Software infrastructure lab

The map and runtime for autonomous software delivery.

A deterministic map of every repository, and a runtime that takes an issue to a merged, deployed change.

● Cartographer — the map● Healer — the evidence● Pipeline — spec a PR

code graph

0 symbols · 0 edges
auth::gatewayevent::busarchivesstorage::engine

map

deterministic

index

0 LLM tokens

runtime

AWS-native

01

Wrong memory is worse than none

Undetectable is null, not a guess. Every narrated claim carries citations.

02

No capability without its consumer

If nothing calls it, it is not finished. Nothing merges to be given a use later.

03

A gate is proven by failing

If you cannot show the red, you do not have the gate. The core is proprietary; a tool that stands on its own gets published.

Works with the agents your team already uses

GitHubGitHubClaudeClaudeGoogle GeminiGeminiCursorCursorGitHub CopilotCopilotWindsurfWindsurfZed IndustriesZedModel Context ProtocolMCPGitHubGitHubClaudeClaudeGoogle GeminiGeminiCursorCursorGitHub CopilotCopilotWindsurfWindsurfZed IndustriesZedModel Context ProtocolMCP

What you can use today

Quartermaster is out, and it is free.

It came out of running out of quota halfway through something. The bar on screen said 8 %; the one actually braking was at 75 %, in a cache Claude Code had already written to disk, and no tool showed it because they all read the two bars that have a name. This one reads all of them, across every profile on the machine. Public, MIT, and installable right now — the three below are not.

The toolPublicv0.1.20

Quartermaster

How much quota is left, across every account.

Reads every Claude Code profile on the machine — not just the default one — plus Codex and whatever providers opencode keeps, and says how much is left in each. No network and no credentials: the number comes from the cache the CLI already wrote to disk.

  • Every profile, not the default one

    Every other tool reads one credential and stops: 1 of N. On the machine this was written on, N was 4 and the 1 was an empty directory.

  • The bar that brakes, not the one on screen

    Inside cachedUsageUtilization there are bars with no key of their own. Reading only five_hour and seven_day showed 8 %; the one actually braking was at 75 %.

  • Four surfaces, one rule

    GNOME, the macOS menu bar, the Windows tray and the browser draw the same thing because none of them decides anything: qm --json emits the resolved answer.

Mission

Nobody should discover their quota is gone by hitting the wall.

quartermaster

›qm

work ███████░░░ 75% stops in 2 h 10 min

personal █░░░░░░░░░ 8%

codex ████░░░░░░ 41%

4 accounts read from disk — no network

  • Never refreshes a token — it reads credentials, it does not write them.
  • Nothing leaves the machine but the quota poll itself, with a 60 s floor.
  • Local transcripts are the floor: there is always a number, even with everything expired.

Products

Three products. One workflow.

Cartographer charts the ground so agents know where they are. Healer turns real incidents into cited root causes. Pipeline takes an approved spec to a PR that verified itself. One artifact — the spec — and everything else produces or consumes it. These three are proprietary and still in the workshop.

The mapPrivate

Cartographer

Charts a repo before anyone sails it.

Scans a repository into a symbolic, machine-queryable context artifact — so agents and humans never re-discover a repo with ls, grep, and cat.

  • Symbolic context artifact

    Nine closed namespaces, closed enums, three-valued semantics. One vocabulary everywhere.

  • Code-symbol index — zero LLM tokens

    tree-sitter over ~25 languages into SQLite. Impact, refs, routes, env, tests — pure SQL, milliseconds.

  • Shared graph, versioned like git

    The graph rides on git — no server to keep running for the team to share the same fresh structure.

Mission

Any agent productive in any repo within one file read.

cartographer

›cartographer scan .

charted: cartographer -> .cartographer/repo-context.json

›cartographer impact buildRuntimePreview .

refs agents/runtime/src/index.ts 10,129,144

impact radius resolved — zero model tokens

  • Deterministic — same tree, byte-identical output.
  • Never invents facts — undetectable means null.
  • No secrets — env names only.
The evidencePrivate

Healer

Real incidents into cited root causes.

Ingests incidents from Sentry and GitHub Actions, normalizes them to one type, and collapses them by signature: N events become M candidate causes before a model is ever called.

  • The funnel, before the model

    Ingest, normalize and cluster are deterministic and model-free. 500 events collapsing into 12 causes are 12 calls, not 500.

  • Zero vendor names outside the adapter

    The core does not know which tracker brought the incident. CI gate, proven red before it was trusted.

  • Nothing self-accepts

    A hypothesis reaches 'accepted' only by human decision. The log is append-only and the first verdict wins.

Mission

Nobody should have to chase a traceback by hand to learn what broke and why.

healer

›healer proponer

141 incidents -> 86 clusters (ratio 1.64:1)

case file: 51 real citations, zero tokens

hypothesis proposed — awaiting human review

  • Deterministic clustering — same input, same output order.
  • Every cause arrives with its citations; undetectable is null.
  • Acceptance rate stays null until a real person decides.
The closePrivate

Pipeline

Approved spec, free executor, verified PR.

