Cloudflare CI SDK, Rspack 2.0 ESM, Ollama fixes - Dev Signal #81

This week's releases cluster around two themes: eliminating configuration ceremony (YAML pipelines, manual cache-busting, speculative decoding setup) and patching real production pain (memory leaks, tool-calling failures, dependency bloat). The Rspack 2.0 and Cloudflare CI drops in particular signal a broader push to make TypeScript and ESM the default substrate for developer tooling, not an afterthought.

Cloudflare ships CI SDK built on Workflows

Cloudflare's new @cloudflare/ci SDK lets you define CI/CD pipelines as TypeScript Workflow classes instead of YAML files. Pipelines trigger directly on artifact pushes - no event subscription wiring, no queue boilerplate. Dependency caching and parallel step execution are first-class, and everything surfaces in the existing Workflows observability dashboard.

Why it matters now: YAML-based CI has always been a leaky abstraction. You eventually end up shell-scripting around it anyway. Defining pipelines in TypeScript means real type safety, composable step logic, and the ability for platform teams to publish a single CI definition as an importable module across every customer repo. The native artifact-event trigger also removes an entire class of webhook/queue infrastructure that most teams cobble together.

Verdict: Ship - for new Cloudflare-native projects, there's no reason not to start here. Migration from GitHub Actions is feasible but not a drop-in replacement; budget time for rethinking your trigger model, not just translating YAML. Requires a wrangler config with the events field and an Artifacts repo wired up.

Vercel Blob adds consistent reads for private storage

Pass useCache: false on blob reads to bypass CDN caching and get guaranteed read-after-write consistency within 60 seconds of an overwrite. Without the flag, you get eventual consistency with CDN-edge performance. With it, you pay slightly more (Fast Origin Transfer costs) and accept slower reads in exchange for freshness.

Why it matters now: Stateful AI systems - agent memory files, session transcripts, real-time report generation - break in subtle ways under eventual consistency. Debugging a stale read that surfaces one in twenty requests is miserable. Having a flag that makes consistency opt-in per call is exactly the right API design: you don't pay the cost everywhere, only on the paths where it matters.

Verdict: Ship surgically - add useCache: false only to hot paths where staleness causes correctness issues. Requires @vercel/blob@2.6.1+. Don't blanket-apply it; the cost and latency tradeoff is real.

Ollama fixes cache leak, improves tool calling

v0.32.1-rc0 patches an MLX model cache memory leak, stabilizes Gemma 4 tool-response continuations across multi-turn interactions, and adds working directory context to agent invocations.

Why it matters now: The MLX cache leak is a genuine production issue for anyone running persistent agent processes - memory creep in long-lived sessions compounds quickly and forces restarts that break continuity. The Gemma 4 tool-calling fix matters if you're building multi-turn reasoning chains; broken tool-response continuations are hard to detect and produce subtle logic failures rather than loud crashes.

Verdict: Ship - this is a drop-in patch upgrade from v0.32.0. If you've observed memory growth in multi-request sessions or use Gemma 4 for tool-calling workflows, upgrade now. No breaking changes for other model users.

Qwen3.5 faster on Apple, OpenAI streaming parity

The MLX engine in `v0.32.6-rc