Skip to main content

Building the CPU for agents

Organized chaos. Verified work.

RepoToire gives agents code facts to work from. We’re building toward a bigger goal: carrying your intent through to verified work, with less guesswork and coordination.

Your goal → RepoToire → verified work

Architecture vision · current execution is serial

01 · Break down the task

One goal. A lot of moving parts.

People bring ideas. Agents bring capacity. RepoToire’s goal is to turn that scattered effort into a clear plan for the work you want done.

Your intent is the starting point: the software you want to build, improve, or repair. Three people and five example agents bring competing requests. Claude takes Part A, Codex Part B, Gemini Part C, Cursor Part D, and Copilot Part E. These illustrate the vision; the private beta is for Codex and Claude Code.

01 / 05

Slide 1 of 5: One goal. A lot of moving parts. People bring ideas. Agents bring capacity. RepoToire’s goal is to turn that scattered effort into a clear plan for the work you want done.

Architecture & tool evidence

The five slides describe the long-term workflow, not the five ticks inside an instruction. Out-of-order execution remains a target. The opcode view is schematic; it is not a compiled program or a recorded execution.

The named agents are examples of this vision. The private beta is for Codex and Claude Code. Completion times are illustrative; they do not compare agent performance.

Code context available today

In a separate recorded example, we chose Hero as the focus. RepoToire found its source location and a use in AnimatedSections. It did not automatically map the rendered headline to code.

This recording predates this slideshow. It is not a live command or a test result. It helps an agent decide what to inspect; it does not prove a change is safe or complete.

Analysis limits: warning: impact authority is unknown (19 parser diagnostic(s)): measured impact recall report does not match this binary's Impact implementation

Context

repotoire context . --focus Hero --budget 1200
scope: {"surface":"context","language":"type_script","files_covered":205,"files_total":205}

# repotoire (TypeScript)

205 files, 730 declarations, 83 external references.
~4750 bytes, ~1188 tokens (1 token ≈ 4 bytes English heuristic).

## src/components/sections/hero.tsx

**Ranked declaration**
- L5 `function Hero()`

## src/components/sections/animated-sections.tsx

**Ranked declaration**
- L7 `function AnimatedSections()`
  > // Keep the overview brief; supporting evidence is available on demand.

[Other ranked declarations omitted from this excerpt.]

showing 40 of 725 top-level declarations, ranked by relevance to `Hero` (~5% of the codebase).

Impact

repotoire impact Hero .
# Impact: Hero

**Definition**
- `src/components/sections/hero.tsx:L5` `function Hero`

**Imports**
- `src/components/sections/animated-sections.tsx:L1` (named in import bindings — needs rename even if unused in file body)

**Compiler Pressure**
- `src/components/sections/animated-sections.tsx:L10` compiler-pressure medium: generic type argument [type_surface]

**Internal Uses In Scanned Files**
- `src/components/sections/animated-sections.tsx:L10` value reference in `AnimatedSections`
- lower bound (repo-wide): this repo has 22 unresolved call/reference sites overall — impact dependents derived from them may be incomplete across this repo (these sites bind to no declaration: dynamic/member dispatch, unresolved or undefined references). Ambient repo-wide context, not a claim that this query's listed dependents were dropped.

**Limitations**
- External consumers (other packages, npm dependents) are not visible to repotoire.
- receiver-patterns: Resolved: `this` / static / `new C()` / `const x = new C()` / typed binder.
- factory-return: Resolved: factory return (plain / const-arrow / generic / static-method).
- property-chain: Resolved: `this.f.m()` / `obj.f.m()` chains, depth ≤ 4, including init/return/arg.
- accessors-fields-conflict-rule: Resolved: accessors / fields / compound assigns; declared-type wins.
- out-of-scope-factory-gaps: Out of scope: aliased / chained / instance-method / body-inferred / union factories.
- out-of-scope-chain-gaps: Out of scope: chains > 4 hops and reassigned fields.
- out-of-scope-inheritance-generics: Out of scope: subclass dynamic dispatch, generics beyond strip-and-capture.
- out-of-scope-dynamic-and-decorators: Out of scope: dynamic bracket keys, eval, computed/abstract decorator gaps.
- out-of-scope-misc: Out of scope: destructured binders, return-value flow, visibility modifiers.
- (full per-variant detail: `repotoire limitations <id>`)
- Dynamic references (`obj[name]`, `eval`, computed property access) are not tracked.
- References through unimported bindings (used without a matching `import`) are not connected to the declaration.

Recording details

2026-09-10 · RepoToire 0.1.0-beta.9
apps/repotoire-web · 205 scanned source files
Marketing worktree with uncommitted website changes

Base revision 443498a579e7f3166f86ad8164a0f2f2eff54aa7

src/components/sections/hero.tsx
SHA-256 cb983104508a287a748c3430d9e4ce078fd573111f004455aba9cdfd8f77db50

src/components/sections/animated-sections.tsx
SHA-256 332d8d1032d952cda5e2cf9f897de317a787151b6af456b90b4e271a3a293b79

Full Context output SHA-256 18856be30b4d9fc920b359720b2d762e28456024e3a0290e94edc6b78ad300e3

Full Impact output SHA-256 d278920c376b99cd1cbec9a308e1b92af518d548de6a9f86f3b4686a924e0f3c

Private beta for Codex and Claude Code.

Less coordination.
More verified work.

That’s why we’re building the CPU for agents. RepoToire’s Agent Processing Unit (APU) carries human intent through evidence, bounded execution, and verified retirement. The aim is more completed work with less time spent guessing, coordinating, and redoing it.

Today’s beta starts with local code context. The current execution model is serial; parallel work with verified results accepted in order is the long-term goal. End-to-end repair proof is still in progress.

Read the architecture

A few essentials.

Independent pieces. One shared form.

Intelligence, in order.

Start with code facts today. Help shape the runtime that carries your intent through to verified work.