feat: enrich_metadata + Dropbox recursive listing (sync spec v0.7.20)

[kdr]

Summary

Syncs the Python SDK to spec v0.7.20 (cloudglue/cloudglue-api-spec#111) and bumps the package to 0.7.23.

Two new API surfaces, both additive:

• recursive — Dropbox-only subtree listing on GET /data-connectors/{id}/files
• enrich_metadata — opt-in source-metadata backfill on metadata imports, plus a files_enriched progress counter

Generated (cloudglue/sdk/)

The regen touches 312 files, but nearly all of it is the The version of the OpenAPI document: 0.7.17 → 0.7.20 header comment. The substantive changes are eight files:

• api/data_connectors_api.py — recursive query param plumbed through all three call variants
• models/metadata_import_filter_set.py — recursive
• models/metadata_import_run_progress.py — files_enriched
• models/{metadata_import,metadata_import_detail,metadata_import_run,create_metadata_import_request,create_metadata_import_run_request}.py — enrich_metadata

Wrapper (cloudglue/client/)

Unlike the JS wrapper (where request bodies are inferred), the Python wrapper passes explicit kwargs, so every new param needs hand-plumbing:

• data_connectors.list_files — new recursive kwarg, forwarded to the generated API and documented
• metadata_imports.create — new enrich_metadata kwarg, forwarded into CreateMetadataImportRequest
• metadata_imports.run — new enrich_metadata kwarg, forwarded into CreateMetadataImportRunRequest as a per-run override

Docstrings refreshed alongside: the filter-set docs described path as "non-recursive, direct children" (now covers recursive), and include_thumbnails still claimed "Grain and iconik today" with the old "materially slows large imports" caveat — v0.7.20 widens it to Grain, iconik, Google Drive, and Dropbox and decouples poster copying from indexing.

No changes needed in cloudglue/__init__.py — this spec bump adds no new models, only fields on models already re-exported there.

Test plan

import cloudglue clean; wrapper signatures confirmed to carry the new kwargs in order; MetadataImportFilterSet.from_dict({'recursive': 'true'}), MetadataImportRunProgress.from_dict({'files_enriched': 3}), and CreateMetadataImportRequest(..., enrich_metadata=True) all round-trip.

Live prod battery: 27/27 passed, mirroring the JS battery check-for-check. Script and full results markdown are saved outside the repo; the substance:

• recursive (8 checks) — against a real nested Dropbox tree: non-recursive root returns 3 folders and 0 files, recursive="true" returns 17 entries / 11 files reaching path depth 5. recursive="false" reproduces the default exactly. Scoped to path="/test": 3 direct children vs 11 recursive, recursive being a strict superset. Ignored (not an error) on a google-drive connector. An out-of-enum value is rejected with 400.
• enrich_metadata (13 checks) — full lifecycle on a throwaway metadata collection against Dropbox, capped at max_files=3, with a {"path": "/test/subfolder", "recursive": "true"} filter set (which also proves recursive round-trips inside MetadataImportFilterSet). The definition persists enrich_metadata=True across create/get/list; a run with no override inherits it; the run completed with files_enriched: 3 of 3 indexed. The payoff is verified, not just the flag: enriched files carry Dropbox media_info — {'duration_ms': 16166, 'width': 4094, 'height': 2014} — which Dropbox's list endpoint has not returned since 2019. A per-run override to False on a refresh pass is honored and reports files_enriched: 0. Import and collection were deleted afterward.
• Clip-window regression (6 checks) — clip_start/clip_end on share.get are not new, but they were absent from the internal spec snapshot and are restored in the spec PR, so they were re-verified rather than assumed: the window lands in the Mux playback-token claims as asset_start_time=3.0 / asset_end_time=9.0, an unclipped read carries no claims, and both invalid forms (half a pair, end <= start) return 400.

Gong enrichment — additionally verified, 9/9. Confirms metadata_imports.create(enrich_metadata=True) and .run() drive the Gong path through this wrapper, with Call Spotlight content modelling correctly through pydantic (parties[0].to_dict() round-trips, brief and topics present). Refresh run reported files_enriched: 1 with all five Spotlight fields populated.

One correction during that run is worth recording as real API behavior: the first attempt used append mode and reported files_enriched: 0 while the Spotlight fields still looked populated. That was not a bug and not a pass — the underlying Cloudglue file for a Gong call outlives the collection it was imported into, so an earlier run's already-enriched file was still present and append mode correctly declined to re-spend Gong API budget on it (files_created: 0, files_updated: 0). The fields on display were stale residue rather than that run's work. Re-running in refresh mode gave a genuine signal (files_updated: 1, files_enriched: 1), and the battery now asserts files_updated >= 1 so stale enrichment cannot masquerade as a pass. Capped at max_files: 1.

The deeper before/after — including proof that enriched documents are re-embedded (same query scoring 0.2580 → 0.4844) — is covered in the JS PR, cloudglue/cloudglue-js#155, since that behavior is server-side and identical for both SDKs.

Merge order

The spec PR (cloudglue/cloudglue-api-spec#111) is already merged, and this branch's spec submodule points at the squashed main commit 23b7b7f — verified byte-identical to the spec these SDKs were generated and live-tested against. No re-pointing needed; this is mergeable as-is.

[cursor]

PR Summary

Low Risk
Additive OpenAPI sync and wrapper kwargs with no auth or breaking signature removals; runtime impact depends on callers opting into enrich_metadata or recursive Dropbox listing.

Overview
Bumps the package to 0.7.23 and regenerates cloudglue/sdk/ from OpenAPI 0.7.20. Most of the diff is version header churn; the meaningful API additions are recursive on data-connector file listing and enrich_metadata on metadata imports (with files_enriched on run progress).

The hand-written client in cloudglue/client/ wires those through explicitly: data_connectors.list_files accepts recursive and forwards it to the generated API; metadata_imports.create and metadata_imports.run accept enrich_metadata on the create/run request models. Docstrings were updated for Dropbox recursive in filter sets, wider include_thumbnails sources (Google Drive and Dropbox), and async poster copying vs indexing.

^{Reviewed by Cursor Bugbot for commit 7c6db93. Bugbot is set up for automated code reviews on this repo. Configure here.}