Portable fleet alpha

One repository.
A coordinated AI engineering team.

TORCH studies your code and project specification, proposes durable areas of responsibility, and gives each specialist its own branch, worktree, context, and route back to main.

Design a fleet Read the source

Request CR-184

A
Owner requestChange terrain streaming without weakening frame checks.
S
Session ManagerRoutes ownership and establishes acceptance.
T
Terrain specialistWorks in its persistent domain context.
QA peer
✓
Exact-commit checksReceipts bind evidence to the measured SHA.
M
Controlled mainAuthorized fast-forward after convergence.

Ready at owner request.

Git required

GitHub optional

Provider independent

Local first

Human approved

The project decides the team.

TORCH does not install a catalog of generic agents. It reads the repository and the specifications you select, then produces a pending organization for you to split, merge, rename, or reject.

Input

Code and intent

Architecture, tests, operations, scarce resources, and external Markdown specifications become evidence—not instructions to execute.

Design

Explicit boundaries

Every domain receives scope, exclusions, owned paths, neighbors, checks, runtime policy, and collision rules.

Operate

Persistent specialists

Stable identities survive runtime sessions. When a durable new boundary emerges, the manager can propose another specialist—but only owner approval wires it into the Fleet.

Keep useful context hot.

A specialist can retain the working set for one coherent area instead of repeatedly loading unrelated project history. TORCH measures cache reads, cache creation, compactions, resumed prompt size, cost, and verified outcomes before calling that an efficiency win.

Context savings are a measured hypothesis, not a marketing claim.

Proof has a commit.

TORCH distinguishes implementation from qualification. Green output without the right scenario, authority, and SHA is not proof.

CapabilityCurrent evidenceState

Portable kernel and reversalScenario testedProven locally

Durable control plane and backlogSQLite + tracked tasksProven locally

Checks, resources, integrationExact-SHA lifecycleProven locally

Claude and Codex coordinationVirtualized adaptersLive gate open

Self-managed successorAtomic candidate mechanicsStable boot open

COMBATRIG migration33 areas, 570 tasks inventoriedBoundaries open

Start with the work you actually have.

Point TORCH at a Git repository, a product specification, or both. Bootstrap gives your Session Architect AI a bounded evidence brief; installation waits for your reviewed proposal.

node bin/torch.mjs bootstrap \
  --repo /path/to/project \
  --spec /path/to/product-spec.md \
  --output fleet-design-brief.json \
  --json

One repository.A coordinated AI engineering team.