claude

flywheel

Nested spec→plan→work→verify→review→compound loop for disciplined AI-assisted development, plus an agent-native runtime pillar (/flywheel:process + /flywheel:run) that turns recurring domain operations into self-improving, persistence-aware process contracts Claude executes as the backend. Also an autonomous metric-driven loop, parallel multi-specialist review, and spec↔code sync. Plans route each task to a model/effort tier so cheap work runs cheap and judgment runs strong. /flywheel:route does the same for work delegated outside a plan: tool, fresh session or subagent, at which model and effort.

0 on GitHubby arazvanv0.87.0MITupdated 1 month agoworkflow
/

Plugin flywheel from marketplace xmarks in arazvan-ec/xmarks. Commands read from the repo's manifest — 1 month ago.

This plugin runs code on your machine — its hooks fire automatically on events like tool use. The exact commands are listed below; read them before installing.

What it installs

Skills (18)

  • autoloop

    Autonomous metric-driven loop — change → measure → keep-if-better against one machine-checkable metric, within a fixed iteration budget. Use only for objectively measurable goals (suite green, lint errors to zero).

  • brainstorm

    Sharpen a vague or fuzzy idea into agreed requirements — key questions, 2–3 approaches with trade-offs, converge on scope. Use before /flywheel:spec when the idea isn't clear yet.

  • compound

    Capture this cycle's decisions, gotchas, patterns, and fixtures into the learnings ledger so the next cycle starts smarter. Use at the end of a unit of work, after review.

  • debug

    Systematic debugging — reproduce, one hypothesis, isolate, fix the root cause, add a regression test. Use when something is broken or a test fails mysteriously.

  • help

    Onboarding and command reference — the nested loop, every /flywheel command, and where to start. Use when someone is new to flywheel or asks how it works.

  • loop

    Run the full development cycle (spec → plan → work → verify → review → compound) for one unit of work, gating between phases; small clear tasks route through a collapsed micro-cycle. Use to start a new feature or task.

  • plan

    Turn an approved spec into ordered tasks, each with its own pass/fail check and a model/effort route. Use after /flywheel:spec sign-off, before writing code.

  • process

    Define or revise an agent-native process contract (fixed rules + output schema + persistence) for a recurring domain operation Claude runs as the backend. Use when a repeatable operation ('analyze a car', 'score a lead') should become something /flywheel:run executes.

  • recall

    Search the learnings ledger on demand — list matches cheaply, expand one on request. Use for a past decision, gotcha, pattern, or fixture the session-start injection didn't surface.

  • review

    Adversarial multi-specialist code review, routed by diff type (correctness always; security/performance when their domains are touched), dispatched in parallel, synthesized into prioritized findings. Use after verify passes, before finishing.

  • route

    Recommend how to run a piece of work you are about to delegate — tool, fresh session, subagent or here — and at which model and effort. Use before create_session, an Agent call or a /flywheel:run execution, whenever the work is outside a plan's routed tasks.

  • run

    Execute a process contract as the runtime — follow its fixed rules, persist to the repo's datastore per DATA.md, prove the write, mature the contract. Use to run an operation defined with /flywheel:process; Claude is the backend.

  • ship

    Close out a unit of work — clean commit, push, pull request. Use at the very end of a cycle, after verify passes and compound is done.

  • spec

    Turn a feature request into a REASONS spec-contract with one machine-checkable success metric. Use at the start of a unit of work, before planning or writing code.

  • sync

    Detect and reconcile spec↔code drift — code the spec doesn't reflect, spec items not implemented — then fix whichever is wrong, with confirmation. Use after code has evolved past its original spec.

  • update

    Update flywheel itself — autodetects marketplace vs vendored install, refreshes it, applies upgrade strategies, reports old → new version. Use to upgrade flywheel or check which version a repo carries.

  • verify

    Objective PASS/FAIL gate — run the app and the test suite against the spec's success metric, with evidence. Use before review/ship, or to confirm a change really works rather than assuming it does.

  • work

    Implement plan tasks with the iterate-until-green inner loop — failing test, minimal implementation, run, observe, fix — never done until the objective check is green. Use when executing tasks from an approved plan.

