Skill pack that drives development from Redmine issues.
These skills shine when mcp-redmine is installed and connected (SSE MCP for the Redmine REST API). Journaling, status / done_ratio updates, refine / split gates, and the everyday ticket-driven loop assume an agent can call Redmine through MCP.
Without mcp-redmine you can still use ./tools/redmine_helper.sh as a CLI fallback, but the full agent workflow is designed for mcp-redmine.
Typical agent path: New → In Progress → Review → Resolved.
- After implementation checks, move to Review (do not resolve from In Progress).
- Run
ticket-evaluatewhile the issue is in Review (評価:notes only; no status change). Plan evaluation after split is also supported. Criteria include quality and security note lanes (blockers only for real safety/data harm; not sole grounds for fail). - When a separate evaluation agent is defined (e.g. host skill
ticket-driven-ask-*-eval), request plan eval right after split and implementation eval after Review — do not self-write評価:(request-eval.md). - On fail: implementer returns Review → In Progress. On pass: close-audit, then Review → Resolved.
- Redmine “Feedback” stays for human pushback; LLM evaluation uses Review. Confirm numeric
status_idviaredmine_metadata(this environment’s Review is often7).
DoD, start notes, and completion notes record what was done, not why. When a ticket involves a design choice (there were alternatives / a value, parameter, or approach was picked / the plan changed mid-work), record the decision, reason, rejected options, and assumptions.
- Record: decisions made during refine go in the description’s
## 判断section; decisions made or changed after start go in a判断:journal note (ticket-driven→sections/journal.md, “判断の記録”). - Refine: sufficiency check includes “判断の根拠” (decision rationale). When rewriting the description, rationale from the old text (measurements, user instructions, reasons for rejecting options) must not be deleted — only moved to
背景(background) or判断(ticket-refine→sufficiency.md). - Evaluate: plan criterion P3 (missing rationale → conditional pass; deleted rationale → fail) and implementation criterion I5 (plan changed without a
判断:note → note-lane finding). - Skip it for tickets with no real choice (no boilerplate). Close-audit and delegate-pack fields are unchanged.
npm installcp .opencode/opencode.jsonc.example .opencode/opencode.jsonc.opencode/opencode.jsonc is gitignored. Edit for your environment:
| Item | Description |
|---|---|
REDMINE_URL / REDMINE_API_KEY |
Local Redmine |
instructions paths |
Machine-local paths (replace example placeholders). Optional memory / other MCP entries are yours to add — not required |
export REDMINE_URL=http://127.0.0.1:3000
export REDMINE_API_KEY=your_key
./tools/redmine_helper.sh get 2cp .cursor/mcp.json.example .cursor/mcp.json.cursor/mcp.json is gitignored. Edit for your environment:
- Project:
.cursor/skills/→ symlinks intoskills/(shipped in-repo) - Personal: symlink the same names under
~/.cursor/skills/for other workspaces - MCP:
.cursor/mcp.json- Connect mcp-redmine for the full ticket-driven loop (URL only; keep API keys out of client config)
- Other MCPs (long-term memory helpers, etc.) are optional — skip them if you do not use them
- Or
cp .env.example .envand fill keys (.envis gitignored) - Reload the Cursor window after changes
Canonical skills live in skills/. Symlink; do not copy:
./tools/sync-skill-links.shWarning: This script writes under your HOME (global agent paths), not only inside the repo (~/.agents/skills, ~/.hermes/skills/software-development, ~/.cursor/skills, ~/.pi/agent/skills). Existing same-named dirs may be backed up then replaced. Skip the script if you only cloned for inspection or do not want HOME changes.
| Runtime | Location |
|---|---|
| OpenCode | ~/.agents/skills/<name> |
| Hermes | ~/.hermes/skills/software-development/<name> |
| pi (host) | ~/.pi/agent/skills/<name> |
| pi (compose) | bind-mount of canonical skills/ticket-* (see compose docker-compose.yml; not this script) |
Restart OpenCode / Hermes / pi if needed.
Hermes agent.system_prompt is a snapshot. When your private bootstrap / core rules change, re-sync with your usual Hermes sync tool (if installed), then restart Hermes yourself.
npm run validate:opencodeCanonical files are skills/<name>/SKILL.md. Index: skills/navigation-protocol.md.
| Skill | Purpose |
|---|---|
ticket-driven |
End-to-end ticket-driven workflow (incl. request-eval when a separate eval agent is defined) |
ticket-create |
Create issues |
ticket-list |
List / filter issues |
ticket-refine |
Refine requirements before split |
ticket-relation |
Issue relations |
ticket-split |
Split by capability → session-sized children |
ticket-evaluate |
Plan or implementation review vs DoD → 評価: journal (種別: 計画|実装; no status change). Note lanes: quality (Q1–Q5), security (S1–S5), UI when applicable. Decision rationale (plan P3, implementation I5) |
ticket-status |
Status-based actions |
ticket-update |
Update fields / description |
| Skill | Purpose | When |
|---|---|---|
git-commit |
Commit procedure | User explicitly asks to commit; use with Cursor commit rules |
git-pr |
PR creation (gh / GitBucket API) | GitBucket curl flow, or GitHub via gh |
git-branch |
Branch naming / switching | Project conventions |
git-rebase |
Rebase / conflicts | Conflict resolution |
git-diff |
Diff / review aids | Review checklists |
git-log |
History search | Investigation |
Cursor user rules own the default commit/PR safety steps. This repo’s git-* skills add GitBucket-specific and project conventions.
- Git host:
http://{git-host}:{port}or GitHub (environment-specific) - Redmine:
http://127.0.0.1:3000(or your own)
skills/<name>/SKILL.md # canonical skills
skills/navigation-protocol.md
.cursor/skills/ # Cursor symlinks
.cursor/mcp.json.example # Cursor MCP template (tracked)
.cursor/mcp.json # local MCP config (gitignored)
.env.example
.opencode/
opencode.jsonc.example
opencode.jsonc # local secrets (gitignored)
tools/
redmine_helper.sh # Redmine CLI when MCP is down
validate-opencode.mjs
AGENTS.md
README.md
README-JP.md
LICENSE