A Claude Code plugin that turns every session into a supervisor of a small engineering team. The supervisor builds small work itself and wakes specialist sections (frontend, backend, database, testing, security, devops, uiux, docs, ai-llm, mobile, plus two read-only advisors) for bigger work. Every piece of code passes an Opus reviewer gate before it is handed back, section reports are saved in the project, and a live Team Board shows who is working.

Measured over 17 benchmark rounds (three tasks of rising size, two runs each, graded blind by 12 Opus judges per round). In the final round:

It costs about 3.6 times as much as plain Claude Code, for cleaner, better-tested and more robust code.

Limits: every setup, plain Claude Code included, passed all the hidden tests, so the gains are in blind-judged quality, not in whether the code works. Each cell is 2 runs (about ±0.3 judge noise), and the team was tuned over 17 rounds on these same three tasks, so expect smaller gains on real projects.

Method: each round builds the same three tasks (a small library, a command-line tool and a web app) twice per setup. Hidden tests check correctness, then the solutions are anonymised and graded blind by Opus judges on correctness, code quality, tests, robustness and docs. The tasks, hidden tests, harness and scorer are in bench/; a round runs real paid Claude sessions. Full results and every round's changes: docs/benchmark-2026-10-24.md (final summary at the end).

• Claude Code with plugin support. The Team Board strip, /team and /chats use the Mods API (hooks/register.tsx).
• Access to the Claude 5 models: sections and the reviewer run on Opus 5.5, the coder helper on Sonnet 5.5, the scout helper on Haiku 5.5.
• Node.js (tested on v26) for the hooks and tests.
• Recommended skills, preloaded by some sections: frontend-design (frontend), tdd-workflow (testing), security-review (security), claude-api (ai-llm, ai-engineer), web-design-guidelines (uiux). Other skills a job names are loaded only if they are installed.

In Claude Code 2.1.275 or later, one command installs it from this repo:

/plugin install section-team --marketplace afel6/section-team

On older versions, run /plugin marketplace add afel6/section-team, then /plugin install section-team@section-team.

To work on the plugin itself, clone it instead:

git clone https://github.com/afel6/section-team.git

• Try it once: claude --plugin-dir <path-to-section-team>.
• Keep it on: in ~/.claude/settings.json, set env.CLAUDE_CODE_PLUGIN_DIRS to the clone's path. New sessions pick it up.
• Off: set "section-team@inline": false under enabledPlugins, or remove the env entry. Emergency: claude --safe-mode.

The section agents are generated: edit the roster in scripts/gen-agents.mjs and run node scripts/gen-agents.mjs. Skills you give a section with node scripts/assign-skill.mjs <skill> <section>... are kept in scripts/roster-extra.json.

• Just talk to the supervisor. Small tasks (about 50 lines, one or two small functions) it builds, tests and gates itself; a feature with its own behaviour to test (a CLI, an endpoint, a module with storage or I/O) is never small. Single-area tasks (up to about 600 lines): it builds the code itself, wakes only the testing section for the tests, and gates its code with the Opus reviewer. Multi-area work wakes the builder sections plus testing. 3+ builder sections (testing, docs and advisors not counted) → it shows a plan first.
• Sections message each other directly by agent id and share contracts as notes in .team/notes/. If an id is missing or a message fails, the supervisor forwards the request with the asker's id, so the owner replies to the asker directly.
• The supervisor names up to 4 skills per job (Skills for this job:), picked from what is installed and fits that section and task.
• Reports: .team/reports/<date>/ in each project. Every REPORT has a Quality: line, the reviewer's grade (see Sections) or n/a (no code).
• Notes: .team/notes/ in each project, INDEX.md plus one file per note. Notes and reports both live under .team/, kept out of git via .git/info/exclude; to keep them in git, create .team/keep-in-git and delete the .team/ line from .git/info/exclude.

