claude

supervisor

Token guardrails for sessions on Fable-tier models: writes the task brief from a short interview, pins spawns to cheap workers, denies forks, requires a brief for expensive spawns, prices a per-session budget from the transcript, and holds workers to an evidence contract.

0 on GitHubby Oleg Konykv2.6.0MITupdated 14 days agoproductivity
/

Plugin supervisor from marketplace konyklabs-plugins in konyklabs/claude-plugins. Commands read from the repo's manifest — 12 days 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 (11)

  • brief

    Interviews the developer about a task with at most five structured questions and writes the brief the supervisor flow needs — task, checkable definition of done, evidence command, out of scope, decisions, assumptions, procedure — then lints it. Use before any non-trivial task in an expensive-model s

  • budget

    Shows this session's priced spend per model and subagent from the supervisor ledger, the tool results that cost the most to read, and sets or raises the expensive-tier budget; also installs the zero-cost status line and switches between off, enforce, observe and explore modes. Use when asked how muc

  • consult

    Asks the expensive-tier architect one structured question — a decision between named options, a decomposition, a diagnosis a cheaper model failed at twice — with the brief the supervisor hook requires (Question, Context, Definition of done). Use from a session on a cheaper model when a decision is a

  • decompose

    Turns a large body of work — an unmerged branch, a messy test suite, a multi-file refactor — into independent slices with a spec each, an order, and a worktree plan, so cheap workers can implement them in parallel. Use when a task touches more than about five files, when a branch is too big to revie

  • delegate

    Hands one slice of work to a cheap worker the right way — a written spec with the files, the definition of done and the tests to run, the matching supervisor agent, and the check of the returned evidence. Use whenever implementation, test writing, or a look-up is about to happen in an expensive-mode

  • explore

    Puts the session into supervisor explore mode for a loosely defined question — workers pinned and forks denied, no report contracts, the budget a checkpoint — frames the question in three lines, and routes the outcome to ship, spike or drop. Use when asked to explore, look into, figure out, poke at

  • off

    Disarms the plugin for this session; use when asked to turn it off, stand down, or stop enforcing.

  • on

    Arms the plugin for this session in enforce mode without a brief, optionally with this session's own budget (a number or a profile name) and a spend reset; use when asked to turn it on, arm it, enable it, give the session a budget, raise the budget, or reset spend.

  • reflect

    Reviews the corrections, standing rules and preferences the supervisor queued from earlier prompts and writes the accepted ones to auto memory with the reason — the user accepts, edits or drops each line; nothing is written to memory or rules without that review. Use at the end of a session, when as

  • start

    The one entry point for a task in an expensive-model session — interviews in one batched round, writes the brief, triages into tiers, decomposes when the work is large, picks a budget profile, and presents one plan card for go / adjust / explore. Use when a task arrives as a sentence, when asked to

  • triage

    Sorts a task into the model tier it deserves before any work starts — the expensive tier for decisions and decompositions, Opus for hard slices, Sonnet for execution, Haiku for look-ups — and writes the delegation table. Use at the start of any non-trivial task in an expensive-model session, when as

Subagents (6)

  • architect

    The expensive-tier consultant — answers one structured design question (a decision between named options, a decomposition of a large change, a diagnosis a cheaper model failed at twice) on Fable at high effort, read-only, and returns a decision record. Spawn only through /supervisor:consult with a b

  • implementer

    Implements a written spec on Sonnet at medium effort — code plus tests, run, with the command and output pasted. Use when the design is decided, the spec names the files and the definition of done, and the work is execution. Stops and reports rather than guessing when the spec is ambiguous. Not for

  • reviewer

    Reviews a diff against its spec on Opus at medium effort and returns findings as JSON, each with a concrete failure scenario. Use after an implementer reports DONE and before the conductor accepts the slice; use with a different model than the implementer. Does not fix anything, does not comment on

  • scout

    Read-only reconnaissance on the cheapest model — locates code, config, tests and call sites, and returns path:line references with short excerpts. Use before any implementation or decision that needs to know where things are, instead of reading files into the conductor's context. Does not review, ju

  • senior-implementer

    Implements a written spec on Opus at high effort — for slices the triage marked hard-but-not-architectural (tricky refactors, concurrency, subtle test fixtures, migrations) where Sonnet has failed or is expected to. Same strict contract as implementer — spec is the boundary, evidence pasted, no git.

