Fulcra CSV

fulcra-csv — Import any CSV into Fulcra

fulcra-csv is a small CLI that maps CSV columns onto Fulcra annotation events. It's the foundation fulcra-media-helpers uses for its generic-csv importer, but it works standalone for any kind of data: weights, moods, expenses, water intake, custom logs.

This skill teaches you (the AI agent) how to import a CSV into the right Fulcra annotation type for a given user data shape.

Where to start — the re-entrancy probes

Before running an import, probe how far this user already got. The states are a prefix of the import flow — authed? → target def exists? → anything-to-import? → already landed? — so enter at the first probe that fails (per the repo's skill-quality pattern, docs/skill-quality-pattern.md). Every state is safely re-enterable: re-running an import produces the same deterministic source_ids, and the importer dedups against a readback of existing source_ids before posting (client-side, in run_import; see "Critical invariant" below), so there is never a penalty for re-probing or re-importing.

All probes pass → the data is already imported; tell the user and point them at Context Web to browse it. A brand-new user fails the first probe. Note the fulcra auth print-access-token command (probe 1, and reused inside the probe-2 curl) belongs to the separate fulcra-api CLI; fulcra-csv has no auth subcommand of its own. Probe 2 hits the Fulcra Life API annotation endpoint (GET /user/v1alpha1/annotation on https://api.fulcradynamics.com, the host the fulcra-api client is what this repo's fulcra-common client resolves from FULCRA_API_BASE; the fulcra-api lib derives the same host from its OIDC audience) directly with curl — the same GET the CLI's annotations_catalog() performs — because neither fulcra nor fulcra-csv exposes a subcommand that lists user-defined annotation definitions.

Three target modes — pick first

Every import targets ONE of three places. Decide first which you're using:

1. User-defined annotation (most common)

You have (or will create) a custom Fulcra annotation definition. The CLI writes events under it. Pass --definition-id <uuid>.

fulcra-csv bootstrap --name "Mood" --description "Mood self-reports" \
  --annotation-type instant --value-type int --tag mood
# → prints the new UUID; save it
fulcra-csv import mood.csv --definition-id <uuid> \
  --annotation-type instant --ts-col timestamp --value-col score \
  --value-type int --note-col context

2. Built-in Fulcra type (BodyMass, HeartRate, etc.) — no definition-id

When the user wants the data to land in Fulcra's native time series (the same place HealthKit imports go), pass --data-type <Name> and SKIP --definition-id. The CLI doesn't append an annotation-def source to the record, so dedup is purely source-id-based and CSV imports can coexist with HealthKit imports of the same kind without duplicating.

fulcra-csv import weights.csv --data-type BodyMass \
  --annotation-type instant --ts-col date --value-col kg --unit kg \
  --tag manual-scale

⚠️ Built-in-type writes assume the receiving schema matches. Check fulcra catalog for known data types. As of this skill's writing, the data-type write API is forthcoming — when shipped, BodyMass/HeartRate/StepCount/etc. are first-class targets.

3. Generic DurationAnnotation / InstantAnnotation — custom definition

Create a simple annotation definition with bootstrap, then import against it with --definition-id. The CLI writes plain DurationAnnotation events (or InstantAnnotation with --annotation-type instant). Useful for "throw a CSV in and forget" cases where the data does not belong to a built-in Fulcra type.

fulcra-csv bootstrap --name "Imported CSV"
# -> save UUID as $CSV_UUID
fulcra-csv import random.csv --definition-id $CSV_UUID

Annotation type — duration vs instant

--annotation-type duration (default) — events have start_time and end_time. For watches, listens, workouts (anything with a span). The CLI uses --end-col or --duration-col; falls back to a 1-second sentinel when neither is given (Fulcra silently drops zero-duration events).

--annotation-type instant — point-in-time. For weights, moods, single readings. recorded_at only has start_time. The --end-col / --duration-col flags error out if you pass them with --annotation-type instant.

Value column — for measurements

If the row has a numeric reading (weight, score, count), use --value-col <name> to lift it into data.value. Coerce with --value-type {float,int,str,bool} (default float). Pair with --unit <string> to add a constant data.unit.

# Body weight in kg
fulcra-csv import weights.csv \
  --data-type BodyMass --annotation-type instant \
  --ts-col date --value-col kg --value-type float --unit kg \
  --tag manual-scale

Empty value cells become None (not zero, not the string "").

Column mapping cheatsheet

Bootstrap a new annotation definition

If the user wants a custom def, mint it first:

fulcra-csv bootstrap \
  --name "Concerts attended" \
  --description "Live music I went to" \
  --tag music --tag tickets
# → prints UUID

For measurement-bearing annotations, set --annotation-type and --value-type:

fulcra-csv bootstrap \
  --name "Resting Heart Rate (manual)" --description "Daily morning RHR" \
  --annotation-type instant --value-type int --unit bpm

Tags are auto-created if they don't exist.

