Repository navigation
Releases
v0.3.3
Flow Lean 0.3.3 refines the instructions for concise replies while preserving requested detail, evidence, and authorization boundaries.
- Aim for about 80 words in ordinary replies, expanding when the request or necessary execution details require it.
- Prefer short prose, avoid duplicate recaps and repeated permission requests, and use tables or diagrams when they help. Sequence diagrams cover runtime exchanges.
- Keep unknowns at their supplied scope and distinguish the assistant's voice from messages drafted for the user.
- Require rollback plans to preserve concurrent edits and prior absence, coordinate all writers or stop, and retain verification records.
The README, changelog, and evaluation guide reflect these rules. The regression catalog grows from 22 to 25 cases. Claude and Codex share the same portable skill source.
Validation and known limits
GitHub's plugin load check passes on the merged commit. Isolated Claude installation matches the tested source, and the three new focused text checks pass on both hosts after correcting the first Claude rollback response.
Local BM25 routing passes 20/20 in-corpus checks and 6/9 additional probes per host. Both Codex full-battery calls timed out; native Codex plugin installation was not tested for this candidate. Known factual limitations remain, and generated rollback commands were reviewed as text rather than executed. These checks do not establish a full factual pass or a new compression benchmark.
The evaluation report links the retained responses and failures.
v0.3.2
First GitHub release for the existing v0.3.2 version, recorded in the changelog on September 7, 2026.
Changes in v0.3.2
- Documents Flow Lean as model-agnostic output control, with native installation paths for Claude Code and Codex.
- Clarifies that discovery, persistence, commands, and automatic routing depend on the host.
- Adds a Mermaid runtime diagram with an ASCII fallback and updates the distributed descriptions.
The skill and evaluation battery are unchanged from v0.3.0. See the versioned changelog for earlier changes.
Verification
Plugin load check passed on release commit 76345bfda088bdfb4a438914e7b4d370d332c71f, validating the evaluation catalogs and Claude Code plugin installation and enabled status. This does not establish fresh-session behavior or Codex runtime behavior.
The current battery contains 17 cases. Published comparison results cover the historical v0.2.0 suite of 13 cases; no new comparative benchmark is claimed for v0.3.2.