han-release — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited han-release (Agent Skill) and scored it 96/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 1 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
<!-- AskUserQuestion is deliberately absent from allowed-tools, and must stay absent. Listing it makes Claude Code's permission evaluator auto-approve the tool through its always-allow path and return empty answers without ever rendering the prompt, so every gate in this skill would silently pass. See han-plugin-builder/skills/guidance/references/skill-building-guidance/allowed-tools-AskUserQuestion.md. The tool still works unlisted; it prompts once for permission. -->
which gh 2>/dev/null || echo "not installed"which jq 2>/dev/null || echo "not installed"which claude 2>/dev/null || echo "not installed"git rev-parse --is-inside-work-tree 2>/dev/null || echo NOIf `gh`, `jq`, or `claude` reads `not installed`, or this is not a git repo: tell the operator which prerequisite is missing and that it must be installed/configured before /han-release can run, then immediately stop. The skill cannot proceed without all four.
The claude CLI is what creates the per-plugin tags in Step 10. Every invocation of it in this skill goes through the shell's command builtin (command claude ...), never a bare claude, because an operator's shell commonly wraps claude in a function or alias that blocks waiting for terminal input. The which probe above resolves the same way a bare call would, so it reports the wrapper's presence rather than the executable's; treat a non-empty result as "the tool is reachable" and let Step 10's first invocation be what proves it runs.
gh repo view --json nameWithOwner -q .nameWithOwner 2>/dev/null || git config --get remote.origin.urlgit branch --show-current 2>/dev/null || echo unknowngit symbolic-ref --short refs/remotes/origin/HEAD 2>/dev/null | sed 's#^origin/##' || echo unknowngit status --porcelain 2>/dev/null || echo NOjq -r .name .claude-plugin/marketplace.json 2>/dev/null!jq -r '.plugins[] | "\(.name)\t\(.source)\t\(.version)"' .claude-plugin/marketplace.json 2>/dev/null
git fetch --tags --quiet >/dev/null 2>&1; git tag -l 'han--v*' --sort=-v:refname | head -n1git tag -l 'v*.*.*' --sort=-v:refname | head -n1grep -m1 '^## v' CHANGELOG.md 2>/dev/nullThe two tag probes are separate on purpose. A single pattern covering both namings returns the wrong tag: version sorting compares the whole refname, so v4.6.0 sorts ahead of han--v5.0.0 and the old suite tag wins on every release after the transition. The fetch runs on the first probe only; both read the same refreshed tag list.
latest parent tag carries the literal han--v* pattern, because a context-injection command is a fixed string and cannot interpolate parent plugin name. If parent plugin name is not han, redo the lookup in Step 2 with the actual name before using either value. Skipping that turns a renamed parent into an empty probe, which Step 2 would read as a first release and silently expand the changelog to the whole repository history.
name (parent plugin name above, normally han). Ithas no skills or agents of its own; it exists to install the children via dependencies. The parent's own per-plugin tag is the one the GitHub release attaches to, so the release tag is {parent plugin name}--v{parent target}.
marketplace.json.plugins[] (han-core, han-github, han-reporting, and anyfuture han-* plugin). Each child has its own version line, bumped independently of the others.
prev (the latest release tag). For the parent this is prev#. For a childit is the version recorded in that child's plugin.json at prev; if the child did not exist at prev, it is a new plugin (see Step 3).
plugin.json.{parent plugin name}--v{parent target}.
{name}--v{target}, for example han--v5.0.0 and han-core--v3.0.0. Every plugin getsone, and the parent's is what the GitHub release attaches to.
prev is the previous release's tag, resolved in Step 2 from the two tag probes. On the first release it is empty.prev# is the part of `prev` after its last `v`. That rule is correct under both namings, which matters becauseprev can be either shape:
v4.6.0 -> prev# is 4.6.0
han--v5.0.0 -> prev# is 5.0.0Do not read prev# as "the number without the leading v". That older rule returns the whole tag string for a per-plugin tag, and prev# is the parent's baseline (see 3a in references/version-plan-rules.md), so a wrong value here silently corrupts the entire version plan rather than failing.
source field in marketplace.json (for example ./han-core), so itsplugin.json is {source}/.claude-plugin/plugin.json. Use {source} verbatim in every git command: the ./-prefixed form works both after a {ref}: colon (git show {prev}:{source}/...) and as a pathspec (git diff ... -- {source}/). Do not strip the leading ./.
