and DIR/schemas, laid out like the bundled one. Never publishable.
--split	dev | test | all	dev is tuned against; only test is publishable. (default all)
--family	TEXT	Only these families. (repeatable)
--case	TEXT	Only these cases, e.g. units/lab. (repeatable)
--row	TEXT	Row label; default: the casts' model.
--out	DIRECTORY	Where the JSON and Markdown reports go. (default bench-results)

schemacast-bench compare

Read REPORT_B against REPORT_A, episode by episode, on the same corpus.

Option	Value	Meaning
REPORT_A	argument
REPORT_B	argument
--row-a	TEXT	Row of REPORT_A; default: its only row, or its only model row.
--row-b	TEXT	Row of REPORT_B; default as for --row-a.
--pair-by-name	flag	Across corpus versions, pair cases a report has no content digest for by name: you assert they did not change.
--json	FILE	Also write the comparison as JSON here.