SlabUploader/AGENTS.md

AGENTS.md — SlabUploader

Working contract for coding agents. Read before writing code.

Status: Gate A freeze · signed 2026-08-31 PT · docs refactor 2026-09-01 · SoT: Gitea Ty_Tech/SlabUploader on gitea-atd (not mirrors).

---

0. Doc precedence (hard)

When docs disagree, use this order:

1. docs/TECH-SPEC-PIPELINE.md — deterministic math, image prep, constants, test vectors
2. docs/ARCHITECTURE.md — hybrid topology, what-runs-where, secrets why, status machine
3. This AGENTS.md — hard rules for agents (verbatim locks below)
4. docs/DATA-MODEL.md, docs/OPENAPI.md, docs/CONTENT-WOO.md, docs/PROMPTS.md, docs/DEPLOYMENT.md. Presentation: docs/UX.md (full rewrite landed 2026-09-02, same day as the deletion). Functional gates stay in this file.
5. docs/IMPL-PLAN.md — hybrid-aligned Phases 0–6. Phase 2 is a deferred stub only (offline PWA / OCR / ruler). Do not implement deferred stubs or any server happy-path measure/crop as SoT.
6. README.md — orientation only
7. docs/archive/PRD-2026-08.txt — frozen archive only. Not a source for new work.

Archived PRD lines that conflict with the files above are non-authoritative (see archive header and known-stale list).

Known stale (archive / historical only)

• Archived PRD FR17 “Tags: fig-* only” → wrong. Locked model is fig-* (1+) + feat-* (0+). See DATA-MODEL.
• Archived PRD FR19 / overview “species and wood category” pricing → wrong. Species $/bdft only. Category-weighted pricing is deferred (v1.5).
• Archived PRD FR40 “server-side processing” and NFR “server normalization” → wrong. Client owns happy-path measure/crop; server stores drafts and runs U2Net on demand only.
• Archived PRD FR42 / deploy bullets hard-coding .201 → superseded by env-agnostic DEPLOYMENT (inventory + APP_HOSTNAME).
• Older “creds client-side” decision → wrong. Server AES-GCM only; never browser storage for secrets.
• Offline-first PWA / OCR / ruler product text → superseded by hybrid + online-only MVP.
• Older “pending” Woo create default → superseded: UAT uses Woo draft + SLAB-UAT-*.

Open doc defects

Landed on docs/refactor-v2: #5 Application Passwords, #8 Call 1 prefill, #9 confidence gate, #11 WOO_BASE_URL, #12 synced-species-only, #15 inline mask knobs (U2Net deferred), #14 taxonomy manufactured last-modified, #2 portable Call1→Call2, #17 shared validation module, #18 409/422 field recovery. #10 UX spec: rewritten 2026-09-02; Ty signed the walk 2026-09-06 PT with one amendment (PC browsers in the compatibility matrix). Presentation: docs/UX.md. The pre-deletion file is history, not a source. #4 browser floor: phone + PC families locked; exact minimum version pins and API inventory depth still open. OPENAPI #3 wording already landed if the body still looks open.

---

0b. Gates (human map)

Label	Human name	What it means
Gate A	Docs freeze	Product locks written into the repo
Gate B	Run-host setup	Inventory the real host, deploy skeleton
Gate C	First draft listing	One Woo draft listing with inference OFF

Phase map: Phase 0 = Gates A+B. Gate C = early Phase 1 claimable publish slice. Call 1/2 after Gate C, still inside POC. Depth: docs/IMPL-PLAN.md.

Freeze provenance (Gate A)

SHA	What locked
a94d81d	gitea-atd Ty_Tech/SlabUploader as remotes SoT
457711b	AGENTS.md + hybrid IMPL Phase 0
4ec478f	Contradiction patches (README/PRD/Woo/DEPLOYMENT)
4d9f307 / 4e97770	Inference review + hybrid capture pipeline (pre-freeze baseline)

Docs refactor stack on docs/refactor-v2 (2026-09-01) applies OPENAPI store-only, env-agnostic deploy, CONTENT-WOO/DATA-MODEL facts, hybrid Phases 1–6, PRD archive. Product locks above are unchanged.