Takes a spec a human approved and carries it to a PR whose own verify: ran and passed. The executor is external and there is exactly one — we do not build an agent.

  • One single artifact

    The spec. Cartographer and healer produce it; the executor consumes it. No interface per pair.

  • Real per-run isolation

    One ephemeral Fargate task per run, with a hard spend cap and state that persists outside the stack.

  • Constrain the what, free the how

    The contract is the outcome, not the procedure. The spec says what must be true and how it is verified.

Mission

A change reaches main from an approved spec, verified by its own commands.

pipeline

›pipeline run spec-142

fargate: ephemeral task — isolation per run

verify: pnpm test && pnpm lint

PR opened — verify green

  • No change counts without its verify: green.
  • No capability merges without its consumer.
  • State persists outside the stack before the stack exists.

The numbers

`null` is an answer. Zero would be a lie.

Every milestone in every repo commits its number to a file, with the date and the command that produced it. This section reads those files — nothing here is typed by hand, and CI fails if a number on this page stops matching its source.

50%

PRs merged without a person rewriting the change

over 2 reviewed · small sample — one run moves this number.

pipeline/numeros/north-star.json

51%

Root-cause hypotheses a person accepted

healer/numeros/h7.json

14186−39%

Incidents collapsed into candidate causes, before any model call

healer/numeros/h1.json

203,43718,889−91%

Tokens to answer “where is X used”, on this very repo

cartographer/numeros/c9.json

95%

Clarify specs passing all four automated checks (prompt v2)

pipeline/numeros/bench-clarify-v2.json

0

Failing verifications in the pipeline’s first real PR

pipeline/numeros/p7.json

`null` is not 0%. “None got merged” and “nobody looked” are two different claims, and conflating them is how a baseline gets lost — it happened to us twice. A number shows up here when someone measured it, not when it looks good.

How it works

Issue in. Merged, deployed change out.

Cartographer charts the ground, healer says what broke and why with citations, and pipeline carries the approved spec to a reviewed change.

01

Incident

Something breaks in production. Healer ingests it and collapses it with everything sharing its signature.

02

Map

Cartographer charts the repo into a deterministic context graph — zero LLM tokens.

03

Spec

A cited cause plus repo context compile into a spec. Nothing moves past here without a person accepting it.

04

PR

The executor takes the spec in an isolated task and opens a PR whose own verify: ran and passed.

About

Precise systems, not demos.

Legios is an independent, founder-led organization. The core — cartographer, healer, pipeline — is proprietary, and there is no open-core boundary to maintain because there is no open product. Tools are another matter: quartermaster is public and MIT because it stands on its own and gives nothing of the core away.

Ground truth

Facts before guesses.

Deterministic maps and zero-LLM indexes give agents ground truth about a repository — byte-identical on every run.

25+ languages0 LLM tokens1 file read

State outlives the run

A dead task loses no data.

Every run is one ephemeral, isolated task with a hard spend cap. What it produces is written outside the stack before the stack exists — a crashed run leaves a record, not a gap.

One artifact

The spec, and nothing else.

Cartographer and healer produce it; the executor consumes it. No interface per pair, and the core stays proprietary.

Operator-grade

Explicit contracts, inspectable behavior.

Small composable tools with typed boundaries — you can see exactly what ran and why.

$ healer proponer

step plan  ok  1.2s

step code  ok  8.4s

step tests ok  passed

Owners

Founder-led by systems builders.

Legios is led by Sol Soletti and Valentin Torassa Colombero, whose public work points at the same center of gravity as the organization: developer tools, backend systems, infrastructure, and automation.

Sol Soletti GitHub avatar

Co-founder

Sol Soletti

Rosario, Argentina

Shapes Legios around practical developer tooling and clear software systems. Public work across TypeScript and Go.

Focus: Developer tools, backend systems, Go, TypeScript, product.

Co-founderTypeScriptGoDev toolingProduct
Valentin Torassa Colombero GitHub avatar

Co-founder

Valentin Torassa Colombero

Argentina

Public work spanning Python, TypeScript, Go, Linux, and self-hosted tooling. Strong bias toward systems and automation.

Focus: Infrastructure, automation, Linux, containers, security tooling.

Co-founderInfrastructureAutomationLinuxAWS

Technical stack

Confirmed technologies across the repositories.

Observed in the legiosai repositories — manifests, workflows, the Terraform under pipeline/infra, and project documentation.

TypeScriptNode.js 22Next.jsReacttree-sitternode:sqliteZodAJV · JSON SchemaMCPExpressBiomeVitestDockerGitHub Actions
Amazon Web ServicesAmazon Web Services

Built on AWS

Provisioned as code with Terraform

TerraformECS FargateAmazon ECRAWS Secrets ManagerAmazon S3Amazon CloudFrontAWS Budgets