Skip to content

Baton v2.0.0

Choose a tag to compare

@shanemhamilton shanemhamilton released this 02 Sep 03:30
· 4 commits to main since this release

A rewrite of the skill file from 475 lines down to 210.

Added

  • Rewrite the skill as one 210-line file: seven rules, eight steps, and a single template with three depth labels (COMPACT, STANDARD, GOVERNED).
  • Write the handoff file before asking the human anything, and never end a session without both deliverables — the file and the closing sentence — even on non-interactive, scheduled, and subagent runs. An unanswered question is recorded as deferred by absence instead of blocking the handoff.
  • Make the closing sentence the file's last line, so it's captured on disk, not only in chat.
  • Classify every claim with one of four evidence classes, with no synonyms allowed, and resolve every cited path from the repository root, so a handoff can't point a receiver to a file that doesn't exist.
  • Add a mechanical Step 8 check (an inline bash block, or evals/check.sh when present) that verifies depth ceilings, evidence-class wording, path resolution, the closing sentence, leftover DRAFT markers, model-identifier provenance, and secrets before a handoff is finalized.
  • Replace the four execution shapes with two — DIRECT or LEAN — plus a named-failure escalation for anything heavier.
  • Discover skills per harness: ~/.claude/skills and .claude/skills for Claude Code, ~/.agents/skills and .agents/skills for Codex, and flag when the receiving session runs on a different machine.
  • Add a "Recommended install/connect" field to Section 0 for a missing capability the human should approve.
  • Rewrite the skill description to lead with its trigger conditions, and add a precedence clause for when another handoff skill is installed.
  • Handle low-context sessions and non-git working directories explicitly.
  • State retention and secrets rules directly in the contract: keep docs/handoffs/ untracked by default, and never write credentials or auth state into a handoff.
  • Add a rule for the author to delegate its own state sweep and challenger to a cheaper subagent when the harness supports one.

Removed

  • The four execution shapes (STRUCTURED and METASWARM), used in under 4% of real handoffs.
  • The six-column model-assignment matrix and the capability-activation table.
  • The five-tier connector ladder.
  • The 27-item self-check and the 23-item anti-pattern list.
  • The second template.
  • The Codex-only $-prefixed skill syntax.

Changed

  • Rewrite the README, NOTICE, and project site to describe what v2 actually does, and state the tested harness roster as Claude Code and Codex.

Verification

  • 36 author runs across eight eval scenarios: every v2 run produced both deliverables and passed all checker assertions on Claude Sonnet, Opus, and Fable and on Codex gpt-5.6-luna, terra, and sol. v1.1 passed 83% of the same assertions and missed the closing sentence in one run.
  • Receiver replay: fresh Sonnet sessions with no skill loaded completed the first milestone with green tests on all three v2 handoffs. One receiver exercised a migration's mutating code path against scratch data despite a hard stop, so the template now asks each hard stop to name the exact forbidden action and the nearest allowed one.
  • Description trigger eval: 59 of 60 decisions correct against near-miss requests (v1.1: 55 of 60).