---

1. Topology (locked — hybrid)

Client (SvelteKit) owns the happy path:

• Capture 1–5 photos
• Sheet detect + chroma-key / black threshold + four knobs inline on capture (sheet, sensitivity, edge offset, feather) with live re-run and mask overlay on the source photo
• Mask overlay + user confirm before measurement
• Length axis overlay (rotate / confirm)
• sqft, bdft, 6" width samples
• 3:4 transparent PNG, 80% fill at extremes, shorter side ≥ 1600 when source allows
• Draft payload: originals + processed PNGs + length/thickness/SKU/sqft/bdft/widths

Server (FastAPI) owns:

• Store drafts / settings / taxonomy cache
• U2Net “Try harder” — deferred from POC, not deleted. Architecture keeps the one-photo server path for awkward backgrounds post-POC. POC mask tuning is client-side only (sheet + four knobs, live re-run). Do not implement U2Net in Gate C / first listing.
• WooCommerce (taxonomy sync, SKU check, create product) — browser never calls Woo
• Inference proxy (Call 1 / Call 2) — keys never in browser
• Prompt files (server text files, read fresh)

Happy path never leaves the phone until the user continues after confirmed mask + numbers.

Depth: docs/ARCHITECTURE.md.

---

2. What NOT to build (early phases / MVP)

Deferred / killed for early phases	Notes
Offline capture / PWA / IndexedDB sync	Online-only MVP
OCR (SKU)	Manual SKU entry
Ruler detection / OpenCV on client for scale	Manual length; scale = longest_axis_px / L
Server-side happy-path pipeline	No server chroma-key / crop / bdft math as SoT
Fake upscale / inventing pixels	Warn + retake if < 1600 after crop
U2Net “Try harder” server mask	Deferred from POC (keep in architecture; do not delete)
LoRA / fine-tune	Deferred
Figure-/category-weighted pricing	v1.5
Auto-dimension overlay	v1.5
Bulk publish, staff roles, analytics, SKU update path	Non-goals

---

3. Deterministic core (non-negotiable)

Encode TECH-SPEC exactly:

• bdft = sqft × thickness_in (identity: 1 sqft @ 1" = 1 bdft). No /12.
• Units: inches in 1/8" steps; sqft/bdft 2 decimals; price half-up 2 decimals.
• Manual length, thickness, SKU, price always win.
• Image prep: sheet/sliders → mask confirm → length axis confirm → 3:4 PNG / 80% fill / ≥1600px when source allows.
• No inventing pixels. No fake upscale. Undersized source → warn + retake; photo not publishable until retake.
• Pricing: species $/bdft from Settings, seeded by Woo species sync; rec = bdft × price_per_bdft; user override sticks; missing rule → empty rec, manual allowed.
• Port TECH-SPEC TV-1 through TV-10 (including TV-10 Rounding). Client (or shared pure modules the client uses) is authoritative for math; server stores client-sent numbers — does not recompute as SoT.

Depth: docs/TECH-SPEC-PIPELINE.md.

---

4. Secrets and settings

Never in browser storage / JS bundles:

• Woo WordPress username + application password (not consumer keys)
• Inference API key (treat inference URL as sensitive)
• AES master key (SLAB_AES_KEY env only; never in DB or repo)

WOO_BASE_URL: compose / host .env only (https:// required). Settings shows it read-only. Not a SQLite settings row. Not user-editable. FastAPI rejects http:// at startup. Do not reintroduce an editable store URL.

Server: AES-GCM at rest in SQLite for Woo Application Password + inference secrets. Settings UI POSTs username + application password to FastAPI; server encrypts. All Woo and inference calls are server-side. Dedicated low-privilege WP user (product create/edit + media). Rotate by revoking the application password in WP Users → Profile and re-entering in Settings. Do not implement WooCommerce consumer keys (ck_/cs_) or the Woo → Settings → Advanced → REST API key UI.

Browser may persist (non-secret): sheet mode, sensitivity, edge offset, feather (localStorage OK; reset-to-default required).

Why/flow: docs/ARCHITECTURE.md §5. Ciphertext key inventory: docs/DATA-MODEL.md settings table.

