github.devotd.
ReleaseCoding agents

opensec-git/pi-bolt bolt-v0.9.5 — request serialization and event delivery path changes

opensec-git/pi-bolt bolt-v0.9.5 reuses request serialization results in long conversations and reduces the streaming delivery path and model configuration reads.

Repositoryopensec-git/pi-bolt ↗

Pi, the coding agent, compiled ahead of time to native code. Starts 3× sooner than Pi on Bun and uses about a third of its CPU.

stars
★ 372
Forks
16
Language
TypeScript
License
MIT
Latest release
bolt-v0.9.5
Read original source ↗View on GitHub ↗

opensec-git/pi-bolt is a project that precompiles the coding agent Pi into native code. Released on October 19, 2026, bolt-v0.9.5 reuses request serialization results in long conversations, and reduces the streaming event delivery path and model configuration reads. It also fixed a resource lookup issue in the built-in extension and a duplicate instance issue in plugins. These changes are recorded in the GitHub release bolt-v0.9.5 of opensec-git/pi-bolt.

According to the repository description, Pi-Bolt starts three times faster than Pi running on Bun and uses roughly one third of the CPU. This is the project's comparison claim as stated in the repository description. What changed in this release is the reuse of request serialization results, the streaming event delivery path, model configuration reads, resource lookup in the extension, and plugin instance handling. What the release summary confirms is a change in the execution path and in configuration/plugin handling; individual performance figures for each fix are not given.

The performance comparison used a local mock model, not an actual model. Therefore, its results cannot be taken as evidence of model inference acceleration. Reusing serialization in long conversations and changing the event delivery path are changes to the processing paths described in the release, but this information alone cannot speak to the response speed of an actual model or the size of the improvement. The changes in this version concern how the code is processed and resource/instance issues, and there is no basis for asserting that model inference itself has become faster.

More in this field

Coding Agent CLIs ↗