Phase 1 foundation for the Stargue Publishing Engine (plan v2, BMAD
panel-reviewed 2026-04-19 — 1 APPROVE, 6 REVISE, 0 REJECT; all principles >=3).
- Governance doctrine adopted from DQMS
(.clinerules/12-foundational-principles.md,
.claude/hooks/gate-plan-exit.sh, .claude/skills/bmad-plan/SKILL.md)
- Bun workspaces + Turbo; apps/{mcp-linkedin,scheduler,admin};
packages/{schema,sanitize,linkedin-client,observability}
- Drizzle schema (content, publications, approvals, metrics,
linkedin_tokens, audit, outlet_feature_flags) with idempotency_key
UNIQUE and kill-switch table per TEA/dev panel revisions
- LinkedIn API canon: Posts API /rest/posts (not legacy UGC); OAuth
auth-code without PKCE; secretbox (not sealed-box); Community
Management API as separate approval gate from MDP
- Frontmatter Zod schema (status, language, outlets[], sanitize,
scheduled, version)
- Pino observability with PII redaction
- Expand-then-contract migration runbook
- Plan + panel verdicts mirrored to docs/plans/
- Deferred gates logged (Dokploy PaaS verification, LinkedIn Dev
Portal app registration)
bun install + bun run typecheck both exit 0 across 11 workspaces.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
31 KiB
Executable File
created, updated, tags
| created | updated | tags |
|---|---|---|
| 2026-04-19T09:45 | 2026-04-19T10:45 |
Plan — Phase 1 Automated Publishing Engine (with LinkedIn MCP Server)
Revision: v2 (consolidated from 7-agent BMAD panel review 2026-04-19)
Verdicts file: /tmp/bmad-panel-verdicts-publishing-engine-phase1-20260419.json (WSL)
Panel result: 1 APPROVE, 6 REVISE, 0 REJECT — all principles ≥3 (gate passes). Revisions below address each REVISE gap, citing originating agent.
Applicability: tier-2 = [RBR, A11y, Testability]
Precedence: CLAUDE.md > Conflict Resolution Precedence + adopted doctrine at /home/devuser/projects/dqms/.clinerules/12-foundational-principles.md (user directive 2026-04-19)
1. Context
User directive 2026-04-19: Full Phase 1 automation, now. No manual posting, no postponing. LinkedIn programmatic refresh is gated to MDP partners per Microsoft Learn > Programmatic Refresh Tokens (retrieved 2026-04-19). Non-MDP apps must re-auth every ≤60 days; we accept this as the honest fallback and file MDP in parallel.
Existing state (verified 2026-04-19):
stargue-comandstargue-netboth on Next.js 16 + React 19 + Bun + Tailwind v4 + remark (verified againstpackage.jsonper dev agent). Note: stargue-net/package.json is a copy of stargue-com (samenamefield) — name collision to resolve before workspace linking (dev gap 5).- Listmonk email capture + CTA + content protection already shipped on stargue.com.
- Dokploy on
sg-paas-s1.stargue.net; Gitea atgit.stargue.net; OpenObserve observability stack. - Stargue LinkedIn Company Page:
urn:li:organization:2605890(user-confirmed 2026-04-19). - Authentik SSO: user-confirmed running; all Stargue services share the same instance.
Capacity reconciliation (PM/SM gap). PRD §6.5 caps Angelo at 4h/week. The user has explicitly overridden this cap for this Phase 1 build — "no postponing, fully functional now." Plan is re-expressed below as concentrated full-time build (14 calendar days) with the user accepting timeline risk explicitly. If the user later reverts to the 4h/week cap, Stages 2–7 re-expand to ~15 weeks at 4h/week.
2. Objectives
- Automated two-channel publish pipeline — Obsidian vault → stargue.com/.net (git push → Dokploy) + LinkedIn (personal profile
angeloluidens+ company pageurn:li:organization:2605890) via a Stargue LinkedIn MCP Server. - Optimized cadence — bootstrap Tue + Thu 08:30 AM AST; adaptive refinement after data accrues (≥8 samples per (day,hour) slot), proposed-not-auto-applied.
- Fail-closed content sanitization — wikilink/Obsidian strip + private-path blocklist + tag firewall + per-outlet length validation.