---

5. Woo UAT and create status (Ty locked 2026-08-31)

• Store URL: WOO_BASE_URL from Docker Compose (https:// only). Intended UAT/POC store is Whidbey Wood Store production; the compose value is the SoT, not a Settings field.
• Mid-phase UAT creates: Woo status draft, SKU prefix SLAB-UAT-*
• Settings field woo_create_status: draft | publish, default draft
• Hard rule: if SKU matches SLAB-UAT-*, force Woo draft even if Settings say publish
• Final POC success: one real (non-UAT) SKU → Woo publish only after Ty review
• Duplicate SKU → stop; UI maps 409 to the SKU field (edit SKU or open existing). No edit-in-place of the other listing in MVP
• After successful publish: purge originals + processed images on server
• Do not conflate with slab lifecycle status (below)

Payload and publish sequence depth: docs/CONTENT-WOO.md (cites this section for safety policy). Key only: DATA-MODEL woo_create_status.

---

6. Inference (inside POC, sequenced)

• Feature-flagged; full suite must pass with inference OFF.
• Gate C claimable path: inference OFF; manual taxonomy + deterministic numbers + Woo draft path.
• Call 1 (vision) / Call 2 (text) only after Gate C is green, still inside overall POC before “done” — not post-POC.
• Call 1 auto only if inference_enabled; Call 2 never auto-fires (user taps Generate text).
• Call 1 body: all inventory photos @1024 (not top-down-only). Taxonomy lists sent to Call 1 are exactly the Woo-synced species / attribute / tag sets. No hardcoded or invented species.
• Confidence 0.7 (default): ≥ 0.7 → pre-populate species, wood categories, edge, figure, grade, and feat-* from Call 1 (mutable). < 0.7 → those fields stay empty. That empty-start subset is not the submit list. Ready still requires the full named set in §7 (species + one figure is not enough). Presentation: docs/UX.md. Inference never auto-publishes.
• Taxonomy cache: taxonomy_last_sync_at vs taxonomy_anchor (bump only on stored-subset hash change, including deletions). Client re-pulls iff the anchor advanced. test-woo always syncs; publish if last_sync >1h. No per-nav sync. No Woo taxonomy timestamp.
• Call 2 portable context: default full resend; never require previous_response_id. test-inference two-turn probe stateful|stateless|fail. Fail blocks Call 2 enablement only.
• LoRA deferred. No numeric/dimension invention by LLM; templates inject deterministic numbers.

Prompt file mechanics: docs/PROMPTS.md.

---

7. Slab status machine (app-internal only)

draft → calibrated → ready → publishing → published (or failed with detail).

• calibrated: mask + axis confirmed; length/thickness/SKU entered; client computed sqft/bdft/widths; draft uploaded.
• ready: SKU, length, thickness, client sqft/bdft/widths, ≥1 inventory PNG, exactly one species, ≥1 wood category, 1 edge type, ≥1 figure term, ≥1 fig-* tag, ≥1 grade, thickness store band (round-up), 1 moisture (default kiln-dried; not inferred), price. feat-* 0+. Title/description may be typed, generated, or templated at publish. Vision may prefill the taxonomy subset only when confidence ≥ 0.7.
• Inference does not gate publish. Below-threshold Call 1 does not prefill; empty mandatory fields block ready (see UX spec).
• Never confuse slab draft with Woo draft (woo_create_status).

---

8. Test and delivery expectations

• No feature without a passing test and shown execution output (PR/commit message).
• Deterministic core unit-tested before UI consumes it.
• One phase = reviewable, shippable increment; Ty exit-gates phases.
• Phase naming: Phase 0–6 stay delivery chapters. Phase 0 = Gates A+B. Gate C = early Phase 1 claimable publish slice.
• Do not write secrets, real Woo application passwords, or AES keys into the repo.
• Prefer fixing docs when code and locked decisions diverge — do not silently reintroduce /12, offline, ruler, or server happy-path pipeline.

---

9. Deploy sketch (see DEPLOYMENT.md)

• Target: inventoried Docker host. Docker Compose (frontend + fastapi) + Caddy → https://${APP_HOSTNAME} (LAN/Tailscale).
• Fill deploy/INVENTORY.md before binding ports or merging Caddy (co-resident apps can collide; bolt.diy is one example).
• Health: GET https://${APP_HOSTNAME}/api/health.
• WOO_BASE_URL in compose (HTTPS). Settings displays it read-only.
• Inference is a remote configurable endpoint, not on the app host by default.
• Historical lab IP/hostname examples in older notes are non-normative.