informs the changelog narrative:
pause_before_publish — true if the argument contains "pause", "review", or "confirm before publish"(case-insensitive). Default false.
draft_release — true if the argument contains "draft". Default false.$release_context, passed into the narrativedispatch in Step 5. May be empty.
working tree from Project Context is non-empty, there are uncommitted oruntracked changes. Stop and tell the operator to commit or stash them first. Releasing an unknown working state is unsafe and a pushed tag is hard to reverse. This is a hard stop, not a pause gate.
current branch is not the default branch, do not stop — note in the Step 7summary that the release is being cut from current branch and the tag will point at that branch's HEAD. The operator chose autonomous; surface the fact, do not block.
latest parent tag when it is non-empty. This is the normal case from the second per-plugin release onward.latest suite tag. This is the transition release, the one release where the previous tag still carriesthe old vX.Y.Z naming. Once a parent tag exists on the remote this arm is unreachable and can be deleted.
all compare links are omitted, and every plugin is treated as new (Step 3).
If parent plugin name is not han, the latest parent tag probe used the wrong literal. Re-run the lookup yourself before applying the precedence above: git tag -l '{parent plugin name}--v*' --sort=-v:refname | head -n1.
prev feeds two different things: the commit range below, and the parent's baseline version at 3a in references/version-plan-rules.md via prev#. Parse prev# with the after-the-last-v rule in the vocabulary block, not by stripping a leading v.
${prev}..HEAD. First release: the full history (HEAD with no range base).git log {range} --oneline. If it is empty, there are no commits since prev.Stop and tell the operator there is nothing to release.
git log {range} --pretty=%s%x00%b | grep -oE '#[0-9]+' | tr -d '#' | sort -unFor each number N, run gh pr view N --json number,title,author,url,mergedAt,state. Keep only entries where state is MERGED. Sort the survivors by mergedAt ascending (newest merge last). This is $pr_list. The PR list is repo-wide and appears once per release; it is not split per plugin. Build the PR lines and the changelog bullets per references/release-notes-format.md and references/changelog-rules.md.
$pr_list is empty (local-only or squash history with no PR refs), record the notable commitsubjects from git log {range} --oneline instead, and use the commits form documented in both reference files.
N in $pr_list, find the issues that PRclosed and credit everyone involved, following references/attribution-rules.md. That file holds the closing-issue lookup, the substantive-comment test that decides who counts as a contributor, the bot-account exclusions, and the shape of each $issue_list entry. If no closed issues are found, $issue_list is empty and the issues subsection/section is omitted everywhere.
Enumerate the plugins from plugins in Project Context (one parent, plus each child). For every plugin, determine baseline, whether it changed in {range}, and its target. Classify changes against docs/semantic-versioning.md. The governing rules:
because every release is a release of the suite.
plugin.json version is its establishedbaseline. Record the introduction in the changelog, but do not increment. This is the general rule for every future han-* extension, not a one-time exception for the current children.
Classify each plugin, compute its bump level, and decide its target by following references/version-plan-rules.md. That file holds the baseline lookup for a parent, an existing child, and a new child (3a); the major, minor, and patch classification for a changed child and for the parent (3b); and the ahead-path versus compute-path decision that determines which plugins need confirmation below (3c).
target = parent target drives the tag {parent plugin name}--v{parent target}.
This gate is conditional and often does not fire. It is not the gate that approves the tags; that one is Step 9 and it always fires.
determined. Do not prompt. Record it for the Step 7 summary and continue.
AskUserQuestion(header: "Release versions"). State prev, and for every plugin a line of the form {name}: {baseline} → {proposed} ({level}; {evidence}), marking new plugins as new at {current} (no bump), ahead plugins as already at {current}, and unchanged children as unchanged at {current}. Name the specific skills/agents and commits that drove each level. Options: accept the proposed plan (recommended, first); adjust the parent level; adjust a child level; enter explicit versions. Apply the operator's answer to the affected plugins. Record the final plan as the version decision in the Step 7 summary.
Run this before Step 4 writes anything. The working tree is still clean here, and that is the point: a failure at this step leaves the repository exactly as the run found it, with nothing written and nothing pushed.