• Nothing opens by itself. A one-line team strip above the prompt shows who is awake, the team's tokens and helpers, and how many sections sleep; its Board button opens the board.
• /team opens the Team Board. Each awake section has a box with its live token use, how many helpers it started by model, and its reviewer grade once it is done. Sections nobody woke are folded into one "N asleep" line.
• /chats opens a dashboard with one card per open chat: needs you, working or idle (a chat quiet for 30 minutes is idle; one that ended or has been quiet for a day drops off). It reads one status file per session in ~/.claude/section-team/sessions/. A status file holds the project folder, the first line of your latest request (cut to 80 characters), the awake sections and the token and helper counts. Status files older than 7 days are deleted at session start.

Builders: frontend, backend, database, testing, security, devops, uiux, docs, ai-llm, mobile. Advisors (read-only): ai-engineer, data-scientist. Section heads work like solo builders. Each reads .team/notes/INDEX.md first, plans its part, then hands independent tasks out in parallel to cheap helpers (not sections, not on the board as cards): section-team:scout (Haiku, read-only: reads and searches, returns file:line facts) and section-team:coder (Sonnet, routine code and tests from the head's spec). Both helpers can load skills with the Skill tool. Heads run helpers in the foreground (run_in_background: false; several in one message still run in parallel), because a head that waits on a background helper ends its turn and is made to hand back unfinished. The plugin refuses a section's Agent call that asks for the background (PreToolUse hook). Where the Agent tool has no such option (the desktop app), helpers run in the background and the head keeps working until their results are in. The head keeps design, tricky bugs and integration for itself. Advisors may use the scout but not the coder.

• Notes: a head writes a note in .team/notes/ for anything other sections depend on (an API, a schema, design tokens, a file layout) and adds one line to INDEX.md (<file> — <owner section> — <one-line summary> — <date>). Only the owner edits its note; shared decisions are appended to .team/notes/decisions.md.
• Skills: each section preloads at most one core skill; every other skill loads on demand, the job's Skills for this job: list first. If no installed skill fits, a section searches the open directory read-only (skills.sh/api/search) and names a Suggested skill: in its REPORT; it never installs. The supervisor lists the suggestion in its summary and asks you; on yes it installs it through section-team:add-skill (a safety review, then the install) for that section. To let it install without asking whenever the review finds no red flag, set env.SECTION_TEAM_AUTO_INSTALL_SKILLS to 1 in ~/.claude/settings.json; a red flag, or an existing skill of the same name, still stops the install and you are told.
• Clean-code gate: once the helpers' code is in, a builder sends its changed files to section-team:reviewer (Opus, never edits). Pass: overall at least 8 out of 10 and no item under 6. On a fail the fix list goes to a coder and the reviewer grades again, at most 2 fix rounds; if it still fails the section ends Status: partial with the grade. Jobs that change no code skip the gate. The plugin enforces it at handback: a builder whose REPORT lists changed code but no grade cannot hand it back (PreToolUse hook on SubagentHandback). It must run the reviewer first, or hand back honestly as Status: partial or blocked. The supervisor runs the same gate on code it wrote itself.

Section heads skip your global CLAUDE.md (it can cost many thousands of tokens per agent) but read the project's own rule files (CLAUDE.md, .claude/CLAUDE.md, CLAUDE.local.md, AGENTS.md, .claude/rules/*.md), and carry a short copy of the house rules.

• npm test — hook logic (node:test)
• npm run test:board — Team Board (claude plugin test)
• npm run validate — plugin validation
• node scripts/gen-agents.mjs --check — agent files match the roster
• node --test bench/*.test.mjs — benchmark scorer. bench/score.mjs records each run's per-agent model split from the subagent transcripts; bench/aggregate.mjs reports the Opus share of cost (opusSharePct).
• Design: docs/specs/2026-10-24-solo-builder-sections-design.md (v3), docs/specs/2026-10-12-claude-team-design.md · Plan: docs/plans/2026-10-24-section-team-v3-plan.md · Phase 0: docs/phase0-results.md