Skip to content

Roadmap ​

Where Aburi stands today, and what comes next.

Status

The v1 schemas are frozen. Every published package is implemented, unit tested, and exercised end to end against a NestJS-shaped fixture project.

Working today ​

Analysis. TypeScript and JavaScript, parsed with tree-sitter WASM. Aburi detects pnpm, npm, Yarn, and Bun workspaces and nx projects on its own, recognises NestJS, Next.js App Router, Express, React, and the tRPC server as frameworks, and reads Prisma, Drizzle, tRPC client calls, and NestJS events as effects.

Dependencies. Aburi resolves calls to the symbols they reach within a file, an import, a component, or the workspace, and builds a symbol-level dependency graph from them. It draws an edge between two components wherever a file of one imports a module of the other. It counts and buckets the calls it could not resolve, so you can see what the graph is missing. Effects propagate along that graph, so a controller that reaches prisma.invoice.create through a service carries the db.write too.

Diff. All seven statuses (added, removed, moved, changed, moved+changed, dropped-toggled, and unknown for a file one revision could not read), a --fail-on gate, and Slice View, which clusters changed symbols along the call graph so a feature cutting through controller, service, and repository reads as one section.

Output. The JSON analysis, a workspace overview, per-component detail, the pull request diff, and per-symbol explain views.

Delivery. @aburi/cli on npm, and @aburi/github-action for pull request comments.

Performance. Measured monthly against nine public repositories at pinned commits, from zustand's 49 files to cal-com's 5,002. On a 4-core runner a 1,081-file scan takes 3.7 s and a 5,002-file one 14.7 s, single-threaded, with no scan passing 470 MiB peak — cost is roughly linear in file count. Three consecutive scans of an unchanged tree produce byte-identical IR. aburi diff runs its two scans at once on a machine with two or more CPUs, so it takes about 1.3× one scan where it took 1.7–2.2×; on one CPU it still runs them in turn.

Not yet ​

LimitationStatus
TypeScript and JavaScript onlyOther languages planned below
Call resolution is syntactic unless a language server is configured — 6–29% of call sites resolve on public repositoriesOpt-in LSP enrichment (lsp.enabled with lsp.servers.ts) resolves this. / super. calls; interface-typed receivers are not resolved yet
No graph visualisation beyond the workspace overview and Slice ViewNo design yet
No LLM integrationDeliberate. Aburi produces the facts, and interpreting them belongs to a separate tool.

Next ​

  • LSP enrichment. Use type resolution to sharpen effect inference and fill in column-level source ranges. The number it has to beat is measured on the benchmark's public repositories: from 6.3% of call sites resolved on shadcn-ui to 29.0% on zod, with NestJS itself at 8.1%. no-match is the largest unresolved bucket everywhere but NestJS, whose constructor-injected interfaces make dynamic larger, and a hand-classified sample of it puts 70% ± 5% of it within reach of a rule that needs no types. Little of that is a missing edge, though: most of it is a global, a binding of the caller or a method of a builtin or package type, filed in the wrong bucket. Rules would turn 4.9% ± 2.3% of no-match into call edges, and type information another 2.0% ± 1.6%.

Later ​

  • More languages. Python and Go, with uv, poetry, cargo, and go.work workspace detection, and a single analysis spanning all of them in one monorepo.
  • More effects plugins. Django, FastAPI, SQLAlchemy, GORM.
  • Large monorepos. A 1,000-file scan finishes in under 30 seconds single-threaded, so the worker pool would buy at best ~4× on a 4-core runner — the ideal scaling, unmeasured — on a workload already inside its budget.
  • A functional language. Scala or Rust, proving that the extension vocabulary can express pattern matching and algebraic data types.

Under consideration ​

Nothing here is committed.

  • A merge-base ref form, aburi diff main...HEAD, comparing the head against the point it branched from the way a GitHub compare URL reads. Today the three-dot spelling is rejected with a message pointing at git merge-base.
  • Expose Aburi as an MCP server, callable from AI coding agents.
  • aburi review, feeding the diff to a model for automated review comments.
  • Name Slice View clusters automatically. The report heads each with the members its calls start from today.
  • Warn when the two analyses aburi diff --base --head compares were made with different plugins, from the plugin list each one records.
  • aburi vocab, listing the effect ids and extension kinds the loaded plugins declare, and which plugin owns each.
  • --quiet for scan and diff, and aburi diff --filter, narrowing the report to chosen statuses.
  • Colour and a progress line on an interactive terminal, honouring NO_COLOR and FORCE_COLOR.
  • Browse the analysis in a web UI.

The design documents specify everything above, and the JSON Schemas live in schema/.

Released under the Apache License 2.0.