1.6.0

Added

Two items adapted from Matt Pocock's skills (main at 5beb35b, after v1.3.1):

• retro: a skill you invoke after a coding session. It reads the session, lists what slowed the agent down or let a mistake through, most severe first, and builds the one you pick as a test, check or doc in the project. A new check has to fail on the mistake before it counts. It loads writing-for-agents for anything it writes.
• git-guardrails.mjs: an opt-in PreToolUse hook that denies git push, git reset --hard, git clean (except a dry run) and git branch -D before they run. It reads the command as shell, so a quoted commit message or a heredoc that mentions git push is not blocked. It is not registered in hooks.json; hooks/README.md has the settings entry.

1.5.2

Changed

Three changes ported from Matt Pocock's skills (main at 5beb35b, after v1.3.1):

• grilling words each question so "yes" accepts the recommended answer, so agreeing never means answering "no".
• tdd gives each proposed seam a one-line note on what it catches and what it misses.
• diagnosing-bugs has the agent diff a forced mutation against an untouched copy before trusting the red, so an edit that changed nothing cannot pass as a failing test.

1.5.1

Hook fix release. Installed plugin users pick it up with claude plugin marketplace update ehnruiz then claude plugin update agent-workflow@ehnruiz.

Fixed

• The hooks read stdin to its end even when it is a non-blocking pipe that is momentarily empty. Before, an EAGAIN on the read made config-protection, commit-secrets, edited-files and stop-check see an empty event and allow the call; the config-protection suite failed this way in 5 of 12 runs on Linux with Node 20.

Added

• docs/examples.md: three examples run against this repository with their commands and exit codes, for regression-proof, skill-collision-audit and supervisor.
• docs/install.md: requirements, what was tested on which platform, the three install paths and how to check each one.
• docs/releasing.md and this changelog.
• Layout tests for the docs: relative links in the README, the changelog and docs/ resolve, the install commands in the README and docs/install.md match the manifests, and the changelog has a section for the current version.
• The CI matrix runs the Python suite on 3.9 and 3.12 and the Node suite on 20 and 22, the floors the docs state.

Changed

• README: a shorter opening on what the library is for, a five-minute quick start, and pointers to the new docs. The skill, subagent and hook tables are unchanged.

v1.5.0

First tagged release. These are the Claude Code skills, subagents and hooks I use every day, rewritten so they work in any repository.

Install

From inside Claude Code:

/plugin marketplace add ehnruiz/claude-skills
/plugin install agent-workflow@ehnruiz

Or with the skills CLI, which can also install into other coding agents:

npx skills add ehnruiz/claude-skills
npx skills add ehnruiz/claude-skills --skill tdd

What's in it

Spend less on each task

• supervisor splits a multi-step task into tickets for cheaper subagents. A subagent that fails twice moves up one model tier.
• Subagents for finding, writing, building, testing, research and review, each pinned to a model tier.

Build, debug and prove changes

• tdd: one vertical slice at a time. New in this release: wait on the condition instead of a fixed sleep, and fake the clock for time-based code.
• diagnosing-bugs: build a command that goes red on the bug before guessing.
• regression-proof: judge every gate by its exit code, and prove a fix's test fails without the fix.
• explain-the-number (new): before crediting a change with a number, check the two runs measured the same thing.
• research and grilling.

Pick up where you left off

• handoff and pick-up, plus an optional session-start hook that shows the newest handoff. Notes live in ~/.claude/projects/<project>/handoffs/, shared by worktrees and found correctly inside git submodules.

Prepare for reviews and meetings

• epic-review, confluence-jira-links and leader-1-1-prep.

Write tickets and messages people read

• ticket-write, ticket-trim and ticket-digest, with a lint script that holds tickets to word budgets.
• plain-writing for messages, emails and PR descriptions sent under your name.

Keep a growing skill library predictable

• skill-collision-audit finds skills that compete for the same request. The repo's own test fails when a new skill overlaps one that's already there.
• writing-for-agents, now with PRESSURE-TESTS.md: prove a rule changes what an agent does before it ships.

Publish under your own name

• no-ai-attribution keeps AI credit lines and co-author trailers out of commits, PRs and comments.
• Opt-in session hooks: block edits to lint, test and CI config, catch secrets before a commit, and run eslint and the type check on edited files before a response ends.

Credits

grilling, tdd, diagnosing-bugs, research and writing-for-agents are adapted from Matt Pocock's skills (MIT). explain-the-number follows Lauren Tan's pstack (MIT). PRESSURE-TESTS.md rewrites the method of obra/clank. Each adapted skill carries its own LICENSE.md.