01
Claude Code
Plugin · real subagents
Five agents with role-scoped tools (only the implementer holds Edit) plus six commands.
/troopers:mission <task>/troopers:brainstorm <task>Five troopers · one mission command · Boyan Balev
Command the work.
Never touch it.
Five specialist subagents under one engineering constitution, run by a mission command that only briefs, reads handoffs, and decides. Every trooper deploys in fresh context, gathers its own evidence, and owns exactly one phase.
Five troopers · one viewportthe fleet moves in unison
The doctrine
Orchestration without contamination.
The commander shares briefs, not context.The constitution remains the law. Repository reality remains the evidence. The original task remains the objective. The main thread carries five thirty-line handoffs, and nothing else.
01
The commander never does the work: no repository reading, no analysis, no code.
02
Troopers deploy strictly one at a time, each as a fresh subagent.
03
Briefs point at artifacts and file:line evidence; they never paste transcripts or file contents.
04
A trooper is never told it is stage N, what others said verbatim, or which conclusion is hoped for.
05
Only the implementer modifies production code.
06
The user is interrupted only for genuine blockers; everything else is decided and logged as a locked decision.
07
Failed verification loops back as bounded defects, twice at most. When the task is solved safely, stop.
Five troopers
One responsibility each. One handoff forward.
The mission loop
Forty lines down. Thirty lines back.
Every trooper writes its full output to a numbered artifact in .troopers/<date>-<slug>/ and replies with only its handoff. The next trooper reads the artifacts it needs directly from disk, so the commander's context never grows.
01 / discover
Problem → options
The brainstormer gets the task verbatim and nothing else, unanchored by anyone's opinion. It returns the framed problem, repository evidence, 2–4 materially distinct approaches, and one recommendation.
artifact
01-brainstorm.md
01-brainstorm.md
02 / commit
Direction → bounded plan
The planner receives the recommendation, the rejected alternatives with reasons, and open risks, then checks all of it against the code before writing scope, non-goals, file-by-file steps, and a definition of done.
artifact
02-plan.md
02-plan.md
03 / execute
Plan → minimal diff
The implementer gets scope and completion criteria, reads the full plan from disk itself, and produces the smallest repository-native diff: tests included, deviations recorded instead of absorbed.
artifact
03-implement.md
03-implement.md
04 / prove
Claims → evidence
The verifier judges the actual diff against the definition of done, re-runs the checks itself, and returns PASS, PASS WITH RISKS, or FAIL, fixing nothing. FAIL loops back as a defect-only brief, twice at most.
artifact
04-verify.md
04-verify.md
05 / preserve
Correctness → durability
The auditor reads the whole chain and the final diff, and rules on erosion: CLEAN, ACCEPTABLE DEBT, or BLOCK. A clean result ends the mission with a twenty-line report.
artifact
05-audit.md
05-audit.md
Task: one-paragraph restatement of the user's task Mission dir: .troopers/2026-09-01-rate-limit-login/ (write your full output to 03-implement.md) Read first: 02-plan.md · src/auth/login.ts:41 (pointers, never pasted content) Locked decisions: token bucket over sliding window; reuse existing redis client Constraints / non-goals: no new dependencies; public API unchanged Definition of done: plan §7 criteria; regression test for burst traffic Open risks: clock skew across instances (carried from the 01-brainstorm handoff) Reply with only your ## Handoff section.
handoff ≤ 30 lines ↗
Mission commandbriefs · reads handoffs · decides
↘ metaprompt ≤ 40 lines
Brainstormer
→
Planner
→
Implementer
→
Verifier
→
Auditor
artifacts on disk: 01-brainstorm.md → 02-plan.md → 03-implement.md → 04-verify.md → 05-audit.md · fix loops append -fix1, -2
Authority matrix
Clear boundaries prevent role drift.
| Capability | Brainstorm | Plan | Implement | Verify | Audit |
|---|---|---|---|---|---|
| Explore alternatives | Primary | Refine | No | No | No |
| Set implementation scope | Recommend | Primary | Escalate deviations | No | No |
| Edit production code | No | No | Primary | No | No |
| Run/inspect verification | Early risks | Define | Run own checks | Primary | Review evidence |
| Block for correctness | No | No | No | Yes | Yes |
| Block for erosion | No | No | No | Flag | Primary |
Core invariant
No trooper owns more truth than the task, the repository, and the constitution.
Documents
Eight files. The exact text every trooper carries.
Rendered from core/, the single source of truth. Each deployed trooper is self-contained: its role, the shared protocol, and the constitution.
Deployment
Same troopers, three theaters.
02
Codex
Prompts · codex exec
The mission runs each trooper as a fresh-context subprocess; briefs and artifacts carry the handoffs.
/troopers-mission <task>/troopers-plan <task>03
pi
Skill · subagent tool
One skill; the first argument picks a lone trooper, anything else launches the full mission.
/troopers <task>/troopers verify <task>git clone https://github.com/balevdev/troopers && cd troopers && ./install.sh --all
github.com/balevdev/troopers · edit core/ · every theater stays in sync
Mission command
Boyan Balev.
Engineer behind Troopers. Writes the constitution, sets the doctrine, commands the mission, and never touches the diff. The system runs the same way he does: smallest safe change, evidence over confidence, stop when it's done.
balevdev · one constitution · three theaters
5→1
Five specialists. One commander. One verified outcome.
The point of multiple agents is not more activity. It is separated judgement under command: explore broadly, plan narrowly, implement minimally, verify mechanically, audit for erosion, then stop.