- Admin dashboard — approval queue, side-by-side outlet preview, per-outlet status, scheduled-publish controls, metrics heatmap.
- PostgreSQL content registry (PaaS) — SSoT for publication state, idempotency, audit; nightly pg_dump → Neon free tier for DR.
- OAuth 2.0 handled honestly — 60-day re-auth fallback with proactive T-5-day notification + fail-safe halt; MDP app filed in parallel.
- Revenue conversion preserved (PM gap). Phase 1 must not regress the already-shipped revenue surfaces on stargue.com (Listmonk email capture, CTAs, service page, Terms of Use). New inbound-inquiry tracking column on
publications.metadata_jsonb.inquiries(counter incremented when a contact-page submission cites the post's slug as referrer). - Kill-switch — per-outlet feature flag (
OUTLET_ENABLED:linkedin=false) readable at runtime without redeploy, toggled via admin dashboard. TEA gap 10.
3. Architecture
3.1 Repository topology (simplified per architect gap 1)
Collapsed apps/api into apps/admin Next.js App Router route handlers. Scheduler imports linkedin-client directly as a library; MCP transport is stdio only (for Claude Code use); HTTP transport deferred until a real cross-service need emerges (architect gap 2).
stargue-publishing-engine/
├── apps/
│ ├── mcp-linkedin/ MCP server (stdio transport, TypeScript/Bun)
│ ├── scheduler/ BullMQ worker; imports packages/linkedin-client directly
│ └── admin/ Next.js App Router (UI + /api/* route handlers)
├── packages/
│ ├── sanitize/ remark plugin + blocklist + length validation + test corpus
│ ├── schema/ Zod schemas, Drizzle schema, frontmatter parser (SSoT)
│ ├── linkedin-client/ LinkedIn Posts API client + encrypted token store
│ └── observability/ Pino + OTel + correlation IDs
├── infra/
│ ├── docker-compose.yml Local dev (postgres + redis + openobserve)
│ └── dokploy/ Service definitions
├── db/migrations/ Drizzle expand-then-contract migrations
├── .clinerules/ Adopted from dqms
├── .claude/
│ ├── hooks/gate-plan-exit.sh
│ └── skills/bmad-plan/SKILL.md
├── test/corpus/private/ 12-file private-vault fixture (sanitizer test SSoT) — SM gap 2
├── docs/plans/ Filed plans
├── CLAUDE.md
├── package.json Bun workspaces
└── turbo.json
3.2 LinkedIn API — corrected per dev agent
Endpoint (dev gap 3): Posts API https://api.linkedin.com/rest/posts — NOT legacy /v2/ugcPosts. Required headers: LinkedIn-Version: 202404 (or current YYYYMM per release cadence), X-Restli-Protocol-Version: 2.0.0, Authorization: Bearer <access_token>. Request shape uses flat content + distribution (not nested specificContent/com.linkedin.ugc.ShareContent).
Auth flow (dev gap 2): OAuth 2.0 Authorization Code flow — no PKCE (LinkedIn 3LO does not document PKCE support per Microsoft Learn > 3-Legged OAuth Flow, retrieved 2026-04-19). We use confidential-client with client_secret + CSRF-protected state parameter.
Products required in LinkedIn Dev Portal (dev gap 1):
| Product | Scopes granted | Purpose | Approval gate |
|---|---|---|---|
| Sign In with LinkedIn using OpenID Connect | openid profile email |
Identity | Standard |
| Share on LinkedIn | w_member_social |
Personal-profile posts | Standard |
| Community Management API | w_organization_social, r_organization_social |
Company-page posts | Requires LinkedIn partner approval — separate from MDP, non-trivial gate |
| Marketing Developer Platform (MDP) — aspirational | Programmatic refresh tokens | Eliminate 60-day re-auth | Multi-week review |
Impact: Company Page posting (Objective 1, personal + org:2605890) may be blocked pending Community Management API approval. Personal-profile posting unblocked immediately. We file Community Management API application in Stage 0.5 and build the tools so they're ready when approved.
