Authexis work log — 2026-03-26

What shipped today

Hashtag pipeline — seed, edit, generate

The social queue now has a complete hashtag pipeline. The first piece was the seeding logic: when content gets added to the social queue, addContentToQueue now extracts any inline #hashtags baked into the LLM-generated post body, strips them from the visible text, and writes them to the social_posts.hashtags column. If the post has no inline hashtags, it falls back to the source content’s tags array formatted as #tag. This resolves the mystery of “my posts already have hashtags” — they were baked inline by the LLM all along, and now they’re a proper first-class field.

The second piece was the UI: every social queue post card now has an inline HashtagEditor component that renders chips with × delete and a + add input. Users can add or remove hashtags without leaving the queue. The third piece was AI generation: a new engine handler social_post.generate_hashtags prompts Claude for 3–5 relevant hashtags given the post body and source content context, writing the result to social_posts.hashtags. addContentToQueue dispatches this command automatically when no hashtags were seeded from inline tags or content tags. Together these three changes (#1742, #1743, #1744) close the loop on hashtag management.

Guided content refresh

Content drift is a real problem — a piece written three months ago about a pending regulation is wrong the moment the regulation passes. The new “Refresh content” flow gives users a frictionless way to handle this. A “Refresh content →” link now appears on content in review or final status. Clicking it expands an inline panel (no modal) where the user describes what changed. On submit, the refresh note is persisted to a new contents.refresh_notes column, a content.refresh_requested event is logged in history, and three content.redo_field commands are queued for draft, intro, and title — each carrying the user’s note as revision context. The existing redo_field handler already supported a revision_notes payload that prepends a “REVISION REQUESTED:” header before generation, so no engine changes were needed. One migration (#1736).

Image work — shipped then reverted

A Vercel Blob upload route and ImageUploader component were built and shipped (#1745, #1746), then reverted after realizing Vercel Blob hasn’t been provisioned for this project. The code was clean and the approach was sound, but pushing a /api/upload/image endpoint that returns 500 in prod was the wrong call. The issues are reopened and parked until the Blob store is set up.

Pipeline housekeeping

Continued running /issue next --auto throughout the day. #1735 (hashtags, too big) was decomposed into three children; #1737 (image uploads, even bigger) was decomposed into three children with a clear dependency order. Several issues were batch-prepped in background agents running in parallel. The pipeline is in good shape heading into tomorrow with a clear queue.

Completed

• #1722 — distribution queue guardrails (previous session, shipped at session start)
• #1736 — guided “refresh outdated content” flow
• #1742 — seed hashtags from content tags when adding to social queue
• #1743 — inline hashtag editor on social queue post cards
• #1744 — AI-generate hashtags engine command (social_post.generate_hashtags)
• #1745 — Vercel Blob upload route + ImageUploader (shipped and reverted)
• #1746 — wire ImageUploader into content detail (shipped and reverted)

Release progress

• v1.5 — 49 closed / 1 open. One lingering backlog item, milestone effectively complete.

Carry-over

• #1747 (ready-for-prep) — Add image_url to social posts: DB column, queue UI, and platform publishing. Needs prep before it’s grindable.
• #1745 (ready-for-prep) — Vercel Blob setup. Blocked on provisioning a Blob store in the Vercel dashboard (Storage → Create → Blob). Once BLOB_READ_WRITE_TOKEN is in env, this can be re-executed quickly — the code was already written.
• #1746, #1738 — both blocked on #1745.

Risks

• Image upload blocked until Blob is provisioned. Three issues (#1745, #1746, #1738) are stacked behind this. The Vercel dashboard action is manual and takes 2 minutes, but it hasn’t happened yet.
• BLOB_READ_WRITE_TOKEN absent from all environments. vercel env ls confirms no blob token exists. If someone tries to run the upload route (still referenced in git history), it will 500.

Flags and watch-outs

• Tester (Jeffrey) received a spurious “outline is ready” email — root cause was a stale notify_content_ready(stage_name="outline") call left over from when outline was a visible stage. Fixed in a previous session by removing the call and adding the correct interview_ready notification instead.
• Hashtags are a TEXT column (space-separated string like "#tag1 #tag2"), not TEXT[]. The issue spec said TEXT[] but the API and DB use plain text. Future work should keep this consistent.

Next session

1. Provision Vercel Blob — go to Vercel dashboard → authexis project → Storage → Create Blob store. Run vercel env pull after to get BLOB_READ_WRITE_TOKEN locally. Then re-execute #1745 (fast, code already written in git history).
2. Prep #1747 (social_posts.image_url + platform publishing) — big issue, needs careful scoping especially for LinkedIn’s Assets API.
3. Continue /issue next --auto — pipeline has ready-for-prep issues to chew through.