  • worker

    A drop-in for a bare general-purpose spawn — Sonnet at medium effort with every tool and no report contract. The supervisor hook routes an unpinned general-purpose spawn here so a cheap worker never inherits the session's effort. Use for research, docs, prose review, or any task that needs no eviden

Hooks (6)

  • SessionStart

    python3 "${CLAUDE_PLUGIN_ROOT}/bin/supervisor.py" session-start --state-dir "${CLAUDE_PLUGIN_DATA}"

  • UserPromptSubmit

    python3 "${CLAUDE_PLUGIN_ROOT}/bin/supervisor.py" user-prompt --state-dir "${CLAUDE_PLUGIN_DATA}"

  • PreToolUse

    python3 "${CLAUDE_PLUGIN_ROOT}/bin/supervisor.py" pre-tool-use --state-dir "${CLAUDE_PLUGIN_DATA}"

  • PostToolUse

    python3 "${CLAUDE_PLUGIN_ROOT}/bin/supervisor.py" post-tool-use --state-dir "${CLAUDE_PLUGIN_DATA}"

  • SubagentStop

    python3 "${CLAUDE_PLUGIN_ROOT}/bin/supervisor.py" subagent-stop --state-dir "${CLAUDE_PLUGIN_DATA}"

  • SessionEnd

    python3 "${CLAUDE_PLUGIN_ROOT}/bin/supervisor.py" session-end --state-dir "${CLAUDE_PLUGIN_DATA}"

View source on GitHub

More from konyklabs/claude-plugins

All 9 plugins →

Python test engineering skills — pytest project structure, Playwright API and browser tests, SQLAlchemy test fixtures — and the untangling workflow for large unmerged test suites.

5 skills · 1 agent

0

Production-readiness and security-scanning checks: a categorized deterministic scan, summarized external scanners, and an auditor agent, ordered by observed impact.

3 skills · 2 agents

0

GitHub Actions workflows authored, preflighted and reviewed the way the infrastructure work is: a deterministic preflight (actionlint and zizmor on PATH reduced to counts, structural checks for the pinning, permissions, pull_request_target and injection traps), a workflow review lens with an Opus reviewer agent, and docs-first authoring references for reusable workflows, permissions and OIDC, supply chain and runtime shape.

3 skills · 1 agent

0

signoff

claude

Reconciles an application's end-to-end coverage against what it does: explore, mine, tile, fill, report, human-readable test cases.

4 skills

0

terrain

claude

OpenTofu on AWS, authored and reviewed the way the Python work is: a deterministic preflight (fmt, validate, the linters on PATH reduced to counts, structural checks for the Fargate and Lambda traps no linter catches, a plan summary that fails closed on destroy), a Terraform review lens with an Opus reviewer agent, and docs-first authoring references for ECS Fargate, Lambda, IAM, networking and state.

3 skills · 1 agent

0

Vitest and TypeScript project skills, the py-testing shape for the TypeScript side: config that shares vite.config.ts, mocking and fake-timer rules, test.extend fixtures, typecheck mode, CI reporters and sharding; tsconfig for Node with nodenext, the TypeScript 6 and 7 default changes, strictness adoption, project references, type stripping; and a Sonnet test-implementer with both skills preloaded.

2 skills · 1 agent

0

Similar plugins

caveman

claude

Talk like caveman. Cut filler. Keep technical details explicit.

22 skills · 1 command · 3 agents · hooks

JuliusBrussee/caveman 119,462

claude-mem

claudecodex

Persistent memory for Claude Code: the next session starts where the last one left off.

29 skills · 1 MCP server · hooks

thedotmack/claude-mem 116,618

Taste Skill

claudecodex

tasteskill 2.5.1: professional company websites at the figma.com and x.ai level, with one section per screen, real photographs, real product UI, effects kept for tech and gaming pages, and a real-browser checker and finish gate.

1 skill

Leonxlnx/taste-skill 114,391

career-ops

claudecodex

Open-source AI job search that runs locally in your AI coding CLI: scan job portals, evaluate listings into a structured A-H report with a 1-5 score, tailor your CV, track applications

5 skills · hooks

career-ops-hq/career-ops 82,253

mem0

claudecodex

Cross-session memory and token savings for coding agents.

6 skills · 1 agent · 1 MCP server · hooks

mem0ai/mem0 68,893

pake

claudecodex

Package any website or local web build into a lightweight desktop app with Pake CLI.

1 skill

tw93/Pake 63,463

Maintain supervisor?

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 plugins9 plugins