Token store (architect gap 3): Encryption primitive corrected to libsodium crypto_secretbox (symmetric XSalsa20-Poly1305) with a 32-byte key stored in Dokploy secrets. Sealed-box was wrong for at-rest DB encryption. Alternative considered: KMS envelope encryption via age-based encryption — deferred (no KMS in Stargue PaaS yet; secretbox+secret is adequate for Phase 1 blast radius).
Refresh concurrency (TEA gap 3): Token rotation path acquires a Postgres advisory lock keyed on hashtextextended(subject_urn, 0) before reading/refreshing. Lock released after commit. Prevents double-rotation race.
Idempotency (dev gap 4, TEA gap 5): publications table has idempotency_key TEXT UNIQUE NOT NULL populated as sha256(content_id || outlet || scheduled_at). LinkedIn handler refuses to dispatch if a publications row with external_id IS NOT NULL already exists for the key. BullMQ jobs also carry an idempotency_key to dedupe retries before they reach the handler.
Rate limits (architect gap 4). LinkedIn does not publicly document a fixed per-member posting cap. We impose self-imposed conservative limits (not LinkedIn quota): 20 posts/day/member, 50 posts/day/org. Stage 3 DoD asserts this via a test (TEA gap 6).
MCP tools exposed (stdio transport, no HTTP):
| Tool | Purpose |
|---|---|
linkedin_whoami |
Returns authenticated subject + scopes + MDP-eligible flag |
linkedin_auth_status |
Access-expiry remaining, refresh-expiry remaining, kill-switch state |
linkedin_create_post |
Text post via Posts API (author_urn param: urn:li:person:<id> or urn:li:organization:2605890) |
linkedin_create_article |
Long-form article (≤125k chars) |
linkedin_upload_media |
Image / PDF carousel; returns asset URN |
linkedin_create_post_with_media |
Combines upload + post |
linkedin_delete_post |
Delete post by URN (soft-reversible within 5-min edit window only; after that, delete-only) |
linkedin_get_post_metrics |
Impressions/reactions/comments/shares/clicks for a URN |
linkedin_get_profile_stats |
Aggregate profile stats for cadence optimizer |
3.3 Sanitization pipeline
packages/sanitize — Unified/remark plugin chain:
- Strip wikilinks
[[Path/Note|Display]]→Displayor plain removal. - Strip embeds
![[file]]→ resolve (blog) or remove (LinkedIn). - Strip dataview blocks.
- Strip callouts.
- Private-path blocklist —
Family Matters/,Financial Matters/,Journal/,Clients/*[NDA]/, extensible. Match triggers build error, not silent strip. - Tag firewall —
#private,#heal-internal,#confidentialtrigger build error. - Per-outlet length validation (LinkedIn 3k / article 125k / Twitter 280 / etc.).
- Idempotency hash output (
content_hash) over sanitized body for dedupe.
Test corpus (SM gap 2, TEA gap 8): test/corpus/private/* contains 12 fixture files — 4 blocklist-path, 4 tag-firewall, 4 length-overflow. Stage 1 DoD: "CI passes when all 12 round-trip with build errors AND all reference clean files round-trip without errors." Promoted from Risks to Stage 1 DoD.
3.4 Content registry (Postgres) + DR
Drizzle schema in packages/schema/db.ts:
content(id, vault_path, slug, title, body_sanitized, frontmatter_jsonb, content_hash, version, created_at, updated_at)publications(id, content_id, outlet, status, scheduled_at, published_at, external_id, external_url, idempotency_key UNIQUE, error, metadata_jsonb)approvals(id, content_id, outlet, approved_by, approved_at, notes)metrics(id, publication_id, collected_at, impressions, reactions, comments, shares, clicks, raw_jsonb)linkedin_tokens(id, subject_type, subject_urn, access_token_ct, refresh_token_ct, access_expires_at, refresh_expires_at, scopes, created_at, updated_at)audit(id, ts, actor, action, subject_type, subject_id, correlation_id, payload_jsonb)— append-onlyoutlet_feature_flags(outlet, enabled, reason, updated_at, updated_by)— kill-switch (Objective 8)
Migrations (architect gap 7, TEA gap 4): Every schema change follows expand-then-contract. Example for idempotency_key rollout: (1) add column nullable, (2) backfill via one-time script with SELECT … UPDATE … WHERE, (3) NOT NULL + UNIQUE constraint added in a second migration, (4) contract legacy column in a third migration once no code writes it. Each migration accompanied by a runbook in db/migrations/README.md. Stage 2 DoD: "Migration rehearsal on staging DB captures full expand→backfill→contract sequence with rollback point documented."
