{
  "skills": [
    {
      "id": "skl_atlas_telemetry_publisher",
      "name": "atlas-telemetry-publisher",
      "plugin": "cwc-data-engineering",
      "description": "Publish the Coworkers board's session telemetry into the REAL Netflix Atlas standalone server on :7101 (tsk_atlas_publish_telemetry, W2 of the Atlas fleet plan). Reads the frozen SessionEvent JSONL wire shape (trace_id/span_id/coworker/sprint_id/kind/ts) from `~/Library/Application Support/Coworkers/telemetry.jsonl`, aggregates counts per UTC step bucket per (coworker, sprintId, kind), and plain-POSTs `/api/v1/publish` payloads (`{\"tags\":{\"name\":\"coworker.session.events\"},\"metrics\":[...]}`) — no spectator-py. Two modes: `uv run atlas-publish replay` (full-history backfill; chronological per-series order is MANDATORY — Atlas silently drops out-of-order older points with HTTP 200, live-verified) and `uv run atlas-publish follow --interval 60` (foreground loop, closed buckets only; no daemon — launchd is a gated doc-only note). Source-list design: W5's cf.requests source plugs in additively. KNOWN SERVER LIMITS (live-verified 2026-07-05, see README table): graph steps >1h return empty legend under the current block-size-60 conf (step=1d presets need W1 conf change or W3 query fallback); consolidation averages so append `:cf-sum` to read raw counts; query index lags publishes by 10-60s; backfill accepted to >=14d. Use for publishing/backfilling fleet telemetry into Atlas, debugging why a published series doesn't graph, or adding a second metric source. Trigger on \"atlas publish\", \"replay telemetry\", \"backfill atlas\", \"follow mode\", \"publish payload\", \"cf-sum\", \"coworker.session.events\". Do NOT use for running the Atlas server itself (scripts/atlas/run-atlas.sh, cwc-engineering), the graph-API prototype (cwc-engineering's atlas-graph-prototype), preset/manifest changes (W3, schemas/atlas-graph-presets.json), or the star-schema ERD (session-telemetry-erd).",
      "summary": "Publish the Coworkers board's session telemetry into the REAL Netflix Atlas standalone server on :7101 (tsk_atlas_publish_telemetry, W2 of the Atlas fleet plan).",
      "gated": 0,
      "source_path": "plugins/cwc-data-engineering/skills/atlas-telemetry-publisher/SKILL.md",
      "created_at": "2026-07-10 18:13:13",
      "cite_as": "https://subagentskills.com/api/skills/skl_atlas_telemetry_publisher"
    },
    {
      "id": "skl_corpus_cleaning_pipeline",
      "name": "corpus-cleaning-pipeline",
      "plugin": "cwc-data-engineering",
      "description": "Typed models (Pydantic + FilterReason enum) and verified, profiled performance/library-quality reference notes for an offline LLM-pretraining-corpus cleaning pipeline — schema validation, vectorized quality filters, exact-hash dedup, datasketch-backed MinHash+LSH near-dedup, real tiktoken token counting, a composite quality-ranking score, sharded parquet output. Grounded in the real pipeline.py at organizations/workspaces/projects/2026-07-06-xai-interview-prep/ take-home/submission/ (an xAI Data Engineer take-home submission), not a hypothetical framework — this skill's SKILL.md and models.py cite that file's actual stage names, actual measured thresholds, and its actual library choices. `FilterReason`'s enum members are grepped 1:1 from pipeline.py's real filter-name strings (tests/test_models.py enforces this as a live drift check, not a static claim) — if pipeline.py's filters change, this skill's tests fail until the enum is updated to match. Deliberately does NOT add pydantic as a runtime dependency to pipeline.py itself; this skill is the typed documentation/verification layer alongside it, scope-honestly framed as a DELIBERATE broadening of cwc-data-engineering's charter beyond its original Coworkers session-telemetry domain (user-approved explicitly, not an organic extension — see README.md's \"Scope\" section). Use when working on corpus/document-cleaning batch pipelines, MinHash/LSH near-dedup, tiktoken-based token accounting, or the xAI take-home submission specifically. Do NOT use for Coworkers board/session telemetry (sibling `session-telemetry-erd` skill's territory) or Cloudflare cache/CDN snapshots (cwc-analytics-engineering).",
