Skip to content
MAESTRO

MAESTRO Metadata

MAESTRO stamps a typed provenance block onto chain.metadata["care"] so a saved chain remembers the task it was generated for, the files attached, who made it, and its tags. You can read/write it with two ReasoningChain methods.

Write & read

Section titled “Write & read”
from mmar_carl import CareChainMetadata, CareContextFile


# kwargs form (handy in code / tests)
chain.set_care_metadata(
    task_description="Summarise the quarterly report",
    context_files=[CareContextFile(path="report.pdf", size_bytes=20480)],
    display_name="Quarterly summariser",
    tags=["finance", "summary"],
)


# or hand over a ready model (CARE's usual path)
chain.set_care_metadata(meta=CareChainMetadata(task_description="..."))


meta = chain.get_care_metadata()   # -> CareChainMetadata | None

Pass either meta= or the individual kwargs — mixing raises ValueError. get_care_metadata() returns None when the chain has no care block (i.e. it wasn’t created by a MAESTRO-aware tool).

CareChainMetadata fields

Section titled “CareChainMetadata fields”

task_description, context_files (list of CareContextFile{path, size_bytes}), generated_by, mage_metadata (dict), display_name, description, tags. The namespace key is CARE_METADATA_NAMESPACE ("care").

Re-prime a context from a saved chain

Section titled “Re-prime a context from a saved chain”

ReasoningContext.from_chain_inputs builds a fresh context from a chain’s MAESTRO metadata — the “re-run from the library” entry point:

context = ReasoningContext.from_chain_inputs(
    chain,
    api=client,
    outer_context=None,             # falls back to the saved task_description
    load_files_from_metadata=True,  # re-read the attached context_files
)
result = await chain.execute_async(context)

Pass files={...} to override file contents, or outer_context= to override the input; any extra **kwargs (e.g. language=, system_prompt=) pass through.

See also

Section titled “See also”