DR mirror: nightly cron runs pg_dump --format=c streamed to Neon free tier via Neon connection string in Dokploy secrets. Retains only last 30 days of metrics on the mirror (respects Neon 0.5 GB free cap). RPO ~24h.
3.5 Admin dashboard + A11y (UX gaps 1–8)
Next.js App Router. Authentik SSO (user-confirmed instance). Routes also host /api/* handlers (architect gap 1 collapse).
Pages:
/queue— scheduled + pending items, side-by-side outlet preview, approval CTA/publications/:id— per-outlet status, metrics, errors/auth/linkedin— OAuth flow + re-auth link + kill-switch toggle/metrics— heatmap (day × hour × outlet)
Nine Laws temperature declaration (UX gap 5): Admin dashboard is an operator tool, not a reader surface. It inhabits the cool / stargue.net (forge) temperature. Forest-green (#4E8211) + Citrus (#99CC00) accents. Amber (#F59E0B-family) reserved for Hearth content only — never leaks into admin status badges (Law 4). Status badges use Forest/slate/red variants.
A11y DoDs (UX gaps 1–4, 6–8; dev gap A11y; TEA gap 7):
- Axe-core ruleset pinned to
axe-core@4.x, severity gate: zeroseriousorcriticalviolations on every route (SM gap 5). - WCAG 2.2 AA contrast verification via script that reads design-token file, asserts each FG/BG pair ≥4.5:1 body and ≥3:1 large-text/non-text-UI on both themes.
- Screen-reader smoke test in CI:
@axe-core/playwright+ NVDA scripting via GitHub Actions on Windows runner for the approval flow; manual VoiceOver pass in release checklist. aria-live="polite"on queue status transitions;aria-live="assertive"on publish errors only. Announcement text specified per surface.- Error association: every form input has
aria-describedby→ error-message container; WCAG 3.3.1/3.3.3/4.1.3 complied. - Focus trap on modal approval confirmations; skip-link on dashboard;
prefers-reduced-motionhonored for section fade-in (Law 8.4). /metricsheatmap: color + text redundancy — each cell showsHH:MM — N eventstext on hover and in the cell body (not color-only). WCAG 1.4.1 satisfied.
3.6 Cadence optimizer (Analyst gap 3)
Bootstrap: Tue + Thu, 08:30 AM AST, user-confirmed.
Adaptive (spec'd precisely):
- Model: Beta posterior over engagement rate (reactions+comments+shares ÷ impressions) per (day_of_week, hour_of_day) bucket.
- Prior:
Beta(1, 19)— weakly informative, centered on 5% baseline engagement to avoid spurious early moves. - Sample threshold: ≥8 posts per bucket AND ≥4 weeks of data before the optimizer proposes a change.
- Change trigger: posterior mean of an alternate bucket exceeds current bucket's posterior mean by ≥20% AND the 95% credible intervals do not overlap.
- Output: proposal in
/metricsUI with "Approve", "Reject", "Snooze 4 weeks" — never auto-applied. Human-in-the-loop (RBR). - Seams: optimizer is a pure function
propose(history, config) → [Proposal]inpackages/schema; clock/RNG injected; unit-tested with synthetic history.
3.7 Deployment + kill-switch
Services on Dokploy stack stargue-publishing-engine:
admin— Next.js standalone, port 3002, Traefikpublishing.stargue.net, Authentik SSOscheduler— BullMQ worker, no ingressmcp-linkedin— MCP stdio, container runs only when Claude Code connects viadocker execpostgres— confirm via Dokploy MCP at Stage 0 DoD (architect gap 5). If no existing cluster, provision a new stack service.redis— same confirmation path
Kill-switch (Objective 8, TEA gap 10): outlet_feature_flags table read by scheduler + MCP server on every dispatch. Toggle via admin dashboard at /admin/feature-flags. Flip takes effect within 30s without redeploy. Audit-logged.
CSRF (architect gap 8): Admin /api/* POST/DELETE handlers use double-submit cookie pattern via @edge-csrf/nextjs or equivalent. Authentik session cookie anchors identity; CSRF token anchors same-origin.
