v2.3.0

The Bash channel release.

The empirical audit behind 2.2 showed Bash carries 48.3% of all tool-result tokens (Read: 35%), results over 2k tokens are a quarter of bash output, and every real duplicate file read happened through cat/sed/head, invisible to Read-tool hooks. This release governs that channel.

Added

• Bash output governor (PostToolUse): oversized stdout is condensed structurally per command family (grep floods keep first matches per file + counts; git show keeps header + per-file diff stats; file re-prints keep head+tail) and the tool result is replaced via updatedToolOutput before it enters context. Full output always preserved verbatim at .wolf/cache/bash/<id>.log with a factual pointer. stderr never modified; test/build output never replaced by default; condensation only when it saves 30%+. Config: openwolf.bash.governor.
• Measured at the rewrite point: per governed call, original vs entered tokens. Unique ground truth - the platform's own telemetry logs tool output before hooks run, so only the rewriting hook can measure what actually entered the context window. The dashboard hero tile now shows tokens verifiably kept out of context.
• Bash-channel read dedupe: simple single-file reads (cat/head/tail/sed -n) are parsed and tracked; a repeated full cat of an unchanged file gets a short factual advisory.

Changed

• The pre-Bash suggestion filter now exempts only genuinely shaped commands (was matching 1 of 2,256 real commands).

Run openwolf update to migrate (backup + install verification included).