Soft-delete a definition

fulcra-csv soft-delete <uuid> --confirm

⚠️ Fulcra has no per-event delete. Soft-deleting a definition removes the def from the user's account but its events stay visible in queries with their source_id pointing at the deleted def. For a true "reset," soft-delete + create a new def with a different source_id_prefix so future imports namespace cleanly. The fulcra-media sibling has a reset command that wraps this for the four media defs (Watched/Listened/Activity/Read).

Critical invariant: source_ids always include the timestamp

When --source-id-col is set, the column value is mixed into the hash with the timestamp, not used verbatim. Two plays of the same Spotify track at different times produce distinct events. Don't try to "preserve" content IDs in source_ids — they're hashed, idempotency is per-row, not per-content. (Content-level identity belongs in --extra content_fingerprint=fp.)

This is why re-running the same import is always safe — same input rows produce the same source_ids, and before posting each chunk run_import reads back the existing source_ids in that time window and skips the ones already present (client-side dedup, not a server-side guarantee). You don't need to track "have I imported this yet?"

Recipes — common shapes

Body weight (HealthKit-compatible)

date,kg
2026-05-01,82.4
2026-05-02,82.1

fulcra-csv import weights.csv \
  --data-type BodyMass --annotation-type instant \
  --ts-col date --value-col kg --value-type float --unit kg \
  --tag manual-scale

Mood entries

timestamp,score,note
2026-05-01T09:00:00Z,7,morning coffee good
2026-05-01T22:00:00Z,5,long day

fulcra-csv bootstrap --name "Mood" --annotation-type instant \
  --value-type int --tag mood
# → save UUID as $MOOD_UUID

fulcra-csv import mood.csv --definition-id $MOOD_UUID \
  --annotation-type instant --ts-col timestamp \
  --value-col score --value-type int --note-col note

Expenses with category tags

date,amount,merchant,category
2026-05-01,12.50,Blue Bottle,coffee
2026-05-01,38.00,Whole Foods,groceries

fulcra-csv bootstrap --name "Expenses" --annotation-type instant \
  --value-type float --unit usd --tag finance
# → save UUID

fulcra-csv import expenses.csv --definition-id $UUID \
  --annotation-type instant --ts-col date --value-col amount \
  --value-type float --unit usd \
  --title-col merchant --tag-col category

Pipedream / IFTTT play log (music)

ts,track,artist,track_id,url
2026-05-01T09:00:00Z,Reelin' In The Years,Steely Dan,1I7zHEdDx8Ny5RxzYPqsU2,https://...

fulcra-csv import plays.csv --definition-id $LISTENED_UUID \
  --ts-col ts --title-col track --subtitle-col artist \
  --source-id-col track_id --tag spotify --extra url=spotify_url

Sleep durations

start,end,quality
2026-05-01T23:00:00Z,2026-05-02T07:30:00Z,8

fulcra-csv bootstrap --name "Sleep" --tag sleep
fulcra-csv import sleep.csv --definition-id $UUID \
  --ts-col start --end-col end --value-col quality --value-type int

Export — round-tripping annotations back to CSV

fulcra-csv export is the inverse of import. Reach for it when the user wants to:

• Round-trip a CSV — re-export what an import landed to verify columns/values look right, or to hand the data to another tool that expects CSV.
• Audit a recent import — pull the last day/week of a definition and eyeball it instead of running ad-hoc API queries.
• Slice for a downstream tool — pull a configurable subset of fields (well-known + data.<key> + external_ids.<key>) into a tidy CSV, optionally with epoch timestamps.

It is NOT a sync/backfill mechanism — it's a one-shot read. If the user wants ongoing sync, point them at a scheduled job that runs export on a window.

Target — same model as import

Pass ONE of:

• --definition-id <uuid> — scope to a user-defined annotation. The CLI fetches the underlying data type (default DurationAnnotation) and filters records whose sources array references the target def. This mirrors the importer's dedup-readback, so what you see in export is what import would dedup against.
• --data-type <Name> — pull a built-in time series (BodyMass, HeartRate, DurationAnnotation, etc.).

--start is required (ISO-8601 or relative — "1 week ago", "yesterday"). --end defaults to now.

Column model — --columns col1,col2,...

Default: start_time,end_time,tag,note,value. Each entry resolves against the record in one of three ways:

source_id is special: it returns the first non-definition source (the per-row dedup key), so you can round-trip dedup hashes back through.

Other knobs

• --date-format iso|epoch|local (default iso, always UTC with Z suffix). epoch writes integer seconds. local honors --tz.
• --tz <iana> — used for parsing relative --start/--end and for --date-format local.
• --out <path> — file output. Omit for stdout.

