Claude Code plugin: a spec-driven delivery pipeline.
/spec → /taskplan → /implement → /cleanup
(PRD) (partitioned (parallel (PR, CI poll,
plan JSON) orchestrated merge, teardown)
build)
Plus a second delivery path for deep, coupled work — GitHub tickets fanned out to parallel Orca worktrees, one agent per ticket:
(mattpocock/skills) (this plugin)
/to-spec → /to-tickets → /frontier → /cleanup
(1 issue) (N issues + (N worktrees, (PR "Closes #n",
blocked_by edges) 1 agent each) CI, merge)
See docs/flows/tickets-to-orca.md for when to take which path.
Plus /project-kit — scaffolds docs/pm/ project management (status, decisions, schemas, dashboard) into any repo.
| Skill | Stage | What it does |
|---|---|---|
/spec |
define | Interview-driven PRD at docs/spec/<slug>.md — objective, structure, testing, boundaries — finished by a delegated fresh-eyes review (opus) that revises the doc in place. |
/taskplan |
plan | Decompose a spec into atomic tasks with file ownership, model tier, and method; derive parallel waves into docs/plan/<slug>.json. |
/implement |
build | Orchestrate the plan: file-disjoint waves of parallel code-implementer subagents, integrate, verify, review. Logs every run (orchlog.py). |
/implement-orca |
build | Same orchestration doctrine, but workers are Orca-dispatched claude CLI terminals (via the official orchestration skill): visible panes + runtime task/dispatch provenance. |
/frontier |
build | Ticket-level parallelism: query GitHub for takeable tickets (open, ready-for-agent, unblocked, unassigned), claim by assignee, dispatch one Orca worktree + claude worker per ticket. Full handoff — no coordinator. |
/cleanup |
ship | Open the PR, poll CI to conclusive (pollci.py), auto-merge on green, tear down the worktree + branches. |
/project-kit |
setup | Scaffold docs/pm/ (status.json, decisions, schemas, dashboard) and the CLAUDE.md block. |
Two independent version numbers, on purpose:
.claude-plugin/plugin.jsonversion— the plugin's semver (what/plugininstalls and updates).WORKFLOW_VERSIONinskills/implement/orchlog.py— the run-log schema version, stamped into everyorchlogrecord. It bumps only when the orchestration doctrine or log semantics change, so runs stay comparable across plugin releases (only compare metrics within the same schema version).
/plugin marketplace add psycoder-sup/workflow
/plugin install workflow@workflow