wise-claude

A coding copilot for Claude Code - flat /wise-* skills, a
multi-harness workflow engine, and autonomous git / PR / ticket-planning
pipelines.

wise-claude is the Claude Code plugin marketplace that hosts the
wise copilot: flat /wise-* skills, YAML workflows run by a
Python engine that drives claude, codex, cursor-agent, gemini and grok
headless, and an SDLC agent roster (CEO / CTO / architect / engineer / QA /
security / SRE / …) that take everyday engineering chores off your
plate — drafting commits, opening and shepherding PRs, planning tickets,
authoring PRDs/TRDs, auditing a scope into an executable backlog. Reach
for a single quick command, or hand off a whole ticket → merged-PR
pipeline to run unattended.

Install

Claude Code:

/plugin marketplace add e1024kb/wise-claude
/plugin install wise@wise-claude

Codex, from a terminal rather than the Codex prompt:

codex plugin marketplace add e1024kb/wise-claude
codex plugin add wise@wise-claude

Then run the wise-init skill to prepare Python and register the workflow
engine for the current host, and /wise to print the command catalog.

For Cursor, Grok, or T3 Code, load Wise's skills through that host's supported
skill installation mechanism, then run wise-init from the loaded installation.
The host setup guide covers
registration and workflow control. The conductor host does not determine which
provider CLI runs workflow steps.

What you get

Skills

• Git & commits — /wise-commit-message (draft only), /wise-commit
  (draft + commit), /wise-commit-push (draft + commit + push), all
  Conventional-Commits aware.
• Pull requests — /wise-pr-create, /wise-pr-add-reviewers,
  /wise-pr-watch (drive CI + review comments to green).
• Planning & docs — interactive ticket planning, /wise-grill (deep-
  research a ticket / doc / prompt into a plan, blueprint, or answer),
  /wise-revise (audit a scope and write executable improvement plans
  into docs/plans/), the model-invoked wise-prd-architect /
  wise-trd-architect document authors, and the wise-estimation
  story-point reference.
• Authoring — /wise-skills-create, /wise-skills-edit, and the
  /wise-workflow-* family for building and running your own workflows.
• The /wise helper — type /wise <free-form text> (e.g. /wise open a PR) and it classifies the request and offers the matching command.
• Autonomous -auto building blocks — decision-free, prompt-free
  variants (/wise-pr-create-auto, /wise-implement-plan-auto,
  /wise-simplify-auto, …) used by the unattended pipelines.

Workflows (multi-step, multi-agent)

• ticket-auto — autonomous ticket → plan → implement → review → PR →
  watch CI → resolve review bots → merge, with no prompts.
• impl-plan-auto — same autonomous pipeline, but fed a ready
  PLAN-*.md (e.g. one /wise-revise wrote): re-plan from the file →
  implement → review → PR → watch → merge.
• pr-watch / impl-plan — the engine behind /wise-pr-watch-auto
  and /wise-implement-plan-auto: watch an existing PR to merge, or land
  a plan's tasks as commits on the checked-out branch, with harness,
  model and effort chosen per phase at pre-flight.
• ticket-plan — autonomous planning you review and adjust before you
  implement.
• code-review — the pre-push branch gate: three parallel reviewers,
  a curator, an optional verification pass, and a fixer that commits the
  kept findings, with harness, model and effort chosen per agent.

Workflows are YAML v2 definitions (agent, bash, approval, ask,
units steps, tuning groups asked in stages). The engine under
plugins/wise/engine runs them as a per-user daemon and talks to Claude
Code through the plugin's wise-engine MCP server; the conversation
renders pre-flight questions and gates through the main client's native GUI/TUI
or rendered MCP forms. When neither structured route is usable, the main harness
collects explicit answers in chat. Missing controls never launch a terminal.

See the plugin README for the full command
reference and docs/wise/ for the workflow engine, the
/wise dispatcher, and the skill-authoring guides.

Requirements

• Claude Code, Claude Desktop, Codex, Cursor, Grok, or T3 Code as the
  workflow conductor.
• git, and an authenticated gh CLI for the PR skills.
• Python 3.11+ for the workflow engine. The launcher manages pinned
  dependencies outside the plugin installation; no Bun, Node or npm is required
  by the engine.
• Authenticate the provider CLI you select (claude, codex, cursor-agent,
  gemini or grok). Subscription and API authentication use that provider's CLI.
• Run wise-init for each conductor host. Its stable managed registration
  avoids host-specific plugin-root expansion and stores dependencies outside
  read-only plugin caches.

Only YAML v2 workflows execute. Import v1 definitions with engine.sh migrate;
legacy run files remain unchanged and cannot resume. Standalone profile,
session and supervision commands use scripts/wise-helpers.py.

Repository layout

wise-claude/
├── .claude-plugin/marketplace.json      # Claude Code marketplace index → plugins/wise
├── plugins/wise/                        # the wise plugin (skills, agents, workflows, engine/)
├── scripts/validate_repo.py             # structural validation
├── docs/wise/                           # workflow engine + authoring reference
├── justfile                             # task runner (validate / test / check)
└── CONTRIBUTING.md                      # full contributor manual

Troubleshooting

• Skills / commands don't show up — confirm /plugin install wise@wise-claude ran, then start a fresh session.
• PR / workflow steps fail on auth — run the wise-init skill;
  make sure gh auth status is green and an origin remote exists.
• /wise can't classify a request — type the /wise- prefix to browse
  every command in the menu.

Contributing

Issues and PRs are welcome — see CONTRIBUTING.md for
the conventions, local-install steps, and validation checks. You can
also file feedback from inside the agent with /wise-feedback.

License

MIT © e1024kb