Repository navigation
Releases
Releases · shanemhamilton/baton
Baton v2.2.0
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
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
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/reportor/api/usersas missing files, in both the inline Step 8 block andevals/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.shruns a request throughclaude -p(six turns) orcodex execin 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
/batonruns are excluded becauseclaude -pdoes 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'shandoffskill won 3 of the 5 Codex misses, so on Codex invoke$batonexplicitly when the choice matters.
Baton v2.0.0
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.shwhen 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/skillsand.claude/skillsfor Claude Code,~/.agents/skillsand.agents/skillsfor 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
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
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.