Claude Code themes and Projects upgrade, DeepSeek API tools, CRAM energy model (AlphaSignal 2026-11-16)
Why this is in the vault
Daily AlphaSignal record of two Claude Code product changes we operate on (mods/themes, Projects coordination) plus a DeepSeek API tooling shift, with the CRAM energy-model item as a research watch-item.
Mapping against Ray Data Co
Most concrete: the Claude Code Projects upgrade (coordinator answers in-chat, /fork, /btw, reminders, read-only thread sharing) overlaps with how we already run parallel threads and sub-agent fan-out, and bears on the 2026-11-05 subagent cost-cut item. Mods/themes are cosmetic but extend the plugin surface already tracked in the 2026-10-02 mods note. The WorkOS Airlock pitch (runtime allow/deny/human-approve on agent calls) echoes our approval-gated-call rules, but is a paid placement, so treat as vendor framing. DeepSeek bundling search/file/URL tools into the API lowers glue-code cost for cheap-model fallback routes. CRAM (1985-style energy function, claimed 1,000x training speed, Sudoku in 42ms) is a headline claim only; unverified, no action. Mapping strength: medium-weak, no decision.
Issue contents
Top news (as reported, figures unverified):
- Claude Code desktop custom themes: built via mods (small plugins), animated backgrounds, theme-examples plugin; local sessions only.
- Claude Code Projects upgrade: coordinator answers directly with live status; read-only thread sharing, thread reminders,
/btwside threads,/forkbranching, artifact thumbnails; public beta on Pro and Max. - DeepSeek V4.1-Flash API tools: web search, deep-thinking search, file input (PDF/Office/images), URL reading, safety checks; billed on top of tokens.
- Sponsor slot (WorkOS Airlock): authorization layer for agents (see Sponsorship).
Signals:
- Qwen4-Max: claimed 60% lower cost, 1M-token context.
- Kinference (Buoyant.io): open-source inference serving for platform teams; sponsored signal.
- Open-source tool that reverse engineers codebases into plain business workflows.
- CRAM: 1980s energy model trains reasoning tasks 1,000x faster than looped models.
- Meta study of 200+ models: vision encoders need parameters, not pretraining.
- Open-source 18 MB CPU text-to-speech model, never trained on real voices.
Events: two Snyk-hosted webinars (Dec 9, Dec 17), promo.
Deep-fetches: zero. Links are tracked redirects and the items were thin on RDCO relevance; the CRAM claim would be the first candidate if it matters later.
⚠️ Sponsorship
Paid placements: WorkOS (Airlock agent authorization, main sponsor block) and Buoyant.io (Kinference signal #2). Snyk events listed under "Our Events" are promo. No sender self-promo beyond house "Work With Us" CTAs. Bias: the sponsored items are product pitches and should not be read as AlphaSignal's editorial picks.
Related
- [[2026-10-02-alphasignal-claude-code-mods-codex-security-deepmind-consciousness]]
- [[2026-11-05-alphasignal-claude-code-subagent-cost-cut]]
- [[2026-11-15-alphasignal-nvidia-cut-ai-context-compaction]]