agent-runtime v0.7 — streaming tool calls
Reworked the execution loop so tool calls stream incrementally instead of blocking on the full model response. Added automatic retries with jittered backoff and a circuit breaker for flaky tools.
Alive signals
A hand-written changelog of intentional milestones, refreshed alongside a live feed of my latest public activity. Both are baked in at build time and rebuilt hourly — no spinners, no stale runtime calls.
Reworked the execution loop so tool calls stream incrementally instead of blocking on the full model response. Added automatic retries with jittered backoff and a circuit breaker for flaky tools.
Trading lab now replays historical sessions tick-for-tick from a content-addressed cache. Backtests are reproducible to the byte, which makes regressions trivial to bisect.
Killed every runtime fetch. GitHub stats, activity, and npm badges are all resolved during the build and refreshed hourly by a GitHub Action. Zero API routes, zero ISR, instant cold loads.
Published a piece on constraining tool selection with typed schemas and a deferred-resolution step, drawn from production failures I hit shipping the agent runtime.
Spun up skill-registry: a small, versioned catalog of agent skills with a manifest format so the runtime can discover and load them at boot. Early days, but the ergonomics feel right.
Every animated surface here now degrades gracefully under prefers-reduced-motion — the contribution heatmap and feed reveal instantly instead of sliding in. Accessibility shouldn't be an afterthought.
Spent a week deep in Rust SIMD intrinsics to vectorize the order-book decoder. ~3.4x throughput on the hot path. Notes are turning into a longer write-up.
New Next.js build with a prompt library, agent-skills section, and a photography gallery. Designed to be fast, legible, and genuinely mine rather than a template.
Showing sample activity — live data appears once the build runs with network access.