Repository navigation
Releases
v0.19.1
- Docs: README mentions the reviewer's extra commands. The review-gate step of the README
says the reviewer is read-only and that a project can allow more commands for it in
.youtrack-task/reviewer-commands, linkingdocs/configuration.md#letting-the-reviewer-run-more;
the Configuration entry of the docs table names them. No behaviour change. tests/skill-text.shchecks that the README links that section and that the heading exists.
v0.19.0
- Reviewer: an allowed way to verify, still read-only. The
ticket-reviewerguard
(hooks/guard-subagent) now also allowsswift test,swift build,go build ./...,
golangci-lint runandscripts/*test*. Options that write or steer a build (--fix,
-o,-toolexec,-ldflags,-Xswiftc,--package-path) stay rejected. - New config key
reviewer_commandsand a per-project file
.youtrack-task/reviewer-commands(one command per line,#comments): extra
commands the reviewer may run, each matched exactly (whole command, no regex, no
prefix). The project file is read from the base branch withgit show, never from the
branch under review, so a branch cannot allow commands for its own review. See
reference/reviewer-allowlist.mdanddocs/configuration.md. - A rejected reviewer command is no longer a code finding. The hook marks it
[blocked-by-guard]; the reviewer reportsBLOCKED_BY_GUARDand the review gate returns
BLOCKEDwithout a fix round, without a stamp, listing the commands and how to allow them.
Before, "build, lint or CI not checkable" ended asCHANGES_REQUESTED. - The base branch of
.youtrack-task/reviewer-commandsis resolved from full ref names
(refs/remotes/origin/HEAD,refs/remotes/origin/main|master,refs/heads/main|master)
and read by commit id, so a tag namedmainororigin/maincannot shadow it. The worker
guard also deniesgit symbolic-refwrites,git remote set-head,git replace,
git fetch/git pullinto refs (also--refmap),git branch --force/--move/--copy
(also abbreviated),git checkout -B,git switch -Candgit worktree add -B, checked on the command with
its quote characters removed (andgit -C <dir>normalised first, the dir read as a whole shell word: quoted, escaped or plain) so quoting an argument
does not hide it; the lookups ignore replace refs. read-configresolvesreviewer_commands(REVIEWER_COMMANDS, entries joined with;).
Tests cover the new built-ins, both extra sources and the self-allowing branch.
v0.18.5
-
Docs: not just a Claude Code plugin. The README header, tagline and badge now say
the tool also runs in Cursor, GitHub Copilot, Codex CLI and other MCP-capable agents.
The install step offers four routes (Claude Code plugin, Agensi, Product Hunt, other agents), and
the README shows the Product Hunt badge (in the header and in the install table) and links
the Agensi listing. The agent setup
moved to the newdocs/other-agents.md(MCP settings per agent, what differs outside
Claude Code, building the Agensi package). -
tests/skill-text.shchecks that the README linksdocs/other-agents.md, the Agensi
listing and the Product Hunt badge, and that the header names other agents. -
Docs: shorter, scannable README.
README.mdis now the entry point only: what
the plugin does, a quick start, the task flow, a command cheat sheet and
links. The detail moved, unchanged in substance, todocs/usage.md,
docs/batch.md,docs/configuration.mdanddocs/troubleshooting.md(the
troubleshooting table is grouped by topic). -
tests/skill-text.shreads the Config overrides section and the fixed-ladder
statement fromdocs/, and checks that the README links every docs page.
v0.18.4
- Docs: end-to-end example. New
skills/youtrack-task/examples/full-lifecycle.md
follows one issue fromnewthrough pickup, plan,--implement,pranddone,
with the state change and comment each step writes. The README's The full
lifecycle section links to it, andSPEC.mdlists it in the file tree. tests/skill-text.shchecks that the example exists, covers every step, and is
referenced fromREADME.mdandSPEC.md.
v0.18.3
- Fix: no improvised go-ahead question before the pickup. The rule "the
invocation authorizes the pickup" lived only in step 4, so a wrong input in step 1
(a mistyped project key such asBUS-187) made the model ask "Shall I continue
with …?" even withconfirm_branch_name = false. It is now a Guardrail ("No
unlisted confirmation questions"), and step 1 stops on an ID that does not exist,
names it and the likely intended<PROJECT>-<number>, and asks nothing. → Branch: <name>is printed in both rows of the step 4 table, so the applied
setting is always visible.reference/branching.mddefers to the step 4 table. Its unconditional wording
("proceed only after the user has seen it") contradictedconfirm_branch_name = false.tests/skill-text.shchecks the new rules; theSKILL.mdline limit is 390.
v0.18.2
- Config limits documented.
