Agent Blueprint
Get started
Browse the pages

DocsHelp

Tested in each tool

Before the release, a smoke test ran a sample project through each installed tool, and an evaluation compared v4 with v3 on the same tasks. Each tool’s support note on GitHub has its own smoke-test results:

Before the release, each of the eight tools installed Agent Blueprint by its own route in a clean profile and listed every skill exactly once (docs/releases/v4.1.0-installs.md). A local smoke test (tests/smoke/) then ran a sample project through each tool’s headless mode. It covered discovery of every skill, the manual-only rule, hooks, debugging, and two checks for this release’s skill fixes (existing code kept under TDD, and an approved design taken as approved). docs/releases/v4.1.0-smoke.md has the table with times and reasons, and each support note under docs/hosts/ repeats its tool’s results.

Tool Version tested Result
Claude Code 2.1.296 Installs with each skill once; every check passes
Codex 0.160.0 Installs with each skill once; every check passes, and all five hooks run once trusted
Antigravity 1.3.3 Installs with each skill once; every check passes
Cursor CLI 2026.10.01 Installs with each skill once, through Claude Code’s plugin or --plugin-dir; every check passes
Grok Build 1.0.34 and 1.0.52 Installs with each skill once; discovery, hooks, the manual-only rule and debugging pass
Pi 0.73.1 Installs with each skill once by its own route
Hermes 0.19.0 Installs with each skill once by its own route
Amp 0.0.1791360091 Installs with each skill once, through Claude Code’s plugin or its own copy

An evaluation then ran the same fixture tasks on v3.8.0 and v4.0.0 in Claude Code (docs/releases/v4.0.0-eval.md). Build, review and debug passed on both versions, with no task getting worse. v4 is slower and costs more per task (review took 7m21s instead of 3m08s), because a v4 pipeline writes a provenance record, runs reviews as a swarm with a validator, and takes debugging all the way to a regression test. The ship pipeline was not part of the evaluation: one run costs about $40 and nearly two hours, and the smoke test above already shows it passing end to end.