jira-sdlc-tools

An SDLC layer for AI coding assistants — the Jira API, the Gitflow model, and git worktrees as isolated per-issue workspaces.

Shipped as a Claude Code plugin (installable from its own marketplace) and as a loose skill set for coding-assistant platforms that respect the Claude or agentskills.io specifications.

The skills are explicit-invocation only by design — never auto-triggered — and carry the corresponding setting on both specs: disable-model-invocation for Claude, allow_implicit_invocation: false for agentskills.io.

Trademarks

Jira and Atlassian are trademarks or registered trademarks of Atlassian Pty Ltd, in the United States and/or other countries. This is an independent, community-built project that integrates with Jira through its public CLI and APIs; it is not affiliated with, endorsed by, or sponsored by Atlassian, and its references to Jira are solely to describe compatibility.

⚠️ Caution

This plugin acts as an authenticated user in both git and Jira. Given credentials, it will commit, create and push branches, open and update pull requests, and take the other actions needed to follow the gitflow strategy — and it will create, update, transition, and comment on issues in your Jira project. Those actions are visible to your team and land under whichever account you configured — your own, or a dedicated one per skill.

Use it with caution: point it at a project you're comfortable having changed, and read Settings files before the first run so you know which repo and which Jira project it's wired to.

What it deliberately never does on its own — merging into your base branch, deleting Jira issues, resolving conflicts — is listed in Safety model.

Environment setup

You will need:

• Software — see Software
  • Install git, gh, and jq
• GitHub — see GitHub
  • Account (can be free)
  • Repository
  • fine-grained PAT scoped to the repository with the following permissions:
    • Contents (read/write)
    • Pull requests (read/write)
    • Meta (read) - added automatically
• Jira — see Jira
  • Account (can be free)
  • Space with a Board (you will have a Project Key, e.g. XYZ)
  • Four configured workflow statuses (names can differ):
    • TODO
    • IN_PROGRESS
    • IN_REVIEW
    • DONE
  • Users: can be only the owner, or additionally a dedicated user per each of the skills - assigner, executor, reviewer (fits the free tier - up to 10 users in org).
  • Classic Jira API token (either for owner, or for each of the 3 users - assigner, executor, reviewer) with the following permissions:
    • read:jira-user
    • read:jira-work
    • write:jira-work
• Coding Assistant — see Coding Assistant
  • Claude or any other compatible solution (see Platform Compatibility Matrix)

Prefer a single tickable list over this section-by-section walkthrough? See Full setup checklist.

Detailed reference for each item: Environment setup.

Quick install

Claude Code

Remote — from the marketplace (recommended)

/plugin marketplace add kantorv/jira-sdlc-tools
/plugin install jira-sdlc@jira-sdlc-tools

Local — clone, then load with --plugin-dir

git clone https://github.com/kantorv/jira-sdlc-tools.git
claude --plugin-dir ./jira-sdlc-tools/plugins/jira-sdlc

See full doc: CLAUDECODE.md

Non Claude Code assistants

This plugin can also be installed as a loose skill set with various coding assistants other than Claude Code, Antigravity, Cursor, Kimi Code, and more. See the Platform Compatibility Matrix for the full list and integration status per platform.

Platform Compatibility Matrix

The skills target the Claude skills spec, so jira-sdlc also works — natively or through the Agent Skills adaptation — in a growing set of other AI coding assistants: Cursor, Kilo Code, Codex, Antigravity, OpenCode, Grok Build, Pi, Kimi Code, and Muse Code. See Platform Compatibility Matrix for the platform-by-platform table — each one's spec, wiring, integration status, and a link to its detailed doc.