4. Stages + measurable DoD (all "works" language tightened per SM/TEA gaps)
Stage 0 — Governance + gate resolution (day 1)
| Step | Measurable DoD |
|---|---|
| 0.1 Create Gitea repo | GET https://git.stargue.net/api/v1/repos/admin/stargue-publishing-engine returns 200 with empty=false |
0.2 Install .clinerules/ from dqms |
diff -q .clinerules/12-foundational-principles.md /home/devuser/projects/dqms/.clinerules/12-foundational-principles.md reports no difference |
0.3 Install .claude/hooks/gate-plan-exit.sh |
Hook file present with exec perms; bash -n .claude/hooks/gate-plan-exit.sh parses clean |
0.4 Install .claude/skills/bmad-plan/SKILL.md |
File SHA256 matches dqms SKILL.md |
| 0.5 Project CLAUDE.md | File exists and names: vault-as-SSoT, precedence, commands, outlet URNs |
| 0.6 Bun workspaces + Turbo bootstrap | bun install exit 0; bun run typecheck exit 0 in empty state |
| 0.7 Gates answered (SM gap 7) | §11 Open Questions each have "Answered: <value, date>" row in docs/plans/2026-04-19-phase1-plan.md |
| 0.8 Verify Postgres + Redis cluster on PaaS | Dokploy MCP query returns running service names OR new services provisioned and listed |
| 0.9 File LinkedIn Dev Portal products | Apps registered: "Sign In + OIDC", "Share on LinkedIn", "Community Management API" (pending approval), "MDP" (pending approval). Screenshot URLs recorded in docs/linkedin-apps.md |
| 0.10 Resolve stargue-net/package.json name collision (dev gap 5) | stargue-net/package.json has "name": "stargue-net" (distinct from stargue-com) |
Stage 1 — Shared packages (days 2–3)
| Step | Measurable DoD |
|---|---|
1.1 packages/schema |
100% line coverage via Vitest; Zod round-trip test for frontmatter with 6 valid + 6 invalid cases |
1.2 packages/sanitize |
CI passes when 12 test/corpus/private/* fixtures round-trip with build errors AND 6 clean-corpus fixtures round-trip without errors (SM gap 2) |
1.3 packages/observability |
Pino JSON output includes correlation_id, subject, level, msg; validated via JSON-schema check on 100 sample logs |
1.4 packages/linkedin-client |
Fake client passes contract-test suite using msw (chosen recorder, dev gap 7) against recorded Posts API fixtures |
Stage 2 — Database + core API (days 3–4)
| Step | Measurable DoD |
|---|---|
| 2.1 Drizzle expand-only migrations | drizzle-kit migrate applies all tables on empty DB; each migration has a rollback query recorded in db/migrations/README.md |
| 2.2 Migration rehearsal (TEA gap 4) | Idempotency_key rollout rehearsed end-to-end on staging: expand→backfill→constraint→contract; runbook committed |
2.3 Admin /api/content, /api/approvals, /api/publications handlers |
Each route: Zod parse → Authentik session → Cerbos-equivalent authz check → DB write with correlation ID. Vitest integration tests assert 401/403/200/422 for each route |
2.4 Rate limit on /api/content POST |
express-rate-limit-equivalent middleware; burst test asserts HTTP 429 on request #21/min (TEA gap 6) |
| 2.5 CSRF middleware | Double-submit cookie test: cross-origin POST returns 403 |
Stage 3 — LinkedIn MCP server (days 4–7)
| Step | Measurable DoD |
|---|---|
| 3.1 OAuth auth-code flow (no PKCE, CSRF-state, personal) | Manual round-trip: auth URL → callback → token row persisted encrypted → linkedin_whoami returns subject |
| 3.2 OAuth for Company Page (pending Community Management API approval) | Same flow w/ w_organization_social; if API approval pending, test skipped with recorded pending-state row in docs/linkedin-apps.md |