      "summary": "Typed models (Pydantic + FilterReason enum) and verified, profiled performance/library-quality reference notes for an offline LLM-pretraining-corpus cleaning pipeline — schema validation, vectorized quality filters, exact-hash dedup, datasketch-backed MinHash+LSH near-dedup, real tiktoken token c...",
      "gated": 0,
      "source_path": "plugins/cwc-data-engineering/skills/corpus-cleaning-pipeline/SKILL.md",
      "created_at": "2026-07-10 18:13:13",
      "cite_as": "https://subagentskills.com/api/skills/skl_corpus_cleaning_pipeline"
    },
    {
      "id": "skl_ledger_quality",
      "name": "ledger-quality",
      "plugin": "cwc-data-engineering",
      "description": "Data-quality checks over the REAL fleet-deploy ledger (`fact_fleet_deploy` in local Postgres, migration `crates/durable-store/migrations/postgres/ 008_fleet_deploy_ledger.sql`) — typed rows (Pydantic `LedgerRow` + a `DeployStatus` enum whose 6 members are the migration's own `CREATE TYPE deploy_status` values, enforced by a live drift test that re-parses the SQL at test time), pure-function invariant checks (all-terminal after a completed batch, unique (domain, wave), verified rows carry verified_at, skipped rows carry notes), and a thin psql-backed CLI (`uv run ledger-quality check`) for live runs. Use when auditing a deploy wave's ledger state after a batch, validating ledger rows from any non-Postgres source (fixtures, CSV export, a merge), or checking the ledger before a closeout report. Do NOT use to WRITE ledger rows (plugins/cwc-deploy/skills/fleet-expansion/ scripts/ledger.sh is the only writer) or for deploy step-log timing analytics (cwc-analytics-engineering's deploy-wave-analytics owns the wave logs).",
      "summary": "Data-quality checks over the REAL fleet-deploy ledger (`fact_fleet_deploy` in local Postgres, migration `crates/durable-store/migrations/postgres/ 008_fleet_deploy_ledger.sql`) — typed rows (Pydantic `LedgerRow` + a `DeployStatus` enum whose 6 members are the migration's own `CREATE TYPE deploy_s...",
      "gated": 0,
      "source_path": "plugins/cwc-data-engineering/skills/ledger-quality/SKILL.md",
      "created_at": "2026-07-10 18:13:13",
      "cite_as": "https://subagentskills.com/api/skills/skl_ledger_quality"
    },
    {
      "id": "skl_parquet_shape_profiler",
      "name": "parquet-shape-profiler",
      "plugin": "cwc-data-engineering",
      "description": "Compute a structured, typed data-shape profile (DatasetShapeProfile: row/column counts, file size, per-column dtype + kind + null rate, numeric percentiles [min/p25/median/p75/max/mean] or categorical distinct-count + top-values) for any parquet dataset, as validated Pydantic JSON — `uv run shape-profiler <path.parquet>` (add `--out profile.json` to write to a file). Every number in the output is a REAL computed statistic, never a placeholder — the same \"capture is source-shaped, never invented\" discipline the sibling corpus-cleaning-pipeline skill applies to FilterReason. This skill owns COMPUTING the numbers; the sibling plugin cwc-design's `data-shape-svg` skill owns rendering the resulting JSON as SVG markup — a deliberate cross-plugin split (profiler computes, design draws) first proven end-to-end on the xAI take-home's real 100MB sample and rendered live on agenttables.com. Use when the user wants to profile a parquet file's shape/structure, needs real column statistics before designing filters or visualizations, or asks \"what does this data look like\". Do NOT use to render the visualization itself (data-shape-svg's territory), to clean/filter the data (corpus-cleaning-pipeline), or for Coworkers board/session telemetry (session-telemetry-erd).",