Platform	Specification	How it loads	Integration status	Compatibility	Documentation
Claude Code	Native Claude skills	plugin marketplace · .claude/skills/ drop-in copy · --plugin-dir	First-class (reference)	✅	CLAUDECODE.md
Cursor	Native Claude skills	shares the ~/.claude/ tree with Claude Code	Verified — Linux/macOS	✅	CURSOR.md
Kilo Code	Native Claude skills	kilo.jsonc skills path	Working	✅	KILO.md
Codex (CLI)	Agent Skills	.codex/skills/ copy + per-skill agents/openai.yml	Working — sandbox & timing caveats, testing needed	⚠️

Compatibility: ✅ works — verified in a live session · ⚠️ caution — works with caveats, not run end-to-end here · ❌ not compatible · ❔ not tested — not yet exercised in this environment. See Platform Compatibility Matrix for the full status legend.

Task lifecycle preview

The three skills map to three phases of a task's life. The Jira states below use the default Kanban board names (To Do / In Progress / In Review) — these are configurable per project, so map them to your own workflow's status names.

See Task lifecycle for the full phase-by-phase breakdown (skills, Jira states, and per-phase steps).

flowchart LR
    Kickoff([👤<br/>Phase 0 · Kickoff<br/>Human<br/>invokes /jira-task-assigner]) -->|feature · task · bug| Plan
    Plan([🤖<br/>Phase 1 · Plan<br/>jira-task-assigner<br/>To Do]) -->|create issues<br/>branches<br/>worktrees| Execute([🤖<br/>Phase 2 · Implement<br/>jira-task-executor<br/>In Progress])
    Execute -->|implement<br/>run tests<br/>open PRs| Review([🤖<br/>Phase 3 · Review<br/>jira-task-reviewer<br/>In Review])
    Review -->|changes requested<br/>back to In Progress| Execute
    Review -->|approved<br/>verdicts posted| Merge([👤<br/>Phase 4 · Merge<br/>Human<br/>Done])

Examples

JIRA-TASK-ASSIGNER

claude
> /jira-sdlc:jira-task-assigner "Refactor the InstantProductViewset create action. The action is currently separated into two perform_create methods. Investigate the code to determine whether this flow could be simplified. Additionally, check for any redundant code. Reference: cropapp/catalog/views.py, lines 1265–1676"

JIRA-TASK-EXECUTOR

# cd into each worktree it creates, run this in each one (no key —
# derived from that worktree's branch):
claude
> /jira-sdlc:jira-task-executor 

JIRA-TASK-REVIEWER

# once the sub-task's PR is up, run from the same worktree:
claude
> /jira-sdlc:jira-task-reviewer 

What's here

This repo currently hosts one plugin, jira-sdlc — three coupled skills (jira-task-assigner, jira-task-executor, jira-task-reviewer) that turn a feature request into Jira issues and git worktrees, implement each piece in parallel, and then review and merge the result as a single unit, leaving only the final release merge for a human. A fourth, jst-install, is the setup skill you run once before any of them.

This page is the front door. Everything about how the plugin actually works — architecture, prerequisites, configuration, a full usage walkthrough, safety model, and troubleshooting — lives in plugins/jira-sdlc/README.md.

The three skills, one per stage of the lifecycle:

• jira-task-assigner — turns a feature/task/bug description into Jira issues with matching git branches and worktrees. Investigates the codebase, asks clarifying questions, decides whether the request is one self-contained task or a multistep split into parallel sub-tasks, and gives every leaf issue its own branch and worktree so parallel work can start immediately.
• jira-task-executor — implements the issue implied by the current worktree's branch, end to end: status transition, investigation, implementation, tests, commit, push, and PR. No issue-key argument — run it from inside the issue's own worktree.
• jira-task-reviewer — run from the parent issue's worktree. Reviews each sub-task PR into the parent branch (approve or request changes), posts findings to Jira, and reviews the parent PR into the base branch once the sub-task PRs are merged. Never merges anything itself.

Plus one that runs before all three, once per project:

• jst-install — guided first-time setup. Walks Step-by-step installation — local tooling, GitHub repo prep, Jira board prep, healthcheck — verifying each with the bundled statuscheck script before moving on, so a missing development branch or a misspelled status name surfaces at setup rather than mid-run.

Full Setup

Everything to have in place before the first run — the three CLIs, both API tokens, the two settings files, and the branches and board your project needs — is a tickable list in Full setup checklist, ending in one command that verifies most of it for you.

