Cyclone

A self-hosted EDI claims management suite for a single billing office. Parses 837P professional claims and 835 ERA remittances (X12 005010X222A1 and 005010X221A1), with a local-only FastAPI backend and a React UI for browsing, filtering, and inspecting the parsed data.

Local-only on purpose: binds to 127.0.0.1, no auth, no internet exposure. Built for one operator, one machine, one trading partner (Colorado Medicaid, currently).

Install

# Backend (Python 3.11+)
cd backend
python -m venv .venv
.venv/bin/pip install -e '.[dev]'

# Frontend (Node 20+)
cd ..
npm install

Dev

Two terminals:

# Terminal 1 — backend
cd backend
.venv/bin/python -m cyclone serve
# (defaults to 127.0.0.1:8000; override with CYCLONE_PORT=...; reload with CYCLONE_RELOAD=1)

# Terminal 2 — frontend
npm run dev
# (Vite on http://localhost:5173)

Then open http://localhost:5173. Drop an .txt 837P or 835 file on the Upload page; navigate to Claims, Remittances, Providers, or Activity to see the parsed data.

The frontend reads its backend URL from VITE_API_BASE_URL (default empty). Create a .env.local at the repo root with:

VITE_API_BASE_URL=http://127.0.0.1:8000

Without that, the UI falls back to its in-memory sample store via the existing data adapter (parses are disabled).

Test

# Backend
cd backend && .venv/bin/pytest

# Frontend type-check + build
npm run typecheck
npm run build

# Frontend unit tests
npm test

Live updates

The Claims, Remittances, and Activity pages stay current without manual refresh. The backend publishes an internal event on every store write, the page opens a streaming HTTP connection to the matching /api/<resource>/stream endpoint, and new rows are appended to the table the moment they hit the database.

Wire format

Each stream endpoint emits newline-delimited JSON. The first batch is the snapshot of currently-known rows; after that comes snapshot_end with the count, then the live events.

{"type":"item","data":{"id":"CLM-1", "...":"..."}}
{"type":"item","data":{"id":"CLM-2", "...":"..."}}
{"type":"snapshot_end","data":{"count":2}}
{"type":"item","data":{"id":"CLM-3", "...":"..."}}   ← live
{"type":"heartbeat","data":{"ts":"2026-06-20T23:17:09Z"}}   ← idle keep-alive

Lines are {"type": ..., "data": ...}; known types are item, snapshot_end, heartbeat, and (rare) item_dropped / error. Heartbeats keep the connection alive when nothing is happening — clients flip to stalled after 30s of total silence (heartbeat or otherwise) and surface a ↻ Reconnect button.

Endpoints

Method	Path	Subscribes to	Default sort
GET	/api/claims/stream	claim_written	-submission_date
GET	/api/remittances/stream	remittance_written	-received_date
GET	/api/activity/stream	activity_recorded	-timestamp (limit 50)

All three accept the same query params as their non-streaming counterparts (status, payer, date_from, …) so a frontend can swap a one-shot fetch for a tail with no URL surgery. Responses are Content-Type: application/x-ndjson.

Status pill

The pill in each toolbar reflects the current connection state:

Status	Badge variant	What it means
live	success	snapshot received, listening for new events
connecting	warning	opening the stream (initial mount)
reconnecting	warning	previous attempt failed, backing off before retry
stalled	destructive	no event (including heartbeat) for 30s — click ↻ Reconnect
error	destructive	stream errored — click ↻ Reconnect
closed	destructive	page unmounted

The reconnect button is only shown on stalled and error. The backoff schedule on error is 1s → 2s → 4s → 8s → 16s → 30s capped.

Knobs

Env var	Default	Effect
CYCLONE_TAIL_HEARTBEAT_S	15	Idle heartbeat interval (seconds). Lower it for tests.

Inbox

/inbox is the working surface. Four lanes, dark by default (Ticker Tape aesthetic):

• Rejected — claims whose 999 set-level response was R or E. Re-submit in bulk.
• Candidates — remits whose CLP-claim-id didn't match exactly; each one shows its top scored claim. One-click manual match or dismiss.
• Unmatched — claims still waiting for a remit, and remits with no candidates above the threshold.
• Done today — terminal state transitions in the last 24 hours.