3.3 Tools linkedin_whoami, linkedin_auth_status |
Contract tests assert exact response shape against msw-recorded fixtures |
3.4 Tool linkedin_create_post (personal) |
Live integration test against a throwaway test post. DoD: "Post URN persisted to publications.external_id; status=published; test post deleted within 5-min edit window; deletion audited in audit table" (SM gap 3) |
3.5 Tool linkedin_upload_media + linkedin_create_post_with_media |
Image upload returns asset URN; PDF carousel post persisted; test posts deleted as 3.4 |
3.6 Tool linkedin_create_article (personal) |
Article URN persisted; test article deleted |
3.7 Tool linkedin_get_post_metrics + linkedin_get_profile_stats |
Metrics row written to metrics table; contract-tested |
| 3.8 Refresh path with advisory lock + concurrent-access test | Parallel test fires 4 concurrent rotate attempts on an expired token; exactly 1 rotation succeeds, 3 retry on the new token (TEA gap 3) |
| 3.9 Fail-safe halt on auth expiry | Forced-expiry test: outlet_feature_flags.linkedin.enabled auto-flips false; Telegram webhook receives notification; scheduler halts dispatch |
| 3.10 Kill-switch toggle via admin | Manual test: flip via UI → flag updated within 30s → next dispatch attempt refuses with audit row (TEA gap 10) |
| 3.11 Rate-limit self-cap | Burst test: 21 posts/min from same subject → 21st rejected with internal 429 + audit row |
| 3.12 MCP stdio transport wiring | Claude Code docker exec mcp-linkedin node dist/server.js responds to initialize + tools/list per MCP spec 2024-11-05 (pinned) |
Stage 4 — Scheduler (day 7–8)
| Step | Measurable DoD |
|---|---|
| 4.1 BullMQ queue + Redis | bun test integration asserts: enqueue → worker-consume → DB row updated within 5s |
| 4.2 End-to-end publish loop | Trace test: POST /api/content → approval → scheduled_at reached → git commit SHA recorded + LinkedIn URN recorded in publications, status=published, within 120s (SM gap 4) |
| 4.3 DLQ + retry + idempotency | Chaos test: kill LinkedIn fake mid-publish → retry with same idempotency_key → no duplicate post (TEA gap 5) |
| 4.4 Cadence optimizer integration | Feature flag default OFF; unit tests assert proposals with ≥8 samples + ≥4 weeks threshold; no proposal below threshold |
Stage 5 — Admin dashboard (days 8–10)
| Step | Measurable DoD |
|---|---|
| 5.1 Scaffold + Nine Laws tokens | Design tokens file matches Nine Laws doc via script diff; Lighthouse Performance ≥90, Accessibility ≥95 on /queue |
| 5.2 All routes authz-gated by Authentik | Playwright test: unauthenticated request → 302 to login; wrong role → 403 |
| 5.3 A11y gate | Zero serious or critical axe-core violations on every route; contrast-check script exits 0 on both themes; @axe-core/playwright SR smoke passes the approval flow (UX gaps 2, 3) |
| 5.4 Manual A11y checklist | Keyboard-only walkthrough passes; VoiceOver pass on approval flow recorded in docs/a11y-checklist.md; focus-trap on modals verified |
| 5.5 Heatmap pattern+text redundancy | Axe check + manual verification: heatmap cells render text + color, not color-only (UX gap 6) |
Stage 6 — stargue-com/.net integration (day 10)
| Step | Measurable DoD |
|---|---|
6.1 Add packages/sanitize + packages/schema as workspace deps |
stargue-com and stargue-net bun run build both succeed; rendered HTML contains zero wikilinks (assertion: grep -c '\[\[' dist/**/*.html = 0) |
