Faster AI-agent foundations
Porcupine v0.1.64 reduces work across model streaming, the agent loop, session persistence, Safe Autonomy, and terminal rendering while preserving ordering, cancellation, replay, approval, and fail-closed behavior.
Highlights
- Faster model streaming: Anthropic, Bedrock, and OpenAI Completions now use stable O(1) content-block indexing instead of repeated linear scans. The OpenAI regression workload performs 12.4x fewer identity probes.
- Smoother terminal UI: Box and Container render each child once on changed frames, stable truncation is cached, trailing trim is single-pass, and unchanged frames emit zero redundant cursor writes.
- Lower agent overhead: no-handler context hooks avoid redundant cloning and dispatch, session persistence avoids repeated history scans and timer churn, and Safe Autonomy formats plan context once.
- Cheaper task-graph updates: repeated active-tool continuations use a zero-allocation fast path; changed steps use targeted copy-on-write updates.
- Safer transcript conversion: incremental conversion is restricted to the proven element-local built-in converter. Custom context-dependent converters always receive the complete transcript.
- Clearer identity: the README now introduces Porcupine as a terminal AI agent.
Verification
- Full GitHub CI build, check, and test workflow passed on the release commit.
- TUI: 826/826 tests passed locally.
- Affected AI, Agent, and Coding Agent suites passed with deterministic regression coverage.
- Signed tag:
v0.1.64.
Install or update:
npm install -g @porcupineai/porcupineai@0.1.64