M3 — task breakdown

Role: breakdown. Ordered task list for v0.5 / M3. Cites requirement IDs; defines none.

Objectives: roadmap.md in this folder. Behavior and acceptance: docs/capabilities/CREATE.md. Scope and IDs: docs/MILESTONES.md. Decisions live in docs/contracts/, not here.

TDD every slice: the failing test precedes its implementation; never parallelize RED and GREEN of the same scenario. Execute slices serially (one local GPU; see AGENTS.md). Frontend ships with the slice that first needs the screen, not as a trailing blob.

Hard dependencies

S0  full schema.sql + factory + health
 └─ S1  crypto + Settings
      ├─ S2  slabs CRUD          (needs S0 tables; tests seed fixture taxonomy rows)
      │    └─ S3  photos         (needs slab_id)
      └─ S4  Woo + taxonomy      (needs S1 credentials only — not S2/S3)
           │
           S5  pricing + title/short-desc + ready     ← S2 AND S4
            ├─ S6  publish                            ← S3 + S4 + S5
            └─ S7  inference last                     ← S2 + S3 + S4 + S5; not S6

Default serial order for Atlas: S0 → S1 → S2 → S3 → S4 → S5 → S6 → S7 → Exit.

S4 does not wait on S2/S3. If S2 is blocked after S1, start S4. S7 does not wait on S6 (METADATA-75).

Slices

Slice	Delivers	M3 IDs	Hard deps
S0	factory; full `schema.sql` from `docs/contracts/DATA-MODEL.md` (including empty taxonomy tables); config/startup guards; health/meta; JSON logs; Dockerfile	APP-30, APP-45, APP-55, APP-65, SETTINGS-10, SETTINGS-25, OPS-25	—
S1	AES-GCM crypto + Settings API/UI (masked, write-only secrets). Missing key refuses startup (S0); this slice encrypts/decrypts. Decrypt failure → owner re-enters credentials.	SETTINGS-5/15/20/25/30/75/85/90, QUALITY-125	S0
S2	slabs CRUD, optimistic concurrency, soft delete. Fixture `woo_categories` rows for FK tests. Frontend: API client + slab list.	APP-5/15/20/35/40, REVIEW-45	S0
S3	photos (multipart → BLOBs, original/processed, 410 after purge). Frontend: photo attach on the draft.	CAPTURE-95/105/130, PUBLISH-40	S2
S4	Woo client (application-password Basic) + taxonomy sync + test-woo. Fills the S0 taxonomy tables. Owns `taxonomy_anchor` / `taxonomy_last_sync_at`. Frontend: Settings test-woo.	METADATA-5/10/15/20/30, SETTINGS-40/50/51/55/60/80, QUALITY-35	S1
S5	species `$/bdft` rates; deterministic title + short_description; server-derived `ready`. Inference off: owner types `description` by hand (`METADATA-70`). Frontend: Review screen.	REVIEW-5/10/15/20/25/30/35/40/50, METADATA-70/80/85/90, SETTINGS-35, QUALITY-90/95	S2, S4
S6	publish sequence, UAT draft, duplicate/stale/partial failure; upload processed PNGs; publish-time taxonomy refresh when last sync is stale (`SETTINGS-60`). Frontend: Publish screen.	PUBLISH-5/10/15/20/25/30/31/35/36/37/38/40/45/50/55/60, SETTINGS-60	S3, S4, S5
S7	inference last: Call 1, Call 2 (one static seed; calculated title/short_description injected into the Call 2 prompt), prompt revision REST. Never blocks create. Frontend: Generate text control.	METADATA-35/40/45/50/55/60/65/66/75/95/100/105/106/110/115, SETTINGS-45/70	S2, S3, S4, S5
Exit	inference-off suite, Woo-mock integration, Compose UI E2E, real `SLAB-UAT-*` draft	QUALITY-20/35/50/105/110/120	S0–S7

IDs split on purpose (do not “move” them to one slice):

ID	Slices	Split
SETTINGS-25	S0 + S1	S0 refuses startup without the key; S1 implements AES-GCM
SETTINGS-60	S4 + S6	S4 owns anchor + cache rebuild; S6 refreshes on publish when stale
PUBLISH-40	S3 + S6	S3 stores PNG BLOBs; S6 uploads them to Woo
QUALITY-35	S4 + Exit	S4 writes the Woo-mock tests; Exit re-runs them

Parked work items

The seven per-leaf Call 2 prose templates are not M3. S7 ships the single static call2 seed in docs/contracts/PROMPTS.md §6. The model keys off the already-calculated title, short_description, fact block, photos, and metadata. Per-leaf templates are later.
Specialty cuts is out of M3: no title pattern, no prose template, no Call 2 prompt. Owner will not inventory that leaf in M3.
S7 ships the prompt revision REST interface (METADATA-106) so UAT prompt tuning is an API call on the trusted LAN. There is no prompt editing UI.
Authentication for the prompt revision interface is later (METADATA-107), not M3.

Frontend layout

New code stays outside the Capture union: frontend/src/api/*, frontend/src/create/*.

With slice	Screen
S1	Settings (store creds, inference flags, secret masking)
S2	Slab list / draft resume
S5	Review (ready set, primary wood category, price, title/short_description)
S6	Publish
S7	Generate text

Verification

Slice bar: docs/contracts/TEST-SPEC.md. Docs gate: python tools/check_requirements.py. Exit bar: docs/MILESTONES.md and docs/capabilities/CREATE.md.

Not yet validated

Long-horizon multi-step agentic loops (more than one or two tool round-trips).
M3 implementation has not started.

5.1 KiB Raw Blame History Unescape Escape