| 6.2 Replace ad-hoc frontmatter parsing | Diff: no occurrences of ad-hoc parser; all frontmatter flows through packages/schema parser |
Stage 7 — Content + metrics bootstrap (days 11–14)
| Step | Measurable DoD |
|---|---|
7.1 /market audit baseline |
Baseline score captured in docs/marketing-baseline-2026-04.json |
| 7.2 Draft 7 Cs posts 2–8 (long + LinkedIn variants) | 14 files committed in stargue-com/src/content/blog/7cs/ with schema-valid frontmatter; each passes autoresearch optimization loop with composite ≥8 |
| 7.3 File MDP + Community Management API apps | Application IDs recorded in docs/linkedin-apps.md |
| 7.4 First scheduled post lands on stargue.com + LinkedIn | publication.external_url returns HTTP 200 from unauthenticated curl; stargue.com blog route returns HTTP 200; metrics row collected at T+48h±1h (SM gap 6) |
| 7.5 DR test | pg_dump to Neon succeeds and restores to verifiable row count |
5. Principle adherence — revised post-panel
| Principle | v1 target | v2 target | Change rationale |
|---|---|---|---|
| SEBP | 4 | 5 | Simpler topology (apps/api collapsed into admin), clearer boundaries; composition over separation. |
| SSoT | 5 | 5 | Unchanged (panel consensus). Name-collision fix in 0.10. |
| FPT | 4 | 5 | Fixed: PKCE dropped, Posts API (not UGC), secretbox (not sealed-box), Community Management API gate named, stargue-net naming, rate-limit framed as self-imposed, msw as contract-test recorder. All per dev + architect verification. |
| DiDSP | 5 | 5 | Idempotency_key + advisory lock + CSRF + kill-switch added. |
| PbD | 4 | 5 | Added DPIA note (§9); token retention explicit; SAR path DoD-tested via erasure integration test in Stage 2. |
| OF | 5 | 5 | Unchanged (panel consensus 5 on OF). |
| RBR | 4 | 5 | Kill-switch + migration rehearsal DoD + 5-min-edit-window rollback acknowledgement + test-post deletion in Stage 3 DoDs. |
| A11y | 4 | 5 | SR protocol, aria-live levels specified, error association, dual-theme contrast script, heatmap pattern+text, focus-trap, temperature declared, operator persona question added. |
| Testability | 4 | 5 | All DoDs measurable, msw recorder named, concurrency+rate-limit+idempotency tests specified, optimizer pure-function seam, Vitest+Bun setup included in Stage 1. |
6. Critical files / paths (unchanged from v1 + additions)
[Existing list preserved]
Additions:
test/corpus/private/— sanitizer test SSoTdocs/linkedin-apps.md— Dev Portal app IDs + approval statesdocs/a11y-checklist.md— manual A11y checklistdb/migrations/README.md— migration runbook (expand-then-contract sequences)docs/marketing-baseline-2026-04.json—/market auditbaseline
7. Verification (expanded per TEA gaps)
- Unit: Vitest + Bun workspace. 100% line coverage on pure cores.
- Integration: real Postgres (Docker), real Redis,
mswfakes for LinkedIn. Frozen clock via@sinonjs/fake-timers; seeded RNG. - Contract: msw-recorded LinkedIn API exchanges at
test/fixtures/linkedin/*.har. Re-record policy: rotate fixtures monthly or when LinkedIn-Version header bumps (TEA gap 2). Stale-fixture detector: fixture file mtime > 45 days → CI warning. - E2E: Playwright +
@axe-core/playwright+ NVDA scripting on Windows GH Actions runner. - Security:
bun auditin CI; gitleaks pre-commit; Dependabot. - Load: k6 at Phase 1 exit — 50 queued publishes drained within SLA.
- DR: weekly pg_dump-to-Neon rehearsal with row-count verification.
