improve-solo

improve-solo

A fork of shadcn/improve — an agent skill that audits any codebase and writes implementation plans for other agents to execute — with pluggable persistence. When Solo is available it stores and dispatches natively through Solo — auto-detected, no flag needed: plans become scratchpad + todo pairs, dependencies become real todo blockers, executors are Solo agents, and review wake-ups use Solo idle timers. With --files it behaves exactly like upstream (plans as files).

The idea is unchanged: use your most capable model for the part where intelligence compounds — understanding the codebase, judging what's worth doing, writing the spec — and hand execution to cheaper models. The skill never implements anything itself. The plan is the product.

you          →  /improve-solo                (expensive model, advises)
plans        →  Solo scratchpads + todos     (auto-detected; --files for plans/*.md)
other agent  →  implements, tests, ships     (cheap model, executes; you merge)

Storage backends

Exactly one primary store per run; --issues is additive on top of either. Solo is auto-detected; --files opts out.

	solo (default when available)	--files
Plan body	scratchpad Plan NNN: <title>	plans/NNN-slug.md
Index / status	the todo list itself (improve-plan tag)	plans/README.md table
Dependencies	native todo blockers	prose in the index
BLOCKED / REJECTED	tags blocked / rejected + reason comments	status column
Audit memory (rejected findings)	per-run audit-report scratchpad	plans/README.md section
Executor (execute)	spawn_agent in an advisor-created worktree	host subagent, isolated worktree
Executor report	todo comment	subagent's final reply
Review wake-up	timer_fire_when_idle_any	inline

One backlog, never two: if prior plans exist in the other store, the skill says so and follows the existing store rather than forking the backlog.

Installation

Claude Code — as a plugin (recommended there: versioned updates via /plugin update):

/plugin marketplace add codeSTACKr/improve-solo
/plugin install improve-solo@improve-solo

Any other agent (Cursor, Codex, Gemini CLI, Amp, OpenCode, …) — as a plain skill via the skills CLI, same as upstream:

npx skills add codeSTACKr/improve-solo

The skill is agent-agnostic: the Solo store works from any MCP-capable agent connected to Solo, and where the host can't spawn subagents the workflow says so and degrades gracefully (direct audit, manual plan handoff instead of dispatched executors).

In Claude Code, pick one mechanism — a skills-CLI copy in ~/.claude/skills/improve-solo and the plugin register the same skill name and will conflict or silently shadow each other. For local development, point the marketplace at your checkout: /plugin marketplace add /path/to/improve-solo.

Usage

/improve-solo                        full audit → prioritized findings → plans (Solo when available, else files)
/improve-solo --files                plans as local files (upstream behavior)
/improve-solo quick|deep             effort level
/improve-solo security               focused audit (also: perf, tests, bugs, ...)
/improve-solo branch                 audit only what the current branch changes
/improve-solo next                   feature suggestions — where to take the project
/improve-solo plan <description>     skip the audit, spec one thing
/improve-solo review-plan <plan>     critique and tighten an existing plan
/improve-solo execute <plan>         dispatch a cheaper executor, review its work
/improve-solo reconcile              refresh the backlog: verify, unblock, retire
/improve-solo ... --issues           also publish plans as GitHub issues

All modifiers compose: /improve-solo deep security --files --issues.

How the Solo store works