Prefer it as prose? Step-by-step installation walks the same ground in the order you actually do it.

Prefer to be walked through it? /jira-sdlc:jst-install covers the same four sections interactively, checking each one before moving to the next.

Applications

Beyond the walkthrough above, the plugin is consumed two ways — as a Claude Code marketplace plugin, or as a loose skillset copied into a project's own skills folder — and this repo ships demo GitHub Actions workflows under docs/github-automations/example-workflows/ showing both consumption modes driving the three skills headlessly in CI, from a standalone reviewer gate on an open PR up to the full assigner → executor → reviewer chain on the feature and hotfix paths. Full detail, including which repository secrets each demo reads, is in GitHub Automations.

Four scenarios, each with its own walkthrough. None of them pauses for approval: each is gated only on who can trigger it (an OWNER comment), then runs start to finish on repository secrets.

• Feature flow — the whole assigner → executor → reviewer chain on the planned path: a GitHub issue becomes a Jira issue and a feature/* branch, gets implemented, and ends as an open, reviewed PR. Nothing is merged. Comment-triggered. demo-claude-feature-flow.yml (Claude Code · /make-feature) · demo-fcc-nvidia-nim-feature-flow.yml (Free Claude Code + NVIDIA NIM · /fcc-make-feature)
• Hotfix flow — the same chain on the emergency path: hotfix/* cut off <PRODUCTION_BRANCH>, PR aimed back at it, assigner forced single-step. Comment-triggered. demo-claude-hotfix-flow.yml (Claude Code · /make-hotfix)
• Review a PR — the reviewer on its own against an already-open PR, posting its verdict to GitHub and Jira and merging nothing. Comment-triggered. demo-claude-reviewer.yml (Claude Code · /review) · demo-fcc-nvidia-nim-reviewer.yml (Free Claude Code + NVIDIA NIM · /fcc-review)
• Issue to task / bug — the assigner alone: a commented issue becomes a Jira Task (or Bug) with its branch and worktree, and the run stops there. Comment-triggered ( / ), gated by the OWNER author check. (Claude Code · ) · (Claude Code · )

Jira states - who can move a card

The four anchor statuses (<STATUS_TODO>, <STATUS_IN_PROGRESS>, <STATUS_IN_REVIEW>, <STATUS_DONE>) are configurable — the tokens you map onto your board's real status names in .jst/jira-sdlc-tools.env. Who moves a card to which state — the three skills, GitHub Actions, a Jira automation app, or direct REST calls — is consolidated in Jira state movements.

Repository layout

jira-sdlc-tools/
├── .claude-plugin/
│   └── marketplace.json        # lists every plugin this repo offers
├── plugins/
│   └── jira-sdlc/              # the plugin itself
│       ├── .claude-plugin/
│       │   └── plugin.json
│       ├── skills/
│       ├── docs/
│       ├── LICENSE
│       └── README.md           # full plugin documentation
├── .jst/                       # settings folder — the only location the skills read
│   ├── jira-sdlc-tools.env     # template — team-shared settings (committed)
│   └── jira-sdlc-tools.local.env.example  # template — machine-specific (gitignored)
├── AGENTS.md                   # repo-wide instructions for AI coding agents
├── CLAUDE.md                   # imports AGENTS.md + Claude Code–specific notes
├── LICENSE
└── README.md                   # this file

It's split into a marketplace layer (this level) and a plugin layer (plugins/jira-sdlc/) so the marketplace can grow to host more Jira/SDLC-related plugins later without another reorganization — right now there's just the one.

Development

Editing a skill needs a tighter loop than a marketplace install gives you: Claude Code copies a plugin snapshot into its cache at install time, so changes to your clone won't show up in an installed copy until you reinstall.