Both checks cover every plugin in plugins, not only the ones being bumped. Version drift in an untouched plugin is the likeliest way this fails, because Step 4 only writes plugins whose target differs from current and so never corrects a plugin nobody changed.
plugin.json version against its marketplace.json entryselected by name:
jq -r .version {source}/.claude-plugin/plugin.json
jq -r --arg n '{name}' '.plugins[] | select(.name == $n) | .version' .claude-plugin/marketplace.jsonAny disagreement stops the run. Name the plugin and both versions. Do not attempt a repair: the correct value depends on which file is wrong, and that is the operator's call.
git status --porcelain must be empty. Step 1.2 already required this, so a non-empty resulthere means something wrote a file since the run started. Stop and say so.
Do not implement either check by invoking command claude plugin tag --dry-run. That command refuses on a dirty plugin folder, and after Step 4 every bumped plugin's folder is dirty by construction, so a dry-run sweep placed after the writes refuses exactly the plugins the release just bumped and cannot tell that refusal apart from real drift. The two reads above answer the same question and work in either state.
For every plugin whose target differs from its current (the compute-path plugins from 3c in references/version-plan-rules.md, and any plugin the operator edited at 3d), set both files so they read target:
{source}/.claude-plugin/plugin.json version to that plugin's target (Edit).marketplace.json entry: set the version of the plugins[] element whose name equals theplugin name (Edit). Select by name, not by index.
Skip any plugin whose target == current (ahead-path or new plugins — their files are already correct). When the entire plan is ahead-path/new (no version differs from current), this step is a no-op; note it and continue.
Re-run the version-agreement check from Step 3.5 over every plugin this step wrote. Each plugin's two files are edited separately, so this step is the one place in the run that can create a half-applied bump. Skip the cleanliness check here: the tree is dirty by design now. A disagreement stops the run, and nothing has been committed yet, so the recovery is discarding the working-tree edits.
Follow references/changelog-rules.md exactly, writing the narrative in the voice at writing-voice.md. From v3.0.0 onward, each release section is a parent ## v{parent target} heading with one ### {plugin} v{version} sub-heading per plugin that changed (the parent always appears; new and changed children appear; unchanged children are omitted), plus the release-level bookkeeping subsections. Every @mention in the changelog (narrative, PR bullets, issue bullets) is a markdown link to the person's GitHub profile: [@{login}](https://github.com/{login}), never flat text.
subsections as the last ### subsections of the ## v{parent target} section, before the next ## v heading, in this order: ### Issues closed in this release (only when $issue_list is non-empty), then ### Pull requests in this release (or the commits form from the fallback). Build the issue bullets from $issue_list and the PR bullets from $pr_list (Step 2), and close the final subsection with the Full changelog: line using the blob link from references/release-notes-format.md. Use Edit.
If a generated bookkeeping subsection is already present under this heading, replace it in place rather than appending a second copy. A re-run reaching this branch is ordinary: the Step 9 tag gate sits after the release commit, so declining it and running again later brings the run back to its own changelog section.
general-purpose agent to write the narrative## v{parent target} section. The skill already holds this context — paste the actual values into the prompt, do not tell the agent to go read them:
baseline → target, and for each changed/new child its name,baseline → target, level, and new/changed status.
git log {range} --oneline and git diff {range} --stat, plus, per changed plugin,git diff {range} --stat -- {source}/ so the agent can attribute each change to its plugin, plus a suite-level stat git diff {range} --stat -- docs/ README.md CONTRIBUTING.md CHANGELOG.md .claude-plugin/ labeled as the evidence for the ### han parent section (repo-root changes outside any plugin directory).
$pr_list (PR numbers, titles, authors).$issue_list (Step 2): each closed issue's number, title, opener, contributors, closing PR(s), and the relevantfix, so the narrative can credit the issue opener where it describes that fix.
$release_context from Step 1 (may be empty).## v{X.Y.Z} sections from CHANGELOG.md verbatim, as the register model.references/changelog-rules.md, pasted in full.
Prompt the agent to: produce only the markdown for the ## v{parent target} section — a one-paragraph summary that names the parent's new version and lists each changed/new child with its version, then one ### {plugin} v{version} sub-heading per changed or new plugin (parent first), each describing only that plugin's changes (using #### for topic subsections when needed), then a release-level ### Deferred (YAGNI) subsection only when work was deliberately cut; when a change closes a tracked issue, name the fix and credit the issue opener inline as a profile link [@{login}](https://github.com/{login}); render every @mention as a [@{login}](https://github.com/{login}) profile link, never flat text; match the register of the two pasted sections; obey every hard voice constraint; attribute every change to the plugin whose directory it touched; never invent changes not present in the commits, diff, PR list, or issue list; return only the section markdown with no preamble. If the agent returns anything else, discard it and re-issue with an explicit "return only the section markdown" reminder.
