Skip to content

v2.0.0 — Progressive Disclosure + Agent Discipline + Pass Rename

Latest
Latest

Choose a tag to compare

View all tags
@REMvisual REMvisual released this 21 Apr 23:48
v2.0.0
b1adf0c

Major refactor focused on cost efficiency and correctness.

Highlights

  • ~30% lower skill-load cost — monolithic skill.md split into a 2,700-token core + four on-demand references/ files. You pay only for what the current mining pass needs.
  • Agent discipline saves 15–45K tokens per run — Step 1A (git/beads/ls) is now parallel Bash only. Previous "judgment based on complexity" wording allowed subagent bootup on cheap external-state checks, which was pure waste.
  • Phase 1 baseline enforcement — Phase 1 MUST hit the pass minimum on first Write; Phase 2 is strictly for gap-filling. Eliminates the 210-line-then-expand-to-300 anti-pattern.
  • Pass rename: Tier 1/2/3 → Quick/Deep/Chunked — numeric tier names collided with chain seq N (real bug where users and the model both misread "Tier 2" as "seq 2").
  • Stricter chain detection — a shared bead is now a CANDIDATE, not automatic proof of continuation. Prevents false-positive chains when one bead hosts multiple work streams (brainstorm → impl → testing).
  • Auto-handoff filter — precompact hook-generated files are tagged **Auto:** true and excluded from chain-parent scanning.

Migration

Re-install from source to pick up the new references/ directory:

rm -rf ~/.claude/skills/handoff
cp -r claude-handoff/skills/handoff ~/.claude/skills/
cp -r claude-handoff/skills/handoffplan ~/.claude/skills/

Existing installs keep working with the old flat skill file — no hard break, but you won't get the cost savings or improved chain detection until you re-install.

Full changelog

See CHANGELOG.md.

Assets 2
Loading