Skip to content

Releases

Baton v2.2.0

Choose a tag to compare

@shanemhamilton shanemhamilton released this 22 Sep 13:04

Highlights

  • Replaces mandatory parallel handoff files with one repository-selected canonical continuation record, plus an exact receiver-start instruction.
  • Adds checkout identity, explicit cleanup authority, worktree and branch ledgers, and receiver-readback proof for cross-checkout continuation.
  • Adds committed multi-worktree receiver coverage and five lifecycle controls for linked-worktree movement, recovery, and retirement safety.

Verification

  • python3 -m unittest evals/test-check.py — 32 tests passed.
  • python3 evals/receiver-eval.py self-test — all controls passed, including multi-worktree.
  • python3 evals/worktree-lifecycle-eval.py self-test — 5 tests passed.

Baton v2.1.0

Choose a tag to compare

@shanemhamilton shanemhamilton released this 05 Sep 21:48

Baton v2.1.0 preserves task context earlier, helps fresh sessions reconcile changed facts, and tests whether receivers complete permitted work safely.

Added

  • Explicit Baton checkpoint requests preserve irreplaceable task context without ending the session. Final handoffs recover essential checkpoint facts and refresh current state.
  • Conditional transfer-readiness guidance records essential local artifacts, repository and checkout mappings, verified receiver access, and pending prerequisites.
  • Public synthetic receiver scenarios for interrupted work, state drift, and deferred consequential actions, with independent outcome and boundary checks.

Changed

  • Receiver boot reconciles current instructions, workspace state, action preconditions, and scoped approvals, revisiting only affected decisions.
  • The first action moves into Section 0 with its expected observation and safe fallback. Intent, boundaries, unfinished work, and decision rationale take priority over optional capability and model advice.
  • Independent challengers reconstruct a safe first action from the handoff and available artifacts.

Fixed

  • Both document checkers reject empty required content, absent truth evidence, and a closing citation that points to another file with the same basename. The inline checker returns a real failure status and scans first-action paths in Section 0.
  • Scenario scoring binds the emitted handoff to its actual contents, final message, skill hash and artifact baseline. Failed authors, unchanged preexisting artifacts, retry reuse and file/message mismatches return failure instead of a successful report command.
  • Receiver grading permits legitimate test extensions and additional explanatory metadata, and checks protected state after executing receiver code.

Verification

  • All 18 local regression tests pass across the repository checker, installed inline checker, and author runner. Receiver-grader self-tests cover valid outcomes, preservation failures, and attempted prohibited actions.
  • A fresh-agent synthetic sequence passed through an explicit checkpoint, final handoff, implementation, second handoff, and verification without repeating completed work or changing protected files.
  • Skill validation, shell syntax, site JavaScript syntax, and whitespace checks pass. Independent reviews were completed and their findings resolved.
  • These checks do not establish comparative model performance, cross-harness parity, or real cross-machine transfer reliability.

Full changes: v2.0.1...v2.1.0

Baton v2.0.1

Choose a tag to compare

@shanemhamilton shanemhamilton released this 02 Sep 18:28

Fixed

  • Low-context handoffs now record each open question as its own claim row with Class Unknown; two authors had left them untagged.
  • The path check no longer flags URL routes such as /v1/report or /api/users as missing files, in both the inline Step 8 block and evals/check.sh.

Changed

  • The description now names "saving where work stands for later" as a trigger and states the precedence rule explicitly: when another handoff skill is installed, including metaswarm's handoff, use baton. On Codex, metaswarm's handoff had won 2 of 20 should-trigger requests.

Added

  • evals/trigger-eval.sh runs a request through claude -p (six turns) or codex exec in a read-only sandbox from a seeded scratch repo that is reset before every run, and records whether the harness invoked baton on its own, with no API key or SDK.

Verification

  • Extended cross-model matrix on the released skill: 31 of 31 author runs (Sonnet, Opus, Fable, and Codex gpt-5.6-luna, terra, and sol across the six scenarios that had only run on Sonnet) produced both deliverables; 30 had zero checker failures; 274 of 275 generic assertions passed.
  • Money-path receiver replay against the final template: 2 of 2 fresh Sonnet receivers stopped at the migration gate with green tests.
  • Harness-native trigger eval (20 requests, 2 reps each, no API key): Claude Code invoked baton in 15 of 18 should-trigger runs and in none of the 20 should-not runs; the two literal /baton runs are excluded because claude -p does not expand slash commands. Codex invoked baton in 15 of 20 should-trigger runs and, in substance, in none of the should-not runs (six flagged rows were Codex reading baton's file to answer questions about baton, or log contamination from an unconfined read-only sandbox). Metaswarm's handoff skill won 3 of the 5 Codex misses, so on Codex invoke $baton explicitly when the choice matters.

Baton v2.0.0

Choose a tag to compare

@shanemhamilton shanemhamilton released this 02 Sep 03:30

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).

Baton v1.1.0

Choose a tag to compare

@shanemhamilton shanemhamilton released this 01 Sep 18:00

Highlights

  • Adds a mandatory Model Efficiency Gate to every nontrivial handoff.
  • Assigns each lane to the least expensive and fastest reliable tier: FAST/Haiku-class for deterministic legwork, BALANCED/Sonnet-class for bounded coding and analysis, and FRONTIER only for judgment-bearing work.
  • Requires a written explanation of why both lower tiers are insufficient for every frontier execution assignment.
  • Prevents DIRECT execution from becoming a frontier-legwork exception.
  • Escalates one tier at a time only after an observable failure or capability gap.
  • Requires boot-time model-route refreshes so Claude and Codex resolve current available models instead of trusting stale IDs.
  • Keeps frontier models focused on architecture, consequential risk, integration, synthesis, and final review.

Verification

  • Release commit: c9bcac3
  • Canonical and installed Claude/Codex skill validation passed.
  • All installed skill copies matched the canonical SHA-256.
  • Embedded website JavaScript syntax check passed.

Baton v1.0.0

Choose a tag to compare

@shanemhamilton shanemhamilton released this 01 Sep 17:54

Highlights

  • Preserves every material result, decision, work-in-progress boundary, recovery state, and next move without transcript dumping.
  • Prevents momentum loss with an exact stop point, do-not-redo boundary, active-work inventory, and executable continuation mission.
  • Ranks installed and installable capabilities, prefers authoritative MCP connectors, and requires provider-primary research before manual fallbacks.
  • Recommends superior missing connectors through a human install-or-defer gate with Claude- and Codex-specific setup and fallback paths.
  • Keeps capable models on orchestration, synthesis, integration, and review while delegating bounded research, mechanical coding, and test execution.
  • Explicitly evaluates efficient-frontier, ponytail, and more specific installed skills for the receiving session.

Compatibility

Baton is a single Markdown skill file for Claude Code, Codex, and other compatible skill harnesses. See the README for global and per-project installation.

Verification

  • Release commit: d955d23
  • Skill schema validation passed.
  • Embedded website JavaScript syntax check passed.
  • Canonical, installed, and remote skill hashes matched before release.