Skip to content

v2.3.0 — README discipline, skill refactor, resume fixes

Latest
Latest

Choose a tag to compare

View all tags
@sebmartin sebmartin released this 04 Jun 14:26
· 11 commits to main since this release
v2.3.0
cac9048
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's new in v2.3.0

README discipline

Thread READMEs are now a lean index — not a content dump. A fixed-section template (replacing the old Problem/Current State/Desired State/Existing Infrastructure/Resolved Questions/Notes sections) with explicit write rules: no new sections, everything substantive goes in a linked artifact or decision. Quick Resume decay rules keep the thread summary tight (5 ideal / 10 max next steps, rolling 3–5 recent progress entries).

Threads skill refactor

SKILL.md split from a 461-line monolith into a 154-line lean index with 7 command reference files loaded on demand — reducing token cost for every non-resume invocation. Resume stays inline since it runs every session.

Resume fixes

Based on analysis of observed resume failures:

  • get_thread_status renamed to resume_thread and now returns the full README (not just Quick Resume), giving the LLM the complete map in one call
  • Session frontmatter (date, summary, keywords, next_context) loaded with a recency gradient on resume for richer context without eager file reads
  • Pre-work confirmation checkpoint before any planning or recommendation
  • Patching/confusion correction rules

Save fixes

/threads save now executes as explicit mandatory steps: README Quick Resume update, README hygiene checks, session log creation/update, session linking. On save, unlogged decisions from the conversation are surfaced and batched into a single prompt.

New get_skill_file MCP tool

Replaces get_template for reading any plugin file (templates and command references) without triggering filesystem permission prompts.

Global permissions setup

When setting a workspace for the first time, the skill offers to write allowlist entries to the global CLI configuration file — no more per-session permission prompts.

CLI detection fix

Detection now uses the LLM's own knowledge of its config file location rather than printenv CLAUDE_PLUGIN_ROOT (which isn't available in bash commands).

Other fixes

  • Codex skill invocation is $threads not /threads
  • MCP fallback rule clarified: fix failures, don't work around them with raw file reads
  • Decision logs created immediately with README links — no orphaned files
Assets 2
Loading