Insert the returned section directly under the # Han Release Notes title, above the previous newest entry (Edit/Write). Then append the generated bookkeeping subsections to it exactly as in the augment case.
Build the GitHub release body per references/release-notes-format.md: the release's summary paragraph first (the one-paragraph overview from the ## v{parent target} narrative, with no heading above it), then ## What's Changed and the PR lines (* {title} by @{login} in {url}, newest merge last), then an ## Issues closed section built from $issue_list (omitted when empty), then every ### {plugin} v{version} sub-heading of the narrative excluding the ## v{parent target} heading itself and the generated PR/commits/issues bookkeeping subsections (the summary paragraph already leads the body, so it is not repeated here), then the **Full changelog:** blob link and the **Full Changelog:** compare link (compare line omitted on a first release). Compute the blob anchor by lowercasing v{parent target} and deleting every character that is not a-z, 0-9, or - (v3.0.0 → v300). Write the assembled body to /tmp/han-release-notes-v{parent target}.md with the Write tool. Do not assemble it with shell echo/printf.
Print to the operator, regardless of mode:
{parent plugin name}--v{parent target}, the parent's baseline → target andhow it was decided (ahead-of-tag → used as-is, or computed-and-confirmed at Step 3), and one line per child (bumped baseline → target, unchanged at version, or new at version).
CLAUDE.md states a "Current version:" that does not equal the parent target, note it as a follow-up the operator may want to make (do not edit CLAUDE.md; it is out of scope).
## v{parent target} section.--latest, or draft (only if draft_release).If `pause_before_publish` is true: use AskUserQuestion (header: "Publish release") — options: publish now (proceed to Step 8), abort (stop, having changed only local files). Do not push or publish until approved.
If `pause_before_publish` is false (default): continue to Step 8 without pausing.
This gate is opt-in and covers the whole release. It is separate from the Step 9 tag gate, which always fires and covers the tags only. When both fire they arrive close together; that is intended, and the distinct header values are what tell them apart.
The operator's request to tag and publish authorizes the commit and push required to do it.
CHANGELOG.md, .claude-plugin/marketplace.json, and every{source}/.claude-plugin/plugin.json that Step 4 changed. Commit with chore(release): v{parent target}. The commit subject keeps the plain version; it is a message, not a ref. If nothing is staged (augment produced no diff and no version changed — unlikely), skip the commit and note it.
git rev-parse HEAD. Every tag must point here, and Step 9 re-checks it.git push origin HEAD. If this fails, stop: no tag has been created, sonothing irreversible has happened. Pushing a tag first would transfer the commit to GitHub without putting it on any branch, and a tag can never be moved afterwards.
The commit is on GitHub and no tag exists yet. Everything past this point is irreversible.
{name}--v{target}, for every plugin in plugins, then split them into twosets using the Step 3 version plan:
target differs from its baseline (the parent always, pluseach bumped child), and every new plugin classified at 3a, whose tag has never existed.
target equals its baseline. Its tag was created by therelease that shipped that version and points at that older release commit. This run does not re-tag it.
A release commonly bumps a handful of plugins and carries the rest forward, so the carried-forward set is routine rather than exceptional. Treating it as an error blocks every release after the first.
${CLAUDE_SKILL_DIR}/scripts/remote-tag-state.sh {release commit} {tag name}...It prints {tag}\t{state}\t{sha} per tag. Exit 2 means the remote could not be read: stop, because a run that cannot see the remote must not guess. The script reports what the remote holds and nothing more; this step decides what each state means, because one of the four means different things for the two sets:
| State | What Step 10 does with it |
|---|---|
absent | Create and push it. |
remote-at-commit | Skip it. Already published at this commit. |
remote-at-other-commit | Depends on which set the plugin is in (see 3). |
local-only | Push it. This is never a skip. |
{release commit} is the normal resting state for a carried-forward plugin and an unrecoverable collision for one being tagged now, so the two cannot share a rule.
git merge-base --is-ancestor {sha} {release commit}. When it succeeds, the tag pointsinto this release's own history, which is what an already-shipped version looks like. Record it as already on GitHub and skip it; Step 10 never touches it. When the check fails, the tag sits on a commit this release does not descend from, so treat it as the stop below.
publish is already taken.