1. Clone the repo:

   git clone https://github.com/kantorv/jira-sdlc-tools.git
   cd jira-sdlc-tools
   

2. Load it manually, pointing at the plugin's own root — not the toolkit repo root, which only holds marketplace.json:

   claude --plugin-dir ./plugins/jira-sdlc
   

   No install step, no marketplace. If jira-sdlc is already installed from a marketplace elsewhere on the same machine, --plugin-dir takes precedence for that session, so you're never testing against a stale cached copy without realizing it.

3. After each edit, reload instead of restarting:

   /reload-plugins
   

   Picks up changes to skills, agents, hooks, and MCP/LSP servers without a full session restart.

There's no build or test suite to run — these are prompt files for an LLM agent plus two JSON manifests, not compiled code. See AGENTS.md for what actually counts as validating a change. Note too that this toolkit repo isn't a valid target for its own skills — you'll need a separate application repo, with its own .jst/jira-sdlc-tools.env, to actually exercise one against.

Run your own

For anyone who wants to customize the skills and run their own version rather than track this one.

This repo is MIT-licensed and is already a ready-to-use Claude Code marketplace — it's only a few clicks from being cloned or forked into your own copy, then registered as a marketplace in your local Claude Code. Fork or clone it, edit the skills to taste, and install from your local folder:

1. Fork it on GitHub (if you want your own upstream to push to), then clone your copy:

   git clone https://github.com/<you>/jira-sdlc-tools.git
   

2. Add the local folder as a marketplace, pointing at the repo root (the directory holding .claude-plugin/marketplace.json):

   /plugin marketplace add ./jira-sdlc-tools
   /plugin install jira-sdlc@jira-sdlc-tools
   

   From here it behaves like the remote install — your fork is the source. While actively editing, use the --plugin-dir edit-reload loop above instead, since a marketplace install copies a snapshot into the cache.

Contributing

Read AGENTS.md first, especially if an AI coding agent is doing the work — it covers the constraints that are easy to break without realizing it (the _shared/ reference-path relationship, what else to update if you rename a skill or the plugin, how to validate a change with no test suite to run).

License

MIT, covering the whole repo, including the plugin.

Lab channel

Everything above describes the main channel — this repo's default branch, and what every install command on this page gives you: the three core skills, reviewed, released, and tagged.

The lab channel is the same plugin sourced from the lab branch instead. It's kept synced with the default branch, so it's never behind main — it's main plus work that hasn't landed yet: more advanced scripts, wider permissions, and an extra skill.

Name	Type	Description	Example reports
conversation-debugger	skill	Post-mortems a recorded run of one of the three core skills against its own prose, verdicting each instruction as followed / diverged / skipped / not-reached.	TBD
feature_report	script	Rolls a whole feature's runs up into a single report.	TBD

Example reports

Report	Generated by	Source	Description
jira-task-assigner-JST-126	ai	conversation-debugger	Single-run post-mortem of a jira-task-assigner conversation for JST-126, verdicting each instruction as followed / diverged / skipped / not-reached.
jira-task-executor-JST-125	ai	conversation-debugger	Single-run post-mortem of a jira-task-executor conversation for JST-125, run on the Windows/PowerShell dispatch path.
jira-task-reviewer-JST-122	ai	conversation-debugger	Single-run post-mortem of a jira-task-reviewer conversation for JST-122, run on the Windows/PowerShell dispatch path.
JST-122-story-report	script	feature_report	Feature-wide rollup for multistep JST-122 (parent + 3 child issues, 15 conversations) built from collect_feature JSON — token/cost/model totals, nothing re-estimated.
JST-126-task-report	script	feature_report	Feature-wide rollup for single-issue JST-126 (3 conversations) built from collect_feature JSON — token/cost/model totals, nothing re-estimated.

To install it, suffix the marketplace repo with @lab:

/plugin marketplace add kantorv/jira-sdlc-tools@lab
/plugin install jira-sdlc@jira-sdlc-tools

Worth knowing before you switch: the extras aren't release-gated, and they reach wider than the core three do — into your whole workspace rather than a single issue's worktree, with the scripts and permissions to match.

See LAB-CHANNEL.md on the lab branch for the full description, both install routes, and the lab-only configuration step.