The page subscribes to the SP5 claim and remittance tail streams (/api/claims/stream and /api/remittances/stream); a new item event refetches the lane payload (debounced ~250ms). When a 999 parses and rejects claims, the inbox reflects the new Rejected rows within a fraction of a second.

Inbox endpoints

Method	Path	Notes
GET	/api/inbox/lanes	All four lanes in one call.
POST	/api/inbox/candidates/{remit_id}/match	Manual match. 409 if the claim state moved out from under us.
POST	/api/inbox/candidates/dismiss	{pairs: [{claim_id, remit_id}]}. Session-scoped.
POST	/api/inbox/rejected/resubmit	{claim_ids: [...]}. 200 with conflicts for non-rejected.
GET	/api/inbox/export.csv?lane=<lane>	Streams CSV of the lane's rows.

Outbound 837 Serializer

The same canonical ClaimOutput model that powers 837 ingestion also drives outbound 837P regeneration. Operators can take a claim that's been parsed, edited, or rejected and emit a byte-faithful X12 837P file — closing the resubmit loop without leaving the app.

Two surfaces:

• Single claim → .x12 download from the ClaimDrawer header. Click the Download 837 icon, get a claim-{id}.x12 attachment. The file is a complete ISA/GS/ST envelope + hierarchy + CLM/SV1 lines, regenerated from the canonical ClaimOutput fields (not by patching raw_segments).
• Multi-claim → .zip bundle from the Inbox rejected lane. Select N>1 rejected claims, hit Resubmit, choose Resubmit + Download in the confirm modal, and the backend hands you resubmit-{N}-claims.zip with one .x12 per successfully resubmitted claim. Each file gets a unique ISA13/GS06 so back-to-back claims don't collide on control numbers.

Design spec at docs/superpowers/specs/2026-06-20-cyclone-serialize-837-design.md. Implementation note: the spec proposed a hybrid "patch raw_segments" approach, but raw_segments only captures post-CLM segments (CLM, REF*G1, HI, LX, SV1 pairs) — the envelope and hierarchy segments (NM1, N3, N4, HL, PRV, SBR, DMG) are not captured, so a pass-through serializer cannot regenerate them byte-for-byte. The shipped implementation rebuilds the entire file from canonical fields ("Approach A" in the plan). The round-trip guarantee still holds: all 113 prodfiles in docs/prodfiles/claims/*.x12 parse → serialize → parse back to the same claim id.

Where to find it

• Backend serializer: backend/src/cyclone/parsers/serialize_837.py
• Single-claim download API: GET /api/claims/{id}/serialize-837
• Bundle API: POST /api/inbox/rejected/resubmit?download=true
• Frontend helper: src/lib/api.ts:serializeClaim837, src/lib/inbox-api.ts:resubmitRejectedWithDownload
• Browser download utility: src/lib/download.ts (downloadTextFile, downloadBlob)
• UI:
  • src/components/ClaimDrawer/ClaimDrawerHeader.tsx — Download icon
  • src/pages/Inbox.tsx — Resubmit bundle modal + progress overlay

Per-Line Adjustment Audit

Every 835 CAS segment is tied back to the specific 837 service line it adjudicates, instead of being aggregated to the claim level. The match runs eagerly at 835 ingest time so the audit trail is stable across re-reads (a re-migration is required if the algorithm changes).

Where to find it

• ClaimDrawer → "Line Reconciliation" tab — side-by-side view of the 837 SV1 lines and 835 SVC composites with per-line CAS reasons.
• ClaimDrawer → ServiceLinesTable — every billed line now shows its paid amount and adjustment sum on the right.
• RemitDrawer → CasAdjustmentsPanel — each CAS row carries a proc · #N reference; claim-level CAS clusters separately.
• ClaimDrawer → MatchedRemitCard — a lines: N/M matched badge lights up amber when at least one line is unmatched.

Match criteria (strict)

A 835 SVC composite matches a 837 SV1 line iff all four criteria align:

Field	Rule
Procedure	Exact match (case-insensitive, uppercased).
Modifiers	Set-equal (order-independent).
Service date	Exact match (both null counts as a match).
Units	Exact match (both null counts as a match).

Unmatched lines surface as a soft warning — the claim still posts to PAID/PARTIAL/RECEIVED, but the Inbox shows the unmatched count and the drawer surfaces a per-line "no 837 line matched" note.

