holt ci

Version updated for https://github.com/Raed2180416/holt to version v0.3.0.

• This action is used across all versions by ? repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

The holt action is an agent-agnostic tool designed to help developers manage multiple worktrees efficiently. It compares the committed and uncommitted states of git repositories to identify which worktrees contain valuable content that should not be deleted, ensuring safe deletion practices in a multi-agent development environment. By leveraging git’s own worktree lock, holt automatically prevents accidental deletions of worktrees that hold important code changes.

What’s Changed

holt 0.3.0 — measured, and fixed where the measurement hurt

Know what your agents made, and don’t lose any of it. You ran a dozen agents overnight; holt tells you what each worktree actually produced, which ones collide, which are safe to delete, and it stops an agent deleting work that exists nowhere else.

0.2.0 could act. 0.3.0 is the first release that has been scored — against an oracle proven to share no code with holt — and then fixed where the score was bad.

Install

npm install -g https://github.com/Raed2180416/holt/releases/download/v0.3.0/holt.tgz

One command — no clone, no build. This release is built, installed and driven against a real repository on Linux, macOS and Windows by CI before the file is attached, and the same smoke test now runs on every commit rather than only at release time. holt.tgz is a stable name, so this URL keeps working; the versioned holt-0.3.0.tgz is attached alongside it.

holt is not on the npm registry yet — npm install -g holt 404s — so this URL is the install.

Then, in any repository with worktrees:

holt integrate       # wire every agent you use — the whole setup
holt auto            # locks what would be lost; tells you what needs a decision

Measured for the first time

50 languages, 900 worktrees, 18,000 labelled claims, scored against an independent oracle that shares no code with holt — the independence is proven, not asserted, by a static import-graph walk, a runtime module-resolution hook, and a probe that proves the hook actually fires (independence-check.mjs; full detail and reproduction command in BENCHMARKS.md §9):

False “safe to delete”: 0 of 900, reproduced across four independent runs. duplicate’s precision and conflict’s recall are the two numbers here below 1.00 on purpose — see “Known limits” below and BENCHMARKS.md §§8–9 for exactly which cases they are and why closing duplicate further would mean overriding holt’s own correct, hand-verified answer.

Recall on disposable was 0.40 when first measured, and that is the headline fix of this release: holt was refusing 60% of the work it exists to do. Perfect precision at 0.40 recall is not a safe tool — it is one that answers “I cannot be sure” to most of its own question.

What’s new

• holt auto — the autopilot. Does everything that cannot lose data by itself (locks what is at risk, releases locks no longer justified) and hands the destructive half over with the evidence and the exact command. It never deletes.
• holt discard <path> — the escape hatch. Captures content to a verified ref first, then removes it, so the guard stays on and the loss does not. A tracked file is reverted rather than deleted. Journalled, with the restore command printed.
• Redundancy-aware disposal — a worktree whose content a living sibling also holds is now correctly disposable. clean --apply re-verifies before each removal, so a redundant set drains to exactly one survivor; gate refuses it, because the rm -rf a script runs after gate never looks again.
• The guard speaks Windows — Remove-Item, rd /s /q, del /f /q, Move-Item, Clear-Content and Set-Content are classified exactly as their POSIX equivalents.
• Shell indirection is never a silent allow — $(echo rm), a variable-supplied verb and eval return ask; sh -c "…" and node -e "…" are read and given a real verdict.
• Cursor blocks deterministically, and ten hosts that claimed MCP now actually get a config — including OpenAI Codex CLI, which needed holt’s first TOML writer.
• Static analysis — 20,749 lines shipped with no type checking; now gated by a ratchet that can go down and never up.
• A rewritten site with a real light mode and a picture of what you actually get.

Fixed

Seven ways holt could report work as “safe to delete” when it existed nowhere else — each reproduced end to end with git fsck confirming the loss, each now covered by a test that was watched failing first: vendor/, logs/ and tmp/ treated as generated; git stash unmodeled; rescue reporting verified:true for a submodule it captured nothing from; discard following a symlink into another file’s work; and node -e "require('fs').rmSync(…)" silently allowed.

Also: a rename/rename conflict reported as “no collisions”; gitignored .env.local files manufacturing false HIGH collisions; an O(N²) merge-tree storm; and a ctags argv injection where a file named -L could leak the contents of files outside the batch.

Known limits

• Very large repositories are slow with symbols on — the Linux kernel takes ~16 minutes. --no-symbols is fast and answers a weaker question.
• A duplicate verdict is symbol-identity based: two workstreams that each declare a function with the same name can read as duplicates even when their bodies differ. Measured precision is 0.75 on the benchmark corpus, and every false positive traces to the one case bench50 plants on purpose, once per language — the same symbol name and body declared in a second file with different surrounding content — which is a real, deliberately unresolvable disagreement between a symbol-identity answer and a content-identity oracle, not an unexamined miss (BENCHMARKS.md §8 has the full breakdown, the fix that did land for a different false-positive class, and why this specific number does not move further). holt duplicates --deep adds token-level clone detection (jscpd) for the same question asked a different way.
• conflict recall is 0.96 (2 of 50 planted conflict pairs missed) rather than 1.00. Both misses are holt collisions --json --all not surfacing a pair git merge-tree says collides; which 2 of the 50 repositories miss has moved between scoring runs taken minutes apart during active work on this exact code path — reported as an open, moving defect with the specific repositories named (BENCHMARKS.md §9), not smoothed into a single number that looks more settled than it is.

• Raed2180416
• GitHub Actions