gh-issue-driven
English | 日本語
Full-lifecycle GitHub-issue-driven dev workflow: propose issues from session context, design-gated start, advisor-gated ship (with opt-in post-PR review) and HITL confirmation at every phase boundary, ceremony-automated tag — with multi-issue batch support, pluggable post-PR reviewer, and per-repo Kagura Memory auto-detect.
gh-issue-driven is a Claude Code plugin — but think of it as an orchestration harness: a thin meta-layer that wires Claude Code's own skills and agents (design advisors, memory recall, /code-review, Copilot) into one governed, repeatable issue → release pipeline with a human-in-the-loop checkpoint at every phase boundary. It's not a code generator or a single command — it's the layer that coordinates those skills with governance gates and memory.
It turns "I'm starting work on issue #142" into a single, repeatable three-phase workflow:
graph LR
P["/propose"] -.-> I[Issue #N]
I --> S["/start<br/>Gate 1: design"]
S --> IMP["implement<br/>+ /code-review"]
IMP --> SH["/ship<br/>Gate 2 + PR"]
SH --> T["/tag<br/>release ceremony"]
T --> R[GitHub Release]
P -.HITL.-> H0((pause))
S -.HITL.-> H1((pause))
SH -.HITL.-> H2((pause))
style H0 fill:#ffe4b5
style H1 fill:#ffe4b5
style H2 fill:#ffe4b5
style I fill:#e6f3ff
style R fill:#d4edda
/gh-issue-driven:start <issue...> — fetch the issue(s), recall related past work from Kagura Memory, run a gate 1 design review (/claude-c-suite:ask cascading to /ceo for complex issues), create a typed feature branch, and hand off for implementation. Pass multiple IDs to batch issues into one branch.- (you write the code, then run
/code-review to review the working-tree diff — the pre-PR review, before /ship opens a PR)
/gh-issue-driven:ship — run a gate 2 parallel review battery (cso + qa-lead + cto advisors by default; an optional binary release gate can be configured via gate2.binary_gate), then create the PR. Post-PR review is opt-in: review.provider defaults to none, so /ship creates the PR and stops unless a provider is configured (Copilot, /code-review, or both) — or /objective drives the review for you. When a provider is set, /ship delegates the review loop to /gh-issue-driven:review (its canonical home), which drives the loop until the PR is approved or no actionable feedback remains. A HITL confirmation gate pauses before the review loop starts, so you can decide whether to invoke it for this PR./gh-issue-driven:tag <version> — release ceremony: compose label-grouped release notes from the milestone, bump plugin.json + marketplace.json, update CHANGELOG.md, commit, annotated-tag, push with --follow-tags, and create the GitHub Release. dry-run previews everything without touching files, git, or GitHub.
The whole flow is bracketed by kagura-memory session-start and session-summary with auto-detect context setup on first run, so each issue's learnings get persisted for future recall.
📖 Read the design philosophy: Qiita article — Issue→Release を自動化したら、逆に人間が重要になった話 · Slides
60-second quickstart
# Step 0 — one-time, in your repo's GitHub Settings page:
# Settings → Code review → ☑ Automatic Copilot code review
# URL: https://github.com/<owner>/<repo>/settings/code-review
# This makes the Copilot review loop work on any gh CLI version.
# Without it, you need gh CLI >= 2.88.0 (see Requirements below).
# In any Claude Code session — install the plugin:
/plugin marketplace add JFK/gh-issue-driven
/plugin install gh-issue-driven
# Recommended companion plugins (gracefully degrade if missing).
# Note: the install target's `@<marketplace>` is the marketplace NAME from
# its marketplace.json, not the GitHub repo slug.
/plugin marketplace add JFK/claude-c-suite-plugin # gate1 + gate2 reviewers
/plugin install claude-c-suite@claude-c-suite
/plugin marketplace add kagura-ai/memory-cloud # session-start/summary + recall
/plugin install kagura-memory@kagura-memory-cloud
# Optional (reserved for v0.2 deep-review modes):
/plugin marketplace add JFK/claude-phd-panel-plugin
/plugin install claude-phd-panel@claude-phd-panel