Line-level endpoints

Method	Path	Notes
GET	/api/claims/{claim_id}/line-reconciliation	Dedicated view for the drawer tab.
GET	/api/claims/{claim_id}	Now includes a slim lineReconciliation[].
GET	/api/remittances/{remittance_id}	Now includes serviceLinePayments[] + claimLevelAdjustments[].
GET	/api/inbox/lanes	matched_remittance payload gains matched_lines + total_lines.

Scoring

4-field weighted (sum = 100):

Field	Weight	Rule
Patient control number	40	Exact match, normalized (case + leading zeros).
Service date	25	Linear decay over ±3 days.
Charge amount	20	Linear decay over ±10%.
Provider NPI	15	Exact match. Missing → 0.

Tiers: strong (≥75, full opacity, Match enabled), weak (50–74, dimmed), hidden (<50, not surfaced).

Persistence

Parsed batches, claims, remittances, matches, and activity events are stored in a SQLite file at ~/.local/share/cyclone/cyclone.db by default. The directory is auto-created on first run.

To use a different location, set CYCLONE_DB_URL:

export CYCLONE_DB_URL=sqlite:///path/to/cyclone.db
# or
export CYCLONE_DB_URL=postgresql://user:pass@host:5432/cyclone

Backup

sqlite3 ~/.local/share/cyclone/cyclone.db ".backup /path/to/backup.db"

This is safe to run while the backend is running (uses SQLite's online backup API).

Project layout

.
├── backend/
│   ├── src/cyclone/
│   │   ├── api.py             # FastAPI app, GET + parse routes, /api/{resource}/stream
│   │   ├── pubsub.py          # in-process EventBus (drop-oldest, per-kind fan-out)
│   │   ├── store.py           # InMemoryStore, mappers, publish-on-write
│   │   ├── __main__.py        # `python -m cyclone serve`
│   │   ├── cli.py             # click CLI
│   │   └── parsers/           # X12 tokenizer, models, validator, writers
│   └── tests/
│       ├── fixtures/          # co_medicaid_*.txt, minimal_*.txt
│       ├── test_api.py        # parse-837/835 round-trip
│       ├── test_api_gets.py   # 6 GET endpoints
│       ├── test_store.py      # store + mappers + iterators
│       ├── test_api_streaming.py
│       ├── test_api_stream_live.py  # 3 live-tail endpoints + disconnect cleanup
│       ├── test_pubsub.py     # EventBus + subscribe/unsubscribe
│       └── test_api_parse_persists.py
├── src/                       # React + Vite + TypeScript UI
│   ├── components/
│   │   ├── ui/                # Skeleton, EmptyState, ErrorState, FilterChips, Pagination, …
│   │   └── TailStatusPill.tsx # live-tail status badge + reconnect button
│   ├── pages/                 # Claims, Remittances, Providers, Activity, Upload
│   ├── hooks/                 # useBatches, useClaims, useRemittances, useProviders, useActivity, useParse
│   │                          # + useTailStream, useMergedTail (live tail)
│   ├── lib/                   # api.ts (6 GET + parse837/parse835/health), format.ts, utils.ts
│   │                          # + tail-stream.ts (NDJSON parser)
│   ├── store/                 # zustand sample-data + parsed-batches store
│   │                          # + tail-store.ts (FIFO-capped live tail slices)
│   └── types/                 # shared TS types
├── docs/
│   ├── reference/             # condensed 837P/835/X12/CO Medicaid notes
│   └── superpowers/plans/     # implementation plan
├── tailwind.config.js         # shimmer, scan, row-flash keyframes
└── package.json

Roadmap

Sub-projects 2 through 8 are shipped. Next up:

