Changelog

All notable changes to loomwork are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.4.1] - 2026-09-17

Added

• skills/split-agents-md/SKILL.md splits one oversized agent guidance file into a tool-agnostic AGENTS.md plus thin host files that import or reference it. The skill is manual-only — its frontmatter description says so at the trigger surface, so it never fires on its own. It carries the classification procedure that decides what is shared knowledge and what is host mechanics, and it moves any marker-guarded block into AGENTS.md so scripts/init.mjs keeps finding exactly one block to update.

  A host file holding nothing but the import is still written when the host cannot read AGENTS.md natively. Claude Code is such a host: dropping its CLAUDE.md as "import-only" would leave it with no shared guidance at all. Only a host that reads AGENTS.md on its own — Codex — gets no file.

  The skill commits only when the user asks. The classification calls are the kind a human wants to review before they land in history.

0.4.0 - 2026-09-16

loomwork becomes a portable Agent Plugins package. Codex can install and run it natively, and the lifecycle gates fire on explicit Codex skill prompts. Claude Code and Cursor are unchanged as supported runtimes.

This release also fixes a bug that made the gates fail silently: they matched with echo "$x" | grep -q under set -euo pipefail, where grep -q exits on its first match and the resulting write failure, combined with || exit 0, turned the gate into a no-op — exit 0, no output, strategy never injected. It read exactly like a gate that chose not to fire. Reproducing it needs a newline after the trigger and a tail past the 64KB pipe buffer; a 200KB prompt dropped the injection 10/10. The Cursor mirrors carried the same defect, where it mattered more, because beforeSubmitPrompt receives the raw user prompt and an ordinary pasted log was enough to disable the gate.

Added

• Root plugin.json on the Agent Plugins 1.0 schema, selecting the Codex hook registration through extensions["com.openai"]. .claude-plugin/plugin.json remains as the Claude compatibility manifest, and a test keeps the two identities aligned.
• hooks/codex-hooks.json registers both gates on UserPromptSubmit. Codex observes explicit $superpowers:... prompts only — implicit skill selection produces no signal the gates can identify.
• skills/init/SKILL.md exposes the initializer as loomwork:init, so initialization no longer requires the Claude-only slash command.

Changed

• The shared gates accept both payload shapes — Claude PostToolUse and Codex UserPromptSubmit — and echo the incoming event name back in hookSpecificOutput.
• Root resolution is ordered. A non-empty CLAUDE_PROJECT_DIR is used exactly as given; otherwise the payload's .cwd walks up to the nearest .git or .loomwork.json, since Codex may start in a subdirectory, falling back to the raw .cwd so uninitialized repositories still get nudged. .git is tested with -e, so a worktree's .git file counts.
• The audit and init skills resolve their CLIs relative to the installed SKILL.md instead of requiring a plugin-root environment variable.
• A fresh consumer repository now receives AGENTS.md, not CLAUDE.md. An existing AGENTS.md still wins, an existing CLAUDE.md is still supported, and marker-block idempotency is unchanged.
• The strategy gate is spill-aware under Codex. Output past additionalContextLimit (2500) is truncated and saved to a file, so the message no longer claims the whole strategy is inline; it points at the saved hook-output file when the host truncated. The Cursor mirror keeps the original wording on purpose — Cursor sets no limit and never spills.

Fixed

• Gate matching no longer uses a pipeline. All matching moved to bash =~ against an unquoted pattern variable, with shopt -s nocasematch preserving the case-insensitivity grep -qiE provided in the Cursor gates. Semantics were diffed against the original patterns across representative inputs with zero divergence.
• hooks/close-out-gate.sh no longer reads /.loomwork.json when no root resolves. It exits silently, matching hooks/strategy-gate.sh.
• Tests no longer leak an ambient CLAUDE_PROJECT_DIR, which would have made every payload-.cwd test pass vacuously.