      "summary": "Compute a structured, typed data-shape profile (DatasetShapeProfile: row/column counts, file size, per-column dtype + kind + null rate, numeric percentiles [min/p25/median/p75/max/mean] or categorical distinct-count + top-values) for any parquet dataset, as validated Pydantic JSON — `uv run shape...",
      "gated": 0,
      "source_path": "plugins/cwc-data-engineering/skills/parquet-shape-profiler/SKILL.md",
      "created_at": "2026-07-10 18:13:13",
      "cite_as": "https://subagentskills.com/api/skills/skl_parquet_shape_profiler"
    },
    {
      "id": "skl_session_telemetry_erd",
      "name": "session-telemetry-erd",
      "plugin": "cwc-data-engineering",
      "description": "The session-telemetry ERD and durable data spine under the Coworkers sprint board — TWO layers, fivetran-style. (1) The CONNECTOR layer (`connector_schema.sql` + connector.py): source-shaped tables named after the REAL fivetran claude-compliance/claude-platform connector entities (organization, chat, chat_message, chat_message_usage w/ exact message_usage_report_result token columns, project, model, skill, claude_code_usage_report) plus the Cowork SDK's own on-disk shapes (cowork_vm_session from bridge-state.json's environmentId/remoteSessionId/processedMessageUuids), scoped to what a claude.ai 20x subscription can ACTUALLY capture — everything else lives in the UNAVAILABLE ledger (rendered as the ERD's dark legend callout) with the API each table would need. `uv run telemetry-erd connector-sync --db <file> [--plugins …] [--email …]` syncs the real Mac sources; `connector-erd --out <svg>` renders the fivetran-grammar ERD; `connector-schema` prints the DDL. (2) The MART layer (`schema.sql`): Kimball fact_/dim_ star schema (fact_task/fact_session_event/dim_coworker/dim_sprint/dim_status) derived downstream for the sprint board, with the cross-language PARITY guard (Swift CoworkersModel ↔ task-session JSON schema ↔ SQL). `load --board <board-state.json>` / `schema` as before. NEVER invent fact_/dim_ names at the capture layer — capture is source-shaped; marts are derived. Use for session/board/telemetry modeling, the claude.ai connector equivalent, ERDs, cross-language parity. Trigger on \"connector\", \"fivetran\", \"ERD\", \"star schema\", \"fact table\", \"Kimball\", \"session telemetry\", \"cowork vm data\", \"usage report\". Do NOT use for WBR analysis (cwc-data), Cloudflare snapshots (cwc-analytics-engineering), or the Swift UI (cwc-swift-engineering).",
      "summary": "The session-telemetry ERD and durable data spine under the Coworkers sprint board — TWO layers, fivetran-style.",
      "gated": 0,
      "source_path": "plugins/cwc-data-engineering/skills/session-telemetry-erd/SKILL.md",
      "created_at": "2026-07-10 18:13:13",
      "cite_as": "https://subagentskills.com/api/skills/skl_session_telemetry_erd"
    },
    {
      "id": "skl_wbr_warehouse",
      "name": "wbr-warehouse",
      "plugin": "cwc-data-engineering",
      "description": "Materialize and refresh the weekly-business-review (WBR) semantic star schema and its deterministic collectors. Use when asked to collect wbr data, load the warehouse, materialize the warehouse, run monday data readiness, or wire a new WBR data source.",
      "summary": "Materialize and refresh the weekly-business-review (WBR) semantic star schema and its deterministic collectors.",
      "gated": 0,
      "source_path": "plugins/cwc-data-engineering/skills/wbr-warehouse/SKILL.md",
      "created_at": "2026-07-10 18:13:13",
      "cite_as": "https://subagentskills.com/api/skills/skl_wbr_warehouse"
    }
  ]
}