• Sub-project 8 (shipped) — Outbound 837P serializer. Closes the resubmit loop: rejected claims can be regenerated back to an X12 837P file (single download from the claim drawer, or ZIP bundle from the inbox rejected lane). See the "Outbound 837 Serializer" section below for details.
• Sub-project 3 (shipped) — More 837P/835 features.
  • 837P validation rules: R034 enforces REF*G1 on frequency-code 7/8 claims; R035 enforces BHT06 transaction-type-code is in the allowed set per payer config.
  • 835 CAS deep-parsing: every CAS adjustment now surfaces with its CARC reason code + label (e.g. CO-29: The time limit for filing has expired), surfaced via GET /api/remittances/{id} and rendered as an expansion row on the Remittances page.
  • 999 ACK transaction set: full inbound parser + outbound serializer. Auto-generated on 837 ingest when ?ack=true is passed to POST /api/parse-837. Persisted to the acks table, browsable on a new /acks page, and downloadable as the regenerated raw 999 text.
  • 270/271 eligibility (API-only): POST /api/eligibility/request builds a 270 from a JSON payload (subscriber, provider, payer, service type); POST /api/eligibility/parse-271 ingests the response and returns structured coverage_benefits.
  • TA1 interchange ACK: parser + persistence + API for the envelope-level TA1 acknowledgment (separate from the 999 transaction-set ACK).
• Sub-project 4 (shipped) — Frontend features.
  • Per-claim detail drawer: click any row on the Claims page to open a side-panel drawer with the full claim context — state + amount header, validation panel, service lines, diagnoses, parties (billing provider, subscriber, payer), raw X12 segments, matched remittance summary, and a vertical state-history timeline. URL is synced (?claim=...) so links and back-button restore the drawer. Keyboard nav: j/k move between claims, esc closes, ? opens the cheatsheet overlay. Skeleton + error + not-found (404) states are all distinct. Powered by a new backend endpoint GET /api/claims/{claim_id} that returns the full drawer payload in a single round-trip.
  • Remit drawer: mirror of the claim drawer for remittances — payer + provider header, financial summary, claim-payments table, CAS adjustments panel, parties grid, raw X12 segments. URL is synced (?remit=...).
  • Batch diff view: side-by-side comparison of two batches (added / removed / changed claims) reachable from the Batches page. Powers the BatchDetail and BatchDiff views.
  • Global Cmd-K search: cross-resource search across claims, remittances, and activity; opens from anywhere via the keyboard shortcut.
  • CSV export: per-page export buttons (claims, remittances, lanes) that stream CSV directly from the backend.
  • Accessibility + print styles: keyboard navigation focus rings, aria roles, and @media print stylesheets for paper-friendly output.
• Sub-project 5 (shipped) — Live updates.
  • The Claims, Remittances, and Activity pages stream new rows in real time over Server-Sent Events encoded as newline-delimited JSON. The backend publishes a claim_written / remittance_written / activity_recorded event on every store write; the frontend opens an HTTP stream and dispatches each event into a per-resource Zustand store that the page reads on top of its base query results.
  • A small status pill in each toolbar surfaces the connection state — live / connecting / reconnecting / stalled / error / closed — and offers a manual ↻ Reconnect button on stalled or error. Stale connections (no event for 30s, heartbeat included) flip to stalled automatically; the back-off ladder on errors is 1s → 2s → 4s → 8s → 16s → 30s capped. See the "Live updates" section below for details.
• Sub-project 6 (shipped) — Inbox workflow automation.
  • Ticker-Tape inbox (/inbox): four lanes ordered by urgency — Rejected (claims whose 999 rejected them), Candidates (remits that didn't auto-match a claim), Unmatched (claims still waiting for a remit), and Done today (terminal transitions in the last 24 hours). The page subscribes to the claim and remittance tail streams so new rejected claims land in the Rejected lane within a fraction of a second.
  • 4-field weighted scoring: every candidate remit is scored on patient control number (40), service date (25), charge amount (20), and provider NPI (15). Tiers: strong (≥75, Match enabled), weak (50–74, dimmed), hidden (<50).
  • Bulk actions: multi-select rows in any lane to dismiss candidates, resubmit rejected claims, or export the lane as CSV in a single round-trip.
• Sub-project 7 (shipped) — Per-line adjustment audit.
  • Every 835 CAS segment is tied back to the specific 837 service line it adjudicates, instead of being aggregated to the claim level. The match runs eagerly at 835 ingest time so the audit trail is stable across re-reads.
  • New endpoints surface the per-line projection: GET /api/claims/{id}/line-reconciliation for the dedicated tab in the ClaimDrawer; the existing claim + remit detail endpoints gained slim lineReconciliation[] / serviceLinePayments[] / claimLevelAdjustments[] fields; the inbox lanes include matched_lines / total_lines badges on matched remittances.
  • The ClaimDrawer gains a Line Reconciliation tab (alongside Details) that pairs every 837 SV1 line with its 835 SVC composite and surfaces per-line CAS reasons + unmatched-line warnings.
  • See the "Per-Line Adjustment Audit" section below for details.

