Work log: 2026-03-27

What shipped today

Image upload infrastructure — Cloudinary end-to-end

The day’s big push was replacing the previously-reverted Vercel Blob upload system with Cloudinary, which was already in use by the engine for DALL-E images. This eliminated a multi-day blocker (Blob store was never provisioned) and gave the whole product a consistent image storage story. The upload route (/api/upload/image) validates MIME type and size, streams to Cloudinary under authexis/user-uploads/{workspaceId}/, and returns a CDN URL. The shared ImageUploader component wraps the whole flow in a click-to-upload zone with thumbnail preview and inline error state.

With the upload primitive in place, three features landed in quick succession: featured images on content detail (replaces the static <img> with a live uploader, optimistically updates the UI), brand reference images in workspace settings (5-slot grid of Cloudinary thumbnails that feed into the AI image generation prompt as visual references), and image_url on social posts (DB column, edit panel in the queue UI, thumbnail on queue cards). The social post image work was the most complex — it required wiring image publishing into all four platform integrations. Bluesky and Mastodon got full blob-upload paths (Bluesky requires uploading the image bytes and embedding via the AT Protocol spec; Mastodon uses /api/v1/media). LinkedIn uses the Community Management API’s asset registration flow. Threads is the simplest — it accepts a public URL directly via the IMAGE media type.

Sentry + engine reliability sweep

Five Sentry-sourced issues closed today. A schema mismatch (c.style column never existed) was generating 855 errors per day — fixed with a migration and the IF NOT EXISTS guard. A dead loop in notifications.py was trying to iterate slots.articles, a field removed in an earlier refactor — one-liner fix. The MCP server was propagating ClientDisconnect as 500s — wrapped at the ASGI boundary to log at INFO and return cleanly. The claude.py service got a proper rate-limit backoff (tenacity, 5 attempts, exponential backoff on RateLimitError) plus a concurrency semaphore (asyncio.Semaphore(5)) to cap parallel Claude calls. Sentry user context was added to the engine’s process_command() so errors now show affected user counts instead of anonymous events.

A pass over eight engine files added exc_info=True to existing log calls and replaced bare swallows with log.warning(..., exc_info=True) — the fix for the widespread (No error message) entries that were making Sentry useless for Python errors.

Completed

• #1745 Add Cloudinary upload route + shared ImageUploader component
• #1746 Wire ImageUploader into content detail featured image section
• #1738 Brand image style library — upload visual reference samples for AI post styling
• #1747 Add image_url to social posts: DB column, queue UI, and platform publishing
  • #1756 DB migration + web layer for social post image attachment
  • #1757 Engine SELECT + Bluesky + Mastodon image publishing
  • #1758 LinkedIn + Threads image publishing
• #1750 column c.style does not exist — schema mismatch breaking queries (AUTHEXIS-14)
• #1748 BriefingSlots object has no attribute ‘articles’ (AUTHEXIS-18)
• #1752 Escalating starlette.requests.stream error (AUTHEXIS-1C)
• #1749 Anthropic 429 rate limit errors — add backoff and request queuing
• #1753 Add Sentry user context so errors show affected user count
• #1751 Add descriptive messages to Python exception handling
• #1274 RateLimitError/429 Sentry issue — closed (fixed by #1749)
• #1754 GitHub API 404 — closed (spurious, token/repo fine)

Release progress

• v1.5: 49/50 closed — one backlog item remaining (#743 dashboard redesign), effectively complete
• v2.0–v2.2: all closed (43 issues total)

Carry-over

None. Queue is empty — no ready-for-prep or ready-for-dev issues. All actionable Sentry errors addressed.

Risks

• Supabase CLI migration sync is fragile — local and remote can drift, causing supabase db push to fail. Workaround is the Supabase MCP apply_migration tool. Upstream issue remains.
• LinkedIn image publishing uses the Community Management API (/rest/posts) which requires the w_member_social scope. If any workspace’s LinkedIn token was issued before this scope was added, their image posts will silently fall back to text-only.

Flags and watch-outs

• CLOUDINARY_URL was only in the engine .env — added to all three Vercel environments (production, preview, development) today. If a new environment is provisioned, remember to add it.
• Hashtags are stored as TEXT (space-separated), not TEXT[]. Keep this consistent.

Next session

Queue is empty. Backlog candidates for the next cycle:

• #927 — Dockerfile uses Python 3.11; dev environment likely uses 3.12+. Low risk, easy win.
• #743 — Dashboard redesign (v1.5’s last open issue). Needs product taste call from Paul before speccing.
• #742 — Alternative layouts. Also needs taste discussion.

Start with #927 (Dockerfile upgrade) if autonomous — it’s concrete and self-contained. For anything bigger, pull Paul in first.