Crate magi 

magi — a blind multi-agent implementation competition.

One task, several agents, and a graph that decides which implementation survives without a human reading the diff:

1. implement — N agents solve the same task in isolated git worktrees, unaware of each other.
2. judge — M judges rank the candidates blind: labels A/B/C, a per-judge presentation order, attribution trailers stripped at write time by a commit-msg hook and again at presentation time.
3. deliberate — if the first choices disagree, the judges argue, with magi as the facilitator. A facilitator made of code cannot leak an author.
4. vote — final votes are collected one-to-one and privately, so nobody can drift toward a visible majority.
5. review — the winner enters a bounded review + real-machine verification loop with a fixer that may reject a finding with an argument.
6. gate — configured commands must pass before anything merges.

Everything is recorded, so the by-products are real numbers: per-agent win rates, per-reviewer precision, and how often execution caught what static review missed. See stats.

Modules

advise

Data and pure logic for the design-deliberation stage.

agent

Driving agent CLIs.

already

Is a branch’s change already in the base under a different commit id?

ask

Questions: what an agent does when the next decision is the owner’s.

blind

Blindness: label assignment, attribution stripping, and leak detection.

blockers

Dependency inventory for blocked tasks.

bump

Release version bumps, opened automatically once a merge lands.

cache

Durable ownership of the shared Cargo build cache (CARGO_TARGET_DIR).

chat_control

Direct task and run control for the standing chat: magi task control ....

chat_notify

Telling a chat how a task it filed ended.

chat_watch

Durable, supervised jobs started by a standing chat. Results use the same draft/turn delivery as task completion notices; jobs never run twice.

choice_effect

One description of each executable question choice, shared by cards and briefs.

clean

The disk janitor: finished runs get their worktrees folded, worktrees whose run record is already gone get reclaimed too, and the shared build cache is pruned to its cap.

conduct

The conductor: a single agent seat that arranges the queue.

config

Run configuration: the agent roster, the shape of the graph, and the blindness / verification policy.

consult

Handing an open question to the chat conversation its task came from.

daemon

The unattended loop: take the next task, run the graph, record what happened, take the next one.

deputy

Deputies: the seat that waits on a conductor’s question, so the conductor never has to - and on land’s merge approval, whose free-text reply used to reach nobody either.

direct

Runs started by hand are tasks.

disk

Disk accounting: how much magi’s own directories occupy, how much space is left on the volume they live on, and how the shared build cache is pruned.

dupes

Duplicate-work detection at the moment work is filed.

fixer

The [roles] fixer fallback chain, shared by every node that asks a fixer.

followup

Follow-up tasks for the findings a merged run left open.

git

Git plumbing.

github_text

Posting gate shared by GitHub titles, descriptions and comments.

graft

The graft phase’s decisions: whether it runs and what it is told.

graft_report

What a graft round was asked to port and what the fixer says it did, for the report.

graph

The competition graph.

handover

Taking a branch over from an earlier attempt at the same task.

land

Landing the winner: watch the pull request, fix what it complains about, and merge it.

lang

One reading of [graph] language, shared by everything that has a translated wording table.

md

GFM markdown, parsed into a serializable node tree instead of HTML.

merge_decision

Persisted owner decisions at the merge gate. Tokens are independent of UI prose.

notices

Notices: what needs the operator’s attention but is not a question.

persona

Personas for the standing chat: a voice, never a behaviour.

persona_icon

Persona avatars: display only, never part of the briefing.

pr_outcome

What a pull request description may say about commits and verification.

pr_recovery

Explicit recovery of a task’s public implementation, independent of historical candidates.

proc

Spawning child processes without putting a window on the operator’s screen.

prompt

Prompt construction.

question_delivery

Admission of operator text: a saved deputy is history, not a listener. Called under the question update lock as well as when rendering its card.

queue

The task queue: what magi should do next, and who asked for it.

rebase

Rebasing the winning branch onto a moved base, with a fixer for conflicts.

reconcile

Deciding what a local branch and its remote twin mean when both moved.

recovery

Carrying an earlier attempt’s evidence into the run that takes its branch over.

refs

Structured implementation source selection. Instruction text is never input.

release_local

Releasing without GitHub Actions ([release] mode = "local").

release_watch

Watches every open release-bump pull request until it lands.

report

Terminal rendering.

report_view

A structured view of the run report, for the web UI.

repos

Local repository discovery: magi repos and GET /api/repos.

retention

Bounded reclamation of retained attempts and Cargo outputs. Source recovery and build eviction have separate authority: evicting a target never retires a run.

review_repair

Durable requests for bounded repair of findings at the merge gate. Old review evidence stays immutable; no merge permission is carried forward.

rng

A seeded PRNG, so that label assignment and session ids are reproducible from a run id.

run

Run state: what happened, where it is stored, and how a run is resumed.

scrub

Last-line scrub for text that lands on GitHub.

settings

The machine-config settings screen: which agent each role resolves to, and a safe way to change those assignments.

stats

Aggregate statistics over every recorded run.

talk

The standing conversation: a place to think out loud with an agent between tasks, reachable from a phone.

triage

Held-task triage: walking held tasks so a hold left by an accident does not sit unread forever next to one a human placed on purpose.

tui

The observation deck.

updater

Self-update, via kaishin.

verdict

Structured answers extracted from free-form agent output.

waiter

The waiter: something that keeps waiting on a question after the process that asked it is gone.

web

The web UI: magi’s queue and run history, readable from a phone.

worktree_setup

[worktree] setup: what a fresh worktree needs beyond its tracked files.