SP3 endpoints

• POST /api/parse-837?ack=true — existing 837 parse, plus optional auto-generated 999 ACK persisted alongside the batch.
• POST /api/parse-999 — parse an inbound 999 ACK and persist it.
• GET /api/acks — list ACKs.
• GET /api/acks/{id} — ACK detail, including the regenerated raw_999_text.
• POST /api/eligibility/request — build a 270 from JSON.
• POST /api/eligibility/parse-271 — ingest a 271 and return parsed coverage benefits.

The UI gains a new Acks page (sidebar entry) that lists persisted ACKs and lets you download the regenerated 999 text.

SP4 endpoints (claim drawer)

• GET /api/claims/{claim_id} — full claim context for the drawer payload: header fields, diagnoses, service lines, parties (billing provider / subscriber / payer), validation result, raw X12 segments, matched remittance summary (or null), and the claim's state history (ordered newest-first). 404 with a structured { "error": "Not found", "detail": "Claim {id} not found" } body when the claim doesn't exist — distinct from a transient fetch failure so the UI can render a dedicated not-found state.

SP5 endpoints (live updates)

• GET /api/claims/stream — NDJSON stream of claim events. Emits a snapshot (filtered by the same query params as GET /api/claims), then snapshot_end, then live claim_written events as they arrive, with a 15s idle heartbeat.
• GET /api/remittances/stream — same shape for remittances (subscribes to remittance_written, default sort -received_date).
• GET /api/activity/stream — same shape for activity events (subscribes to activity_recorded, default limit=50).

SP6 endpoints (inbox)

• GET /api/inbox/lanes — all four lanes (rejected, candidates, unmatched, done_today) in a single round-trip, with row-level scoring and matched-remit context.
• POST /api/inbox/candidates/{remit_id}/match — manual match of a candidate remit to one of its scored claims; 409 if the claim state moved out from under us.
• POST /api/inbox/candidates/dismiss — {pairs: [{claim_id, remit_id}]}. Session-scoped, no DB persistence.
• POST /api/inbox/rejected/resubmit — {claim_ids: [...]}. 200 with conflicts for non-rejected claims in the batch.
• GET /api/inbox/export.csv?lane=<lane> — streams CSV of the lane's rows.

SP7 endpoints (per-line reconciliation)

• GET /api/claims/{id}/line-reconciliation — dedicated view for the ClaimDrawer's Line Reconciliation tab; returns the slim per-line projection (summary + lines[]).
• GET /api/claims/{id} — now includes a slim lineReconciliation[] so the ServiceLinesTable can show paid + adjustments columns without a second fetch.
• GET /api/remittances/{id} — now includes serviceLinePayments[]
  • claimLevelAdjustments[]; the CAS panel renders per-line + claim level separately.
• GET /api/inbox/lanes — matched_remittance payload gained matched_lines + total_lines so the MatchedRemitCard can badge partial matches.

SP8 endpoints (outbound 837P)

• GET /api/claims/{id}/serialize-837 — regenerate the persisted claim as a byte-faithful X12 837P file (text/x12 body, Content-Disposition: attachment; filename="claim-{id}.x12"). 404 if the claim doesn't exist, 422 if the stored raw_json can't be revalidated as a ClaimOutput. Drives the Download 837 icon in the ClaimDrawer header.
• POST /api/inbox/rejected/resubmit?download=true — same mass-resubmit endpoint as SP6, but the response is a ZIP archive (application/zip) containing one claim-{id}.x12 per successfully resubmitted claim. Each file uses serialize_837_for_resubmit so back-to-back files in the bundle get unique ISA13/GS06 control numbers (back-to-back resubmits would otherwise all share 000000001). Conflicts and missing ids are deliberately omitted from the ZIP — the user already saw them in the JSON response on prior calls. Per-claim regenerate failures (rare — usually means raw_json is corrupted for one claim) are surfaced via the X-Cyclone-Serialize-Errors response header (JSON-encoded array) so the UI can show "10 resubmitted, 2 couldn't be regenerated" without parsing the binary. Drives the Resubmit + Download button in the Inbox rejected-lane BulkBar (N>1 modal prompt).

License

No license file yet; this is internal-use software. Add a LICENSE file when one is decided.