On a stop, stop before a single tag is created. Name the tag, the commit it points at, and {release commit}. Say plainly that pushing is not the recovery: the push is rejected identically every time, GitHub does not allow a published release's tag to be moved or deleted, and this skill never forces one. The way out is a different version number, which is free right now and costs a burned version once the walk starts.
AskUserQuestion (header: "Tag plugins"). Show:resolved at 3. Say that none of them is moved or deleted.
{release commit} and the branch it is on, with the Step 1.3 note if it is not the default branch.draft_release is true, that the draft flag holds back the release page only. Every tag is still created andpushed, and none of them can be moved afterwards.
Options: create and push the tags (recommended, first); abort.
version needed computing.
AskUserQuestion has a knownfailure mode that returns empty answers silently, which is why it is absent from allowed-tools (see the note under the frontmatter). This is the one gate in this skill where that failure would be unrecoverable.
so, name it, and say that re-running later picks up from there: the changelog section already exists, so the run will augment it rather than regenerate it.
{release commit}, stop. The tags arecreated against the current checkout, so a commit made while the gate was waiting would tag something that was never pushed.
Walk plugins, taking the entry whose name equals parent plugin name first, then the rest in listed order. The parent goes first so the tag the GitHub release attaches to is the one most likely to exist if the walk stops partway. Order by name, not by position in the file.
For each plugin, act on the outcome Step 9 recorded for it:
shipped that version, and this run leaves it exactly where it is.
git push origin refs/tags/{tag name}. Record it as newly pushed. command claude plugin tag {source} --pushThe command builtin is required, not stylistic: a bare claude runs whatever function or alias the operator's shell defines, which commonly blocks waiting for terminal input. The command derives the tag name from the plugin's manifest and marketplace entry and re-checks that the two agree, so a disagreement that reached this point surfaces here.
Never pass `--force`. It skips both the dirty-tree and tag-exists checks, and this skill does not move tags.
Reading a failure. The command exits 1 for both a benign refusal and a real failure, so the exit status alone does not tell them apart. Match the message:
already exists locally — the tag is on this machine but was not on the remote at Step 9, so a push is what isneeded. Push it as in the local-only case above. Do not re-run the tagging command; it will refuse identically.
folder, and a version disagreement that Step 3.5 should have caught. A stop here is unexpected rather than routine.
On a stop mid-walk, do not publish. Report the full tag state: which tags are on GitHub, which exist only on this machine, and for each of the latter the literal recovery command git push origin refs/tags/{tag name}. Re-running /han-release does not retry a failed push, so never offer that as the recovery.
tag name, holding each result to the rule for itsset. Every plugin being tagged this release must now read remote-at-commit. Every carried-forward plugin must read either remote-at-commit or the same remote-at-other-commit sha that passed the ancestor check at Step 9.3. No plugin in either set may read absent or local-only.
If any result fails its rule, stop and report as in Step 10. Publishing with a tag missing from GitHub is the failure this gate exists to prevent. Demanding that every tag sit on {release commit} is not that check: it would stop every release that carries an unchanged plugin forward, which is nearly all of them.
/tmp/han-release-notes-v{parent target}.md (the filename keeps the plain version; it is a local path, not a ref):
gh release view {parent plugin name}--v{parent target} fails):gh release create {parent plugin name}--v{parent target} --verify-tag --title "v{parent target}" --notes-file /tmp/han-release-notes-v{parent target}.md plus --latest, plus --draft only when draft_release is true (never --latest together with --draft).
--verify-tag is required. Without it gh silently creates a missing tag at the default branch's head, which would mint a permanent tag at a commit nobody released. The --title keeps the plain v{parent target} so the releases page reads continuously across the naming change.
gh release edit {parent plugin name}--v{parent target} --notes-file /tmp/han-release-notes-v{parent target}.md. Add --draft=false only when the operator asked to publish an existing draft (and draft_release is not set). Report it as updated, not created.
Report concisely:
{release commit}, and whether it is on the default branch.remote-at-commit or carried forward from an earlier release), or present only on this machine. Those three are different, and "created" plus "skipped" cannot express the third, which is the state a partly-failed release leaves behind.
git push origin refs/tags/{tag name} recovery line.generated.
CLAUDE.md version drift, non-default release branch).If the operator aborted at Step 7 or Step 9, report exactly what was changed locally, what was pushed, and what was not.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.