Workout Parse — Data Model

One table, in libs/db/src/lib/schema/ai-parse.ts. One row per transform attempt — this is where “freeform text retained as source of truth” lives (D5), and the audit trail for cost, timings, and the commit-time correction delta.

`ai_parse_jobs`

Column	Type	Notes
`id`	`uuid` PK
`organization_id`	`uuid` FK → `organizations.id` (cascade), not null	Scope.
`user_id`	`uuid` FK → `users.id` (cascade), not null	Requesting coach (internal `users.id`, same referencing style as `ai_conversations`).
`source`	`varchar(20)` not null	`'paste'` \| `'freeform_form'` (D12 entry point).
`status`	`varchar(30)` not null, default `'draft'`	Lifecycle below.
`input_text`	`text` not null	Normalized (LF) source text — every span in `draft` is an offset into this; becomes the committed workout’s description by default.
`language`	`varchar(10)` nullable	BCP-47-ish primary tag from extraction (`'en'`, `'he'`, …).
`multi_workout_detected`	`boolean` not null, default false	D4.
`workout_count`	`integer` nullable	Extraction’s count of distinct workouts in the paste.
`draft`	`jsonb` nullable	The persisted `ParseDraft` IR (`parse-draft.schema.ts`). Null on jobs that failed before assembly.
`auto_payload`	`jsonb` nullable	`draftToSectionsPayload(draft)` at parse time — auto-resolved movements only. The baseline for the correction delta.
`final_payload`	`jsonb` nullable	The committed `set_sections`-shaped payload (post-adjustment). Set on commit.
`delta`	`jsonb` nullable	`PayloadDelta` from `diffSectionsPayload(auto_payload, final_payload)` — the headline product metric (zeroEdit, editDistance, swap/add/remove counts). Set on commit.
`workout_id`	`uuid` FK → `workouts.id` (set null), nullable	Input `workoutId` for the “Structure this” flow; set on commit for the paste flow.
`model`	`varchar(64)` nullable	e.g. `claude-sonnet-4-5`.
`input_tokens` / `output_tokens`	`integer` nullable	Extraction usage.
`cost_usd_micros`	`bigint` nullable	Same math as Spotter (`computeCostUsdMicros`, `apps/api/src/ai/cost.util.ts`); also upserted into `ai_usage_daily` via `AgentCostTracker.recordTurn`.
`duration_ms`	`integer` nullable	Wall clock of the whole parse.
`stage_timings`	`jsonb` nullable	`{ extractionMs, resolutionMs, totalMs }`.
`error_code`	`varchar(64)` nullable	`parse_english_only` / `parse_failed` on terminal-failure rows.
`created_at` / `updated_at`	`timestamp tz` not null	House pattern.

Indexes:

ai_parse_jobs_org_created_idx — (organization_id, created_at DESC) (org-scoped listing / lookup).
ai_parse_jobs_workout_idx — partial on (workout_id) WHERE workout_id IS NOT NULL.

Relations (aiParseJobsRelations): job → organization, user, workout.

Status lifecycle

The row is inserted first (status: 'draft') before the LLM call, so even a crashed extraction leaves an audit row. Terminal states are never re-entered — commit/discard require status = 'draft' and 409 otherwise.


                    POST /workouts/parse
                            │
                  insert row (status: draft)
                            │
                     LLM extraction
                            │
        ┌───────────────────┼──────────────────────┐
        │                   │                      │
  language ≠ en       pipeline error          success
        │                   │                      │
        ▼                   ▼                      ▼
rejected_non_english      failed          draft (IR + auto_payload persisted)
  (422, cost still     (500, error_code)          │
   metered)                              ┌────────┴────────┐
                                         │                 │
                                   POST /commit      POST /discard
                                         │                 │
                                         ▼                 ▼
                                     committed         discarded
                              (final_payload + delta
                               + workout_id recorded)

Why the payload triple (`draft` / `auto_payload` / `final_payload`)

draft is the client-facing IR — spans, candidates, confidence. It’s what the preview renders and what a future re-parse feature would compare against.
auto_payload freezes “what the machine got right on its own” at parse time.
final_payload is what the coach actually shipped. delta between the two is the correction distance — computed server-side at commit through the exact same diffSectionsPayload the eval harness uses, so telemetry and eval numbers are on one scale.

Multi-org isolation & PII

All reads go through ParseJobsRepository.findById(orgId, jobId) — org-scoped; cross-org job ids 404.
input_text and draft contain coach-authored content and are never emitted to observability (scrub() blocklist in parse-observability.service.ts drops text/mentions/titles/payloads — counts and enums only).
Cascade on org/user delete; workout_id is set null so deleting a workout keeps the audit row.

Retention posture

No automatic deletion; rows grow with usage (one per parse attempt, including rejected/failed ones). Payload jsonb is bounded by the 10,000-char input cap and the section/movement caps. Revisit alongside the Spotter retention backlog item if volume warrants.