5.4 KiB
Morning notes for Ty — docs refactor v2 (Dev Manager)
Date: 2026-09-01 (before morning PT)
Read: /workspace/slabuploader-docs-refactor-proposal-v2.md + /workspace/slabuploader-docs-refactor-interrogate.md
Branch check: docs/refactor-v2 tip = main @ e1e087b (PM-DESIGN unslop only). v2 commits 1–4 are not on the branch yet — these notes judge the proposal, not a landed stack.
Constraints honored: no Gate B / host inventory; main not touched by this pass.
1. Are Gate C slice + env-agnostic deploy + OPENAPI B1/W1 enough?
Short answer: almost — for architecture. Not for a green “hand to coding agents” flag.
| After… | Agents stop shipping wrong architecture? | Still broken if they follow leftover text |
|---|---|---|
Commit 1 only (B1/W1 + Gate C slice + kill processing + Phase 2 stub + Phase 6 offline strip + TV-10 + AGENTS known-stale rewrite) |
Mostly yes, if AGENTS keeps the “IMPL 1+ incomplete — implement only Phase 0 + rewritten sections” guard until Phases 3–5 are cleaned | CONTENT-WOO JPEG→lose alpha; PRD FR19 category pricing / FR40 server processing if still live; Phase 3–5 top-down vision / offline flavor; .201 hard-coding (deploy, not architecture) |
+ Commit 2 (env-agnostic + deploy/INVENTORY.md + Caddy handle order) |
Architecture same; Gate B framing becomes safe | Same feature landmines as above |
| + Commit 3 (CONTENT-WOO / DATA-MODEL / ARCHITECTURE / IMPL 3–5) | Yes — this is the real agent-safe checkpoint | Nits / archive hygiene |
| + Commit 4 (de-dupe, PRD archive, PM-DESIGN delete after DoD) | Hygiene / reader load; not required to unblock scaffold | — |
Recommendation: Do not unleash coding agents on Gate B scaffold until commit 1 is on the branch and grepped. Do not unleash feature agents (capture/Woo/inference UI) until through commit 3. Env-agnostic (commit 2) can land with or right after commit 1; it does not by itself stop wrong architecture.
Gate C slice before Phase 2 delete is non-negotiable (interrogate A6). v2 has that right. B1 alone without Gate C still leaves capture homeless.
2. Deploy / IMPL landmines v2 still misses or underweights
v2 covers the big ones (B1, processing, Phase 2/6, env-agnostic, Caddy order, RemBG/1000px, TV-10, Gate C home). Add or keep visible:
- OPENAPI dedupe language — “SKU already published” should mean “SKU already exists in Woo (any status).” Else agents skip draft collisions.
- Confidence 0.7 (FR22a) — v2 folds into Phase 5/PROMPTS; make that an explicit commit-3 checklist line so it does not die with PRD archive.
- IMPL Phase 1 remainder — after Gate C slice, leftover “sandbox Woo / pending-era” sentences in Phase 1 body still need a one-pass scrub or they reintroduce wrong publish defaults next to the new slice.
- Compose↔Caddy service names — fragment must match compose service name (
fastapi) and sit on the same Docker network; design text should say that in commit 2, not only handle order. - Partial Woo failure / orphan media — ops note, not architecture; keep in CONTENT-WOO/IMPL Risks so agents do not invent auto-delete of Woo media.
- Do not treat workspace
/workspace/*.mdas SoT — already true; worth one AGENTS line when env-agnostic lands so agents do not read proposal files as repo authority.
Nothing here reopens product locks.
3. Commit staging feedback (v2 commits 1–4)
Order is correct. Wrong-code-first → env-agnostic → factual → de-dupe/archive matches how agents fail. Keep it.
| Commit | Staging feedback |
|---|---|
| 1 | Keep atomic: OPENAPI B1+W1 including examples, Gate C slice, Phase 2 stub, processing gone, Phase 6 offline strip, TV-10, AGENTS stale-list rewrite. Do not ship a HEAD where Phase 2 body is deleted without Gate C written. Prefer one commit; if split, Gate C + Phase 2 stub must move together. |
| 2 | Fine as second. Unblocks Gate B framing only — still no inventory/SSH until you open Gate B. Include corrected Caddy fragment + empty deploy/INVENTORY.md template in the same commit DEPLOYMENT claims them. |
| 3 | Highest ROI after 1 for “agents won’t follow stale Phase 5.” Prefer full hybrid-align of Phases 3–5 here (agree with PPM) over a long incomplete-rewrite guard. |
| 4 | Last. PRD archive + reference rewrite same commit. PM-DESIGN delete only after DoD greps green + Gate table + freeze provenance in AGENTS. |
Agent-safe checkpoints (for CoS/you): after 1 (docs contract); after 3 (feature docs); after 4 (cleanup done). Do not hand implementers a mid-commit-4 tree.
Branch reality check: overnight commits 1–4 are not on docs/refactor-v2 yet (tip = e1e087b). Morning decision is “sign v2 proposal / open execution on that branch,” not “review four landed commits.”
Bottom line for Ty
- Sign v2 direction (not v1). Product locks stay put.
- Execute commit 1 first on
docs/refactor-v2— that is what stops the worst wrong architecture. - Coding agents: after commit 1 for Phase 0/contract only; after commit 3 for capture/Woo/inference work. Still no Gate B host work until you explicitly open it.
- Open Qs: I align with PPM — archive PRD with superseded header; delete PM-DESIGN after DoD; hybrid-align IMPL 3–5 in commit 3.
Dev Manager standing by under HOLD for Gate B / unrelated main writes.