CSV-injection guard

By default, cells starting with = + - @ \t \r are prefixed with a single quote (') so Excel/Sheets/Numbers don't interpret them as formulas. This is OWASP-grade defense-in-depth and on by default. The library exposes ExportOptions.guard_formulas=False for callers feeding CSV into a downstream parser that doesn't need it; the CLI does NOT expose a flag — keep it on for spreadsheets. Booleans render as lowercase "true"/"false" so they round-trip through coerce_value.

Worked example — audit yesterday's mood imports

The user just ran fulcra-csv import mood.csv --definition-id $MOOD_UUID ... and wants to see what landed:

fulcra-csv export \
  --definition-id $MOOD_UUID \
  --start yesterday \
  --columns start_time,value,note,tag,source_id \
  --date-format local --tz America/New_York

Stdout is a CSV with five columns, timestamps in the user's wall-clock TZ, plus source_id so they can match rows against the importer's deterministic hashes.

Pitfalls

• Export does NOT trigger a re-fetch from the source. It reads what Fulcra has. If a recent import is still propagating through ingest, you may see fewer rows than you posted — wait a minute and re-run.
• --definition-id requires the underlying data_type to match. The export defaults to DurationAnnotation; if the user bootstrapped an instant annotation, pass --data-type InstantAnnotation alongside --definition-id.
• Don't promise byte-identical round-trips. Booleans go to lowercase, lists/dicts JSON-encode to single-line, and the formula-guard ' prefix is a real cell value. The data round-trips through coerce_value, but the CSV bytes differ from the original import file.

Periodic ingestion

CSV imports are inherently one-shot ("import this file"), not incremental. There's no watermark layer (that lives in fulcra-media for its API-poll importers).

For agent-driven recurring sync of a CSV that grows over time (e.g. a Pipedream workflow appending rows to a Drive sheet):

1. Download the latest CSV to a temp file.
2. Run fulcra-csv import <tempfile> ... — re-imports are idempotent.
3. Skipped rows are silent (the importer's client-side source-id readback dedup). Posted rows are the delta.

The --dry-run flag is useful for a cheap "any new rows?" check (it parses and prints the first 5, doesn't ingest).

Fulcra Life API auth

Same as fulcra-media: shells out to fulcra auth print-access-token from the fulcra-api package, or honors FULCRA_ACCESS_TOKEN=... for non-interactive contexts.

Common failure patterns

"No definition_id found" exit

You forgot --definition-id <uuid> AND didn't set --data-type AND --annotation-type is duration. Either bootstrap a def, or pass --data-type DurationAnnotation explicitly.

Events post but don't appear in fulcra get-records

Fulcra's ingest-to-query indexing lag (seconds to minutes for bulk imports). Re-run shouldn't help — wait a few minutes and re-query.

Same content keeps producing duplicate events on each import

You're probably using --source-id-col for a per-row id (database row uuid) and expecting per-row uniqueness. The CLI ALWAYS mixes timestamp into the hash. For truly stable per-row identity, ensure the (timestamp, id) pair is stable across imports. If the row's timestamp changes between exports, you'll get distinct source_ids — that's intentional but maybe not what you want.

"ValueError: unparseable timestamp"

dateparser can't parse the timestamp column. Try --tz America/New_York if rows are local-time-no-tz. Worst case, pre-process the CSV to ISO 8601.

Architectural references (for the curious agent)

• fulcra_csv.events — GenericEvent, ColumnMap, coerce_value
• fulcra_csv.parser — parse_csv(path, column_map, tz, annotation_type)
• fulcra_csv.fulcra — FulcraClient (auth, ingest, dedup readback, per-chunk verification)
• fulcra_csv.confidence — apply_cluster_policy, find_low_conf_twins, apply_twin_decisions (used by fulcra-media's Trakt importer, but available standalone)
• fulcra_csv.cli — click-based CLI entry

Don't

• Don't bake --definition-id into a script as a constant — the user might soft-delete + re-bootstrap, changing the UUID. Always read it from a fresh fulcra-csv bootstrap or from the user's state.
• Don't try to dedup at the agent layer — the importer's run_import already reads back existing source_ids per window and skips duplicates before posting. Just re-run the import; already-landed rows are skipped silently.
• Don't write data.note from a column the user didn't intend to expose — --note-col is opt-in.
• Don't pass --value-type bool for "0/1" rows unless that's truly the schema — bool collapses any non-truthy string to false; for diary-style yes/no events use int (0/1).
• Don't print the user's access token to stdout — neither in error envelopes nor in success messages. The CLI scrubs URL params in exception messages, but raw token blobs are sensitive.