8. Authorization gates — ANSWERED
| # | Question | Answer (2026-04-19) |
|---|---|---|
| 1 | Stargue LinkedIn Company Page URL | https://www.linkedin.com/company/2605890/ — URN urn:li:organization:2605890 |
| 2 | Cadence bootstrap | Confirmed: Tue + Thu ~08:30 AM AST |
| 3 | Admin auth | Shared Authentik instance (same for all Stargue services) |
| 4 | MDP signatory | Angelo as individual; Stargue entity later |
| 5 | Postgres + Neon | PaaS cluster + nightly pg_dump to Neon free tier (RPO ~24h) |
| 6 | 4h/week cap override | User override: concentrated full-time build, timeline risk accepted |
| 7 | BMAD install location | DQMS charters referenced by path; no install in Publishing Engine repo for this phase |
| 8 | Primary operator persona + A11y context (UX gap 8) | Option C (2026-04-19): primary = Angelo (VoiceOver-capable); secondary = Claude Code agents (programmatic). A11y target: human-first WCAG 2.2 AA; agents inherit semantic markup for free. |
9. Risks (expanded per panel)
| Risk | Mitigation | Raised by |
|---|---|---|
| LinkedIn denies MDP | 60-day re-auth operational baseline; proactive T-5d notification | plan v1 |
| LinkedIn denies Community Management API | Personal-profile posting unblocked; org-posting tools built but gated until approval | dev |
| LinkedIn algorithm suppresses automated-posted content reach | Monitor engagement delta between automated vs manual-posted baselines for first 4 weeks; if suppression detected, fallback to draft-and-prompt workflow via admin | analyst |
| LinkedIn suspends app (ToS) | Conservative rate limit; approval on every publish; no messaging; no engagement automation; own-content only | v1 |
| Refresh token revoked mid-series | Fail-safe halt + Telegram notification within minutes | v1 |
| Refresh rotation race | Advisory lock on refresh path (TEA gap 3) | tea |
| Duplicate post via retry | Idempotency key on publications + BullMQ jobs (dev gap 4) | dev, tea |
| Dokploy outage during scheduled post | DLQ + exponential backoff + alert | v1 |
| Private vault leak via sanitizer gap | 3-layer fail-closed + CI corpus (sanitize gap 8) | v1, TEA |
| OneDrive sync conflict | Scheduler reads from Postgres, not OneDrive; checksum on ingest | v1 |
| Admin PII leak | No PII in dashboard; Authentik SSO; audit view | v1 |
| Bad LinkedIn post past 5-min edit window | Runbook: delete via linkedin_delete_post tool; audit; post-incident review documented in docs/incidents/ (architect, SM) |
architect, sm |
| Scope creep | Stages gated by measurable DoD; Phase 2 = separate plan | v1 |
| 4h/week → full-time override creates personal-capacity risk | User-owned timeline risk; checkpoints at Stage 3, 5, 7 to re-evaluate (PM gap 1) | pm |
| MDP + Community Management API approval timeline unknown | Plan works without either; track application IDs in docs/linkedin-apps.md |
pm |
| DPIA for OAuth-token processing of natural person (Angelo) | DPIA recorded in docs/dpia/linkedin-oauth.md — purpose (publication automation), data (encrypted tokens only), retention (lifetime-of-authorization), lawful basis (contract-performance + legitimate interest), user rights (erasure via token revoke + hard-delete row) (TEA gap PbD) |
tea, sm |
10. Out of scope (deferred to Phase 2 with separate plan)
- Multilingual routing + hreflang (PAP/NL/ES) — F-021. Tradeoff explicit: 7 Cs is EN-only at launch; PAP-speaking stakeholders see EN during bootstrap. Phase 2 adds language-aware routing (Analyst + PM gaps).
- Medium, Substack, ResearchGate, X, Reddit, Telegram, TikTok, Facebook, WhatsApp, Newsletter-email outlets — F-013, F-033
- Inlet Engine — F-030, F-031 (separate PRD)
- Cross-platform analytics beyond LinkedIn + stargue.com/net
- Higgsfield media pipeline
- Library-vs-MCP decision (scheduler uses library directly; MCP reserved for Claude Code — if future needs require HTTP transport, separate plan)
11. Open questions — ALL RESOLVED
All pre-execution gates answered 2026-04-19. Proceeding to Stage 0.
Related
- Publishing Engine PRD
- 7 Cs LinkedIn Series - Continuation Plan
- Design Philosophy - The Nine Laws of the Hearth
- Outlet Profiles/LinkedIn
- Implementation Progress
Sources
- LinkedIn Programmatic Refresh Tokens (Microsoft Learn, retrieved 2026-04-19)
- LinkedIn 3-Legged OAuth Flow (Microsoft Learn, retrieved 2026-04-19)
- LinkedIn Posts API (Microsoft Learn) — current replacement for deprecated UGC Posts
/home/devuser/projects/dqms/.clinerules/12-foundational-principles.md/home/devuser/projects/dqms/.claude/hooks/gate-plan-exit.sh/home/devuser/projects/dqms/.claude/skills/bmad-plan/SKILL.md/home/devuser/projects/dqms/_bmad/bmm/agents/*.md— 7-agent charters (SSoT for panel roles)- Panel verdicts v2:
/tmp/bmad-panel-verdicts-publishing-engine-phase1-20260419.json