Subagents (6)

  • evaluator

    Independent cross-check for autoloop's stop decision — re-runs the metric command itself (does not trust the working agent's transcript) and returns continue/stop + reason. Invoke on ambiguous keep/discard results and before autoloop declares its target met.

  • executor

    Executes one fully-specified mechanical plan task (rename, move, config/doc edit, fixture from a recorded recipe, run-and-report) on a fast model at low effort, and refuses with ESCALATE when the task actually needs a judgment call. Invoke for a plan task routed `haiku/low+delegate`.

  • reviewer-correctness

    Adversarial reviewer focused on correctness — logic bugs, edge cases, error handling, race conditions, and test adequacy. Invoke (usually in parallel with the other reviewers) to review a diff before shipping.

  • reviewer-performance

    Adversarial reviewer focused on performance — hot paths, N+1 queries, unnecessary work, and resource use. Invoke (usually in parallel) to review a diff before shipping, especially when it touches loops, queries, or data volume.

  • reviewer-security

    Adversarial reviewer focused on security — credential/secret handling, injection, authn/authz, unsafe input, and data exposure. Invoke (usually in parallel) to review a diff before shipping, especially when it touches auth, external APIs, or user data.

  • verifier

    Runs the actual application and test suite and returns an objective PASS/FAIL verdict with evidence, checked against a stated success metric. Invoke to gate a phase or to confirm a change genuinely works (not just looks right).

Hooks (12)

  • SessionStart

    "${CLAUDE_PLUGIN_ROOT}"/scripts/session-start.sh

  • SessionStart

    "${CLAUDE_PLUGIN_ROOT}"/scripts/git-tracking-refs.sh

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}"/scripts/read-prime.sh

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}"/scripts/write-allow.sh

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}"/scripts/bash-allow.sh

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}"/scripts/delegation-guard.sh

  • Stop

    "${CLAUDE_PLUGIN_ROOT}"/scripts/gate.sh

  • Stop

    "${CLAUDE_PLUGIN_ROOT}"/scripts/toolbar.sh stop

  • Stop

    "${CLAUDE_PLUGIN_ROOT}"/scripts/compound-due.sh

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}"/scripts/delegation-record.sh

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}"/scripts/read-meter.sh

  • UserPromptSubmit

    "${CLAUDE_PLUGIN_ROOT}"/scripts/toolbar.sh remind

View source on GitHub

More from arazvan-ec/xmarks

All 14 plugins →

Mod: the Ministry watches every byte — live bytes read, cost and context in the status line; toasts for oversized reads and full rewrites; /ministry dossier.

0

Mod: every exception is contraband — SKIP_*/--no-verify/Release-Exception use is ledgered with its reason, and /black-market audit totals the standing permits.

0

Mod: your file, opened — /expediente shows a flywheel cycle's run record live: every transition, route and escalation, with totals.

0

Mod: three reds of the same check with edits between and no green — the workers strike; edits stop until you debug systematically.

0

Mod: a list you were handed is a file, not a memory — a prompt with a list holds edits until it is written to a plan or scratchpad file.

0

newspeak

claude

Mod: ungood items — a list item with no success criterion is named, and the model is told to ask for one before doing it.

0

Similar plugins

Matt Pocock's agent skills for real engineering (grilling, spec/ticket flows, TDD, code review, domain modelling and more).

39 skills

mattpocock/skills 323,108

superpowers

claudecodex

Core skills library for Claude Code: TDD, debugging, collaboration patterns, and proven techniques

15 skills · hooks

obra/superpowers 316,836

Backend API design, GraphQL architecture, workflow orchestration with Temporal, and test-driven backend development

9 skills · 1 command · 8 agents

wshobson/agents 42,461

beads

claudecodex

AI-supervised issue tracker for coding workflows

1 skill · 1 agent · hooks

gastownhall/beads 28,698

PRD generation and conversion skills for the Ralph autonomous agent loop

2 skills

snarktank/ralph 22,224

deepeval

claude

Skills for adding DeepEval evaluations, tracing, datasets, Confident AI reports, and iterative improvement loops to AI applications.

3 skills

confident-ai/deepeval 19,466

Maintain flywheel?

This page updates itself from your repo. Add the badge to your README, or get it Featured to pin it above the organic results with a followed link to your site.

claude & codex plugins14 plugins