0.3.0 - 2026-08-21

STRATEGY.md belongs to compound-engineering:ce-strategy. loomwork reads it, points at it, and reminds people to run ce-strategy — it does not define the file's shape and does not write to it. Close-out had been appending one ## Milestones bullet per merge, creating that section when absent, and bumping last_updated on every ship, which is the content ce-strategy documents as optional and skip-by-default. This release removes every loomwork write path to the file.

Changed

• loomwork:close-out Step 3b is a decision, not a write. It asks whether the ship introduced an externally visible milestone, changed what an investment area covers, retired a direction, or shifted the target problem. A yes routes to a ce-strategy run on the feature branch so the update rides the same PR; a no touches nothing. No last_updated bump, no Milestones append, no Tracks or ## Not working on edit.
• Doctrine restated in references/PLAYBOOK.md and README.md, with the governing rule recorded so future changes to loomwork's handling of the strategy file can be tested against it.

Added

• Missing-file nudge in both hook gates. hooks/strategy-gate.sh and templates/cursor/loomwork-strategy-gate.sh previously exited silently when no strategy file existed, giving no reminder at the moment grounding matters. Both now inject a one-sentence pointer to ce-strategy, each through its own output envelope. The gates check existence only, never shape; behavior when the file exists is unchanged.

Removed

• templates/STRATEGY.md — a loomwork-authored variant of ce's template that drifted independently from it.
• Strategy-file seeding in /loomwork:init. initRepo now reports the file as missing and names ce-strategy instead of creating one. An existing strategy file is still left byte-identical, and no already-initialized repo loses content — accumulated ## Milestones entries are left in place.

Fixed

• loomwork:close-out Step 5 ran git add docs/... STRATEGY.md, which aborts with fatal: pathspec 'STRATEGY.md' did not match any files (exit 128) when the file is absent, so the commit on the next line never ran. Now guarded on existence. Reachable in practice as of this release, since init no longer seeds the file.
• hooks/strategy-gate.sh resolved its root as ${CLAUDE_PROJECT_DIR:-$PWD} and could act on an unrelated directory's strategy file. It now requires a resolved CLAUDE_PROJECT_DIR, matching the Cursor gate.

0.2.1 - 2026-08-02

Fixed

• Cursor hook commands now invoke gate scripts via explicit bash .cursor/hooks/loomwork-*.sh instead of the bare path — the bare path let Cursor/Windows open the .sh as an editor tab instead of executing it, stealing focus and letting stray keystrokes corrupt shebangs.
• Reworded the strategy gate's injected prefix (Claude Code + Cursor) so agents are told the content is already inlined and not to Read/open the strategy file or any loomwork hook script.
• /loomwork:init migrates legacy bare-command .cursor/hooks.json entries in place on re-run instead of appending duplicate gates.

0.2.0 - 2026-08-02

First working release. 0.1.0 was a skeleton; everything below is what turned it into a usable plugin.

Added

• SDD drift audit engine — scripts/sdd-audit.mjs plus scripts/lib/, ported from weavereads, with unit tests. Flags stale spec status, missing plan DONE banners, and missing/stale verified dates.
• loomwork:audit skill — runs the drift linter over the SDD docs dirs.
• loomwork:close-out skill — pre-merge close-out that commits and pushes to the feature branch, so the close-out lands in the same PR.
• /loomwork:init command — backed by an idempotent scripts/init.mjs scaffolder that lays down the SDD docs dirs, a STRATEGY.md seed, Cursor hooks, and the CLAUDE.md block.
• Hook gates — hooks/strategy-gate.sh and hooks/close-out-gate.sh, wired through the plugin's hooks.json.
• Cursor hook templates — dual-event support for beforeSubmitPrompt and postToolUse, with TDD tests.
• Repo templates — templates/STRATEGY.md and templates/claude-md-block.md.
• Configuration — .loomwork.json loading with repo-root walk-up resolution, configurable paths, and a CLI entry point.
• Docs — playbook doctrine and README.

Fixed

• Force LF line endings via .gitattributes; bash scripts break on CRLF.
• Resolve gh invocations against the correct working directory.