Releases: yongjip/mergetrain
Release list
mergetrain 3.0.7
- Replace the removed duplicate-enqueue option in error messages with supported
status and retry guidance. Identify the owning job when branch changes, failed
gates, or semantic conflicts require a clean repair commit and a retry. - Align agent instructions and recovery documentation with the existing retry
workflow, including the distinction between queued, validated, and deployed. - Include reproducible parallel-development and daemon lifecycle experiments;
these diagnostic records do not establish a general speed improvement.
mergetrain 3.0.6
-
Make the benchmark runners use platform temporary directories and make
their synthetic subprocess test portable across Linux, macOS and Windows. -
Clarify existing-queue handoff even when only one branch is ready. Queue
counts alone do not establish health, runner ownership, or recovery needs. -
Add the current v3 status, diagnostic and job-evidence command reference to
generated agent guidance. Identify removed doctor syntax and preserve the
distinction between explanation, inspection and execution authority. -
Record a frozen 32-run paired operator-guidance comparison and the earlier
rejected discovery-copy experiment. Public discovery descriptions remain
unchanged; this release does not claim automatic-activation improvement.
Validation: paired operator-guidance diagnostic; normal integration gates and cross-platform CI. Discovery metadata is unchanged.
mergetrain 3.0.5
-
Update the dashboard build's Browserslist dependency to 4.28.9 and refresh
its compatible browser-data dependencies, resolving the reported cache-growth
and custom-stats advisories without changing the packaged dashboard bundle. -
Fix dashboard and Hub snapshots remaining stale after SQLite reuses WAL
space. Share one read-only change observer per cached repository, detect
concurrent commits and database replacement, and release observers on Hub
removal or server shutdown without pinning a read transaction. -
Preserve diagnostic truth across human and structured output: distinguish
semantic conflicts from textual merge conflicts, show blocked reasons in
inspect, and always render repository health plus a next action in
status. -
Make the Python quickstart gate checkout-clean by default, include dirty
paths when a gate changes the assembled tree, and protect the interactive
deployment approval path with targeted tests and a critical coverage floor. -
Remove design PNGs from the current source tree, keep historical
visual evidence linked to the immutable v3.0.4 tag, and exclude design-only
assets from source distributions.
mergetrain 3.0.4
Full Changelog: v3.0.3...v3.0.4
mergetrain 3.0.3
Full Changelog: v3.0.2...v3.0.3
mergetrain 3.0.2
Full Changelog: v3.0.1...v3.0.2
mergetrain 3.0.1
Full Changelog: v3.0.0...v3.0.1
mergetrain 3.0.0
A smaller, durable product grammar
mergetrain 3.0.0 reduces the everyday product to six lifecycle verbs:
init → status → enqueue → validate → deploy → inspect
The default help, generated agent instructions, Claude plugin, MCP adapter,
quickstart, convenience scripts, and LLM reference files now use that same
grammar. Internal queue states are projected into five operator states:
Waiting, Running, Ready, Attention, and Done.
What changed
deploynow owns the full human flow: validate if needed, show one exact
plan, confirm, recheck, atomically push, and verify.- Normal users no longer handle train IDs, SHAs, plan hashes, preview modes,
or reuse switches. - MCP is reduced from twelve tools to five: status, inspect, validate, enqueue,
and deploy. - Starter config contains only the schema version, project name, and gates;
safe defaults remain built in. - Removed v2 commands fail with a typed migration error instead of silently
behaving differently.
Long-lived compatibility commitment
Contract 3 freezes the six command names, five MCP tool names, config-schema-2
readability, error envelope, and existing JSON meanings. Future 3.x releases
are intended to be additive and there is no planned successor command grammar.
All existing exact-SHA, combined validation, semantic-conflict attribution,
destination and policy binding, atomic-push, and durable-recovery guarantees
remain intact.
See the changelog
for the complete migration details.
mergetrain 2.4.2
Release trust-chain safety patch. Production publishing is now dispatched from protected main, verifies the signed annotated tag against main's signer policy, requires the tagged commit in main history, pins downstream work to that exact commit, separates verification/build from OIDC authority, and requires an immutable GitHub Release. TestPyPI and production environment policies are main-bound. See CHANGELOG.md for details.
mergetrain 2.4.1
What changed
- Normal ready-checked enqueue now rejects copied
--base-shaor--head-shavalues when they do not exactly match the current integration ref and clean task branch. Ordinary agent handoff continues to capture both SHAs by default. - Release publishing now requires an SSH-signed annotated tag trusted by the repository's tracked allowed-signers file before artifacts are built or PyPI credentials are exposed.
- The adoption harness distinguishes launcher/setup failures from agent behavior failures and its documented macOS launcher is independent of the current worktree directory.
Evidence
- Controlled Codex trials against the 2.4.1 candidate: 3/3 safe handoffs.
- Author-external PyPA
sampleprojectpilot: task implemented and tested, one exact-SHA manual enqueue, no runner/push/cancel mutation. - Local release gate: 583 tests passed plus critical coverage checks.
- GitHub CI passed across lint, package/sdist, MCP clean-install handshake, E2E, dashboard, and Python 3.10-3.14 on macOS, Ubuntu, and Windows.
See CHANGELOG.md for the full details.