Skip to content

Agent - #166

Merged
parashardhapola merged 6 commits into
masterfrom
agent
Aug 6, 2026
Merged

Agent#166
parashardhapola merged 6 commits into
masterfrom
agent

Conversation

@parashardhapola

Copy link
Copy Markdown
Member

This PR introduces the optional scarf.agent workflow for evidence-grounded single-cell data preparation and characterization.

  • Add structured, validated LLM decisions behind the scarf[agent] extra.
  • Add multi-format ingest for H5AD, Loom, Seurat, Matrix Market, 10x, and existing Zarr stores.
  • Characterize covariates, units of inference, confounding, and technical nesting.
  • Characterize feature identity, species, gene families, and exogenous candidates using Ensembl references.
  • Add reusable metadata, association, feature-identity, storage-safety, and read-only summary utilities.
  • Record decisions, accepted actions, and audit events without mutating stores during characterization.
  • Fail safely when input is ambiguous, normalized-only, overlapping with the destination, or requires explicit overwrite approval.

LLM use is limited to selecting among deterministic, precomputed evidence. Core Scarf remains independent of agent dependencies and can be imported without installing the optional extra.

parashardhapola and others added 6 commits August 4, 2026 21:47
…sions.

Introduce scarf[agent], decide/check_runtime/load_env, and isolation tests so later ingest and reanalysis stages can call a validated structured-choice primitive.
Keep imports deterministic: block prenorm-only H5AD unless matrixKey is forced, and require an explicit HTO vs ADT choice when Antibody Capture names look like hashtags.
…nding.

Label cell metadata domains and coefficients, collapse same-domain aliases, and report design-table association with an intercept-aware estimability check that skips technicals varying within the observation unit.
Introduce Ensembl-backed species/family observation with catalogSuspect warnings, and drop constant columns plus invalid finer independent units so design confounding stays general without clustering metadata. Also fix UTF-8 cell metadata, biotype assay-split gating, and decide() coercion for live H5AD runs.
@parashardhapola
parashardhapola merged commit 90b7c08 into master Aug 6, 2026
0 of 5 checks passed
@parashardhapola
parashardhapola deleted the agent branch August 6, 2026 11:25
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant