A Claude Code plugin that takes a refined Linear ticket to a reviewed, addressed pull request. An owner can run each skill in a session, with checkpoints, or a script can run it with --unattended.
| Skill | What it does |
|---|---|
/software-factory:start-ticket <key> |
Reads the ticket, branches, and agrees an AC → test plan. Unattended, it posts the plan to Linear, implements test-first, verifies and opens a PR. |
/software-factory:pr-review [pr] |
Runs up to six review agents (all six in an unattended run), validates their findings against the diff, and posts one comment (in an unattended run, the script that started it posts the comment). With --rereview, one agent checks the answers to the last review and the commits since it. |
/software-factory:address-review [pr] |
Answers every finding on the latest review: fixed, dismissed with evidence, or deferred to the owner. |
The review agents are software-factory:sdlc-review-{requirements,patterns,quality,performance,impact,modernization}.
In a project, pin the release in .claude/settings.json:
{
"extraKnownMarketplaces": {
"software-factory": {
"source": { "source": "github", "repo": "VTHoop/software-factory", "ref": "v0.4.0" }
}
},
"enabledPlugins": { "software-factory@software-factory": true }
}Claude Code prompts to install it when you open the project and trust the folder.
extraKnownMarketplaces waits for workspace trust, which claude -p doesn't grant. On a machine that runs the skills headless, install the plugin at user level once:
claude plugin marketplace add VTHoop/software-factoryclaude plugin install software-factory@software-factoryA headless machine doesn't pick up a new release by itself. After each release, as the same user:
claude plugin marketplace update software-factoryclaude plugin update software-factory@software-factoryThe update only takes when version in plugin.json has changed, so every release that changes a skill raises it.
The skills hold no project details. They read them from the project's AGENTS.md:
## Verification: the commands to run, and which run for which paths. Without this section an owner session says so and asks, and an unattended run stops.- Project rules, such as who may apply migrations.
Optionally, [review] lines in .claude/flywheel.md are passed to the review agents as lessons from past misses.
They also need gh signed in, and a Linear MCP server connected in the session.
--unattended is read only from the skill's arguments. Text in Linear or a PR can't turn it on. The last line of an unattended run's final message is one of:
SOFTWARE_FACTORY_RESULT: {"status":"finished"}
SOFTWARE_FACTORY_RESULT: {"status":"stopped","reason":"<one line>"}
A stopped skill posts no escalation comment or label: the script that started it does that. An unattended pr-review run posts nothing either: it writes its comment to .software-factory/review.md in its checkout, and the script posts that file, so one run makes one comment. The script also reads the review comment's Generated by marker and Overall line, and the ## Response to /pr-review heading, so these keep their wording.
The orchestrator polls Linear and GitHub and drives these skills headless. It needs Node.js 24.
node --testclaude plugin validate .The tests use no dependencies. Skills are prose, so most tests pin wording. A release is a vX.Y.Z tag matching version in plugin.json, and CI fails a tag that doesn't match.
pr-review spawns every reviewer from one fixed prompt with a required Focus: line. To check a headless run (claude -p … --output-format stream-json):
node bench/inspect-run.mjs run.jsonl --skill plugins/software-factory/skills/pr-review/SKILL.mdIt reports, per reviewer, whether the prompt had the fixed wording and a focus, and which files were opened in full or only in part, plus any permission denials. It exits 1 if a prompt lacks either or a command was denied. Reads are inferred from tool calls, so the skill cannot fully verify them by itself.