config.example.tomland the README say that arrays
must be written on one line; a multi-line array falls back to the default. - State ladder documented as fixed.
reference/writeback.md, the README and
SPEC.mdstate that only the state names are configurable, not the order. CONTRIBUTING.mdhas the tag step of the change checklist.- README Config overrides names
list_query,local_plan_copyand
default_new_type.
v0.18.1
- One definition per rule. The language rule is worded only in
SKILL.md;
writeback.md,issue-template.md,batch.mdandsuperpowers.mdpoint to it.
confirm_branch_name/confirm_prare decided by the tables inSKILL.mdstep 4
andreference/pr.mdstep 4, and the never-force / never--D/ dirty-worktree
limits by the Guardrails;branching.md,subcommands.mdandbatch.mdrefer
to them. No behaviour changed. - Confirm tables lead with the effective value (flag over config) instead of a
condition in the header cell. review.mdis the runtime procedure only. The stamp contract and the hook
internals moved toSPEC.md.- Examples match the spec:
newshows Context and Scope,prshows the review
gate,pickupshows the branch line, the plan comment,--implement, the review
gate and the completion comment. scripts-fallback.mdis marked as the path for a missingscripts/only and
names what it cannot resolve ([type_prefix], array keys).tests/skill-text.shchecks the above.
v0.18.0
SKILL.mdloads only what the call needs. It shrank from 813 to 372 lines:
new,pr, the small sub-commands (testing,done,log,comment,link,
worktree), the project-resolution procedure, the "scripts missing" fallback
and the workspace set-up of the primary flow moved toreference/new.md,
pr.md,subcommands.md,projects.md,scripts-fallback.mdand
pickup.md;allpoints tobatch.md. Guardrails, Language, Config, Issue IDs,
the dispatch table and the primary flow stay inSKILL.md. Text moved, rules did
not change.- Hard read instructions. Every dispatch row names its reference file with
"readreference/X.mdbefore acting";prandallcarry it in the table. - Shorter description, triggers only. 355 characters instead of about 590; no
workflow summary, with triggers for a pasted issue URL or ID, booking time,
status changes, attaching a PR and batch runs. tests/skill-text.shchecks the size limit, the description, that every
dispatch row names an existing reference file, that the Guardrails stay in
SKILL.md, and that the moved sections are gone from it.
v0.17.2
- New logo. A dark metallic tile with the branch-and-check mark in front of two
pink shapes, on a fully transparent background.assets/logo.svgis the source;
assets/logo.png(512 px, RGBA) is exported from it and replaces the previous
logo in the README header.
v0.17.1
- Fix: one issue-ID regex.
ISSUE_ID_RE([A-Z][A-Z0-9_]*-[0-9]+) is defined
once inSKILL.md("Issue IDs") and is the only form in the skill,reference/,
SPEC, README,review-stampandhooks/require-review. Before, a branch like
fix/P2X-5-fooread asX-5in the sub-commands, so a comment or state change
could land on the wrong issue. Newtests/skill-text.shkeeps the forms
identical;P2X-5cases added totests/review-stamp.shand
tests/require-review.sh. - Fix: scripts are called by absolute path.
<skill-dir>(${CLAUDE_SKILL_DIR}
in Claude Code) is defined once; scripts run as<skill-dir>/scripts/<name>
with the target repo as working directory instead of "from this skill's
scripts/directory". Step 0 checkstest -x <skill-dir>/scripts/read-config
(SCRIPTS=present|missing) to decide on the plain-git fallback. - Fix:
prorder. Checks → clean tree (commit / stash, asked once) → review
gate or waiver → diffstat and confirmation → push → PR → state change → comment.
The comment's, moved to Testing.suffix now reflects whether the state really
changed.review_before_pr = neverand--no-reviewboth record aSKIPPED
waiver (only the flag asks first). - Fix: first-token parsing. For
comment/log/pr/link/testing/
donea leading token matching the ID form is the issue, anything else is text
and the issue comes from the branch. - Fix: resume. On
existing-worktree/existing-branchthe skill checks the
comments for the pickup and plan headings and only skips steps 5 and 7 when
those comments exist. - Fix: bare issue numbers read the issue before asking for confirmation, so
the question shows the title. - Removed the "dominant project prefix" guess; only "Resolving a project"
applies. - Changed: step 8 needs
--implement. The "proceed now?" question is gone.
Without the flag the flow ends after the plan write-back and prints the command
to continue. - Changed: the local plan copy is always English (a git artifact); the
YouTrack comment follows the issue's language. - Docs:
confirm_prapplies topronly (allalways asks its batch
confirmation); worktrees are entered withEnterWorktreeor per-call working
directories instead ofcd;worktrees.mddescribesdone's automatic
sweep; the completion comment is described one way everywhere.