Stage 1 Decisions — WO-0008 (Maya)

Stage 1 Decisions — WO-0008 (Maya)

Owner: Pablo (Production Line Architect) Date: 2026-05-26 Closes: Stage 1 of WO-0008 Inputs: order.md, TFD-0023, RD-0017 flashcard, sandbox-report.md

Open Questions resolved

Q1 — Corpus filters for M1 Maya-Factory

Decision: Include the following paths from C:\Projects\talent-factory\:

  • company/ (decisions, vision, services, requests-registry)
  • departments/**/requests/**/*.md (only request.md, *-brief.md, key outputs)
  • departments/**/*.md at root level (role definitions, intake logs)
  • references/videos/RD-*/notes/*.md and references/videos/RD-*/out/flashcard_*.md (skip raw transcripts — too verbose for synthesis)
  • production-lines/orders/WO-PROD-*/order.md and production-lines/orders/WO-PROD-*/process/*-complete.md
  • KB lessons / framework spec files

Exclude:

  • .claude/ (commands and hooks — not knowledge)
  • node_modules/, dist/, _publish/ (build artifacts)
  • intranet/dist/ (rendered HTML — duplicates the source markdown)
  • OneDrive client folders (M2 scope per deployment)
  • Anything .git-tracked as binary

Rationale: Sandbox proved markdown-only synthesis works. HTML deliverables are out of M1 because they're rendered from sources already in scope.

Q2 — Wiki location

Decision: Wiki lives in a sibling folder to the corpus, not committed to the main repo.

Path: C:\Projects\talent-factory-maya-wiki\ (separate folder, optionally its own git repo for history).

Rationale: Avoids git noise on every ingest (wiki updates would dominate commit history). Sibling-folder keeps re-index simple (Maya watches talent-factory/, writes to talent-factory-maya-wiki/). Optional .git in the wiki folder gives audit trail without coupling.

Q3 — Re-index cadence

Decision: Hybrid — git post-commit hook for incremental updates + nightly full rebuild as safety net.

  • Post-commit hook (~5s typical) — detects changed files in the corpus filter, runs ingest only on those files. Fresh wiki within 60s of any commit.
  • Nightly 02:00 cron — full re-ingest + /lint-wiki automated report → posted to Telegram channel (factory ops).

Rationale: Cheap to try both. Hook gives freshness, nightly gives a known-good baseline + automated lint visibility.

Q4 — Widget styling

Decision: Match the factory docs design system (Anthropic warm cream + Trustworthy Blue #2563EB + Instrument Serif headings per docs-design-system memory).

Widget contract:

  • Fixed bottom-right launcher (50x50 px, blue circle).
  • Open state: 380x600 panel, cream background.
  • Citations render as inline [TFD-0019] pills, click-through to source HTML on intranet.
  • FR/EN agnostic (no language selector in chrome).

Per-client deployment can override via CSS variables; factory ships the defaults.

Decisions captured from sandbox (sandbox-report.md §"What needs decision before M1 build")

Q5 — Wiki page ID stability

Decision: Filename = stable slug, H1 = current display name. Lint detects orphans on rename. Adopt as schema convention in sandbox/CLAUDE.md (already drafted).

Q6 — Update vs append on re-ingest

Decision: Append-with-edit-log section per page. Each page maintains a ## Edit log tail section listing {date} — {source TFD} — {change summary}. Maintains auditability; lint flags pages where the edit log grows beyond ~10 entries (signal for restructuring).

Q7 — Cross-language synthesis policy

Decision: One bilingual page per concept/entity. Source-language quotes preserved verbatim. Synthesis prose written in the source-language-dominant tone (FR if majority of sources are FR). No separate FR/EN tree.

Q8 — Answer pages — generated vs persisted

Decision: Persist after first asking when Maya answers from wiki. Stored in wiki/answers/. /lint-wiki flags answer pages whose cited TFDs have been updated since the answer was generated (= candidate for regeneration). Answers from raw (i.e., wiki had no synthesis) are not persisted — they trigger a backlog item to add the missing concept page instead.

Frontmatter convention adopted for M2 (deferred)

Concepts and TFDs in wiki/ will eventually carry frontmatter:

---
type: concept | entity | answer | decision-record
slug: stable-slug
refines: [other-slug]
supersedes: [other-slug]
sources: [TFD-0019, TFD-0021]
last-ingest: 2026-05-26
---

Not implemented in M1 — measure pain at 30+ pages first per sandbox §"Pattern friction".

Stage 1 sign-off

  • Sandbox proof complete (sandbox-report.md)
  • 4 order-level open questions answered (Q1-Q4)
  • 4 sandbox-surfaced decisions made (Q5-Q8)
  • Adapted CLAUDE.md schema committed → stage-1-claude-md-draft.md ← next file
  • Stage 2 solution spec opened → stage-2-solution-spec.md ← next file

Target Stage 1 close: 2026-06-02. On track.