What we shipped recently.

The honest list. Every product change, infrastructure migration, and fix worth telling a customer about. If you want the noise too, the full git history is on git.attevia.dev (coming soon).

1. May 23, 2026·v0.9.3 feature

   Memo library — search every memo your desk has ever signed.

   A new /app/memos view with full-text search across thesis, memo title, and tickers. Filter by thesis stance (supported / mixed / not supported), status, and signature presence. Sort by Sharpe, ann. return, conviction, or date. The first feature most analysts open in the morning.

2. May 23, 2026·v0.9.2 ui

   Animated pipeline reel on the landing page and console.

   A 22-second auto-playing visual walkthrough of the agent pipeline — thesis input, six-agent fan-out, quant tournament, signed memo, verifiable receipt. Pauses when offscreen. Replaces the static terminal preview on the hero and now lives at the top of /app too.

3. May 23, 2026·v0.9.1 fix

   Equity-chart range slider rewritten with real mouse + keyboard + touch support.

   The previous overlapping <input type="range"> trick broke on Safari and was useless on touch. The replacement is a proper dual-thumb component: drag either handle to move one endpoint, drag the filled band to translate the whole window, arrow keys to nudge by one step (Shift for 5%), Home / End for the extremes. Plus five evenly-spaced date ticks under the track.

4. May 23, 2026·v0.9.0 ui

   Sticky left column on the thesis result page.

   The agent trace and strategy card now pin to the viewport on desktop, with their own thin internal scrollbar, so the right column (metrics, chart, memo) can scroll freely without leaving a tall blank gutter. Bloomberg-terminal pattern, finally applied.

5. May 22, 2026·v0.8.6 infra

   Upstash Redis is now the primary persistence backend.

   After our Vercel Blob store was suspended mid-week, we built a routing layer that prefers Upstash Redis (zero-setup, free tier, sub-15ms round trips) and falls through to Supabase Storage and Vercel Blob. Every run, receipt, and API key now writes to whichever backend the environment exposes — no manual setup for users. /api/diagnostics/blob always reports which backend is active.

6. May 22, 2026·v0.8.5 feature

   Per-run configuration — currency, lookback, variants, arXiv depth.

   The desk now has an Advanced disclosure that lets you override the lookback window, the number of strategy variants the quant tournament explores, the arXiv search depth, the display currency (INR, USD, EUR, GBP, AED, SGD, JPY), and even the model used by each agent. Tune per-run; defaults persist via /app/settings.

7. May 22, 2026·v0.8.4 model

   GPT-5.5 promoted to consensus + fallback tier.

   With the OpenAI budget under-utilised and GPT-5.5 measurably stronger on the consensus task than GPT-5.4, every run now ships a GPT-5.5 second-opinion against the Claude Opus 4.7 primary. The Critic still runs on Sonnet 4.6 + extended thinking; the Scribe still drafts on Opus 4.7.

8. May 22, 2026·v0.8.3 feature

   Model provenance recorded on every run.

   Every ThesisRun now carries a models field with the exact model id that touched the Researcher, Critic, Scribe, Consensus, and signing step. Surfaced on the run page in the new "Built with" card — no more "powered by AI" hand-waving.

9. May 22, 2026·v0.8.2 feature

   Browser Use is now a fallback for arXiv rate-limits.

   When direct arXiv hits a 429 or 503, the Researcher now spins up a Browser Use cloud session that scrapes arxiv.org/search and returns a clean abstracts list. Configure with BROWSER_USE_API_KEY. Also available as an interactive panel on the desk for arbitrary web automation.

10. May 21, 2026·v0.8.1 fix

    Critic JSON parsing fortified with a three-level retry.

    Critic failures used to leak into the executive summary ("The critic output failed to parse…"). Now: Sonnet 4.6 + thinking → Haiku strict-JSON → OpenAI JSON-mode → generic baseline. We always return a useful red-team result, even when the primary model misbehaves.

11. May 21, 2026·v0.8.0 feature

    Smart date parsing — "from 2020 to 2024" now actually means it.

    When your thesis contains an explicit date window in natural language, the backtest honours it. A pill at the top of the equity chart shows which window the thesis named, so it's obvious when the engine is respecting your range vs falling back to the default 7-year lookback.