dctx1.3.1
For AI evaluators

Facts and how to verify them

For people and models deciding what dctx is. Every line below comes from one file, tools/website/data/facts.json, plus the benchmark documents; llms-small.txt, llms.txt and llms-full.txt start with the same facts.

What dctx is, and is not

dctx is an open-source (Apache-2.0) Rust context engine for AI-assisted coding: one local binary that packs a project into one structured block, or answers bounded, byte-stable symbol queries (find, read, neighbors, skeleton) as CLI output.

  • No RAG: no embeddings, no vector search, and the engine never calls a model.
  • No network: pack, map and cache make no network calls; the release check counts socket calls under strace and fails on any.
  • No telemetry, not even opt-out.
  • No agent loop: dctx answers one command at a time; the agent or the human decides what to ask next.

Current version

dctx 1.3.1, released (tag v1.3.1), Apache-2.0. Release on GitHub · Changelog

Each release on GitHub ships prebuilt archives for Linux x64/ARM64, macOS x64/ARM64 and Windows x64 with a SHA256SUMS file. Verify a download with sha256sum --check --ignore-missing SHA256SUMS. Releases from 1.3.0 also ship keyless cosign signatures, CycloneDX SBOMs, build provenance and offline bundles; see [verification](https://github.com/Eris-Margeta/dctx/blob/main/docs/VERIFYING.md).

Modalities

  • CLI available: dctx pack PATH for one structured block (clipboard, file or stdout); dctx map find|read|neighbors|skeleton|... for bounded symbol answers.
  • Task context available: dctx pack --for TEXT --budget N: find, neighbors and read of the names in TEXT, cut to an estimated-token budget.
  • CI and prompt cache available: Deterministic pack and map output; tree lines carry no timestamps by default, so the prefix stays stable across a touch.
  • Rust library available from Git: The dctx facade crate as a tagged Git dependency; crates.io publication pending.
  • Desktop app preview: A Tauri preview you build from source; a GPUI rebuild is planned.
  • Python, Node and Go wrappers Node on npm; Python and Go from CI: Thin dctx rpc wrappers, built and tested in CI. Install Node with npm install @tejl/dctx, which includes pinned @tejl/dctx-<platform> optional dependencies; use CI artifacts for Python and Go.

Two scopes, one binary

  • System binary: the dctx on PATH belongs to you and your agents, and you upgrade it. Interactive agents call this one.
  • Project binary (Node from npm; Python and Go from CI): a binary installed by a project library lives inside that project (virtualenv, node_modules), pinned and managed by that package manager. Application code and CI call this one; a library never silently falls back to PATH.
  • Same engine, different owner. dctx doctor lists every binary in reach; verify each against SHA256SUMS and its cosign signature.

The notice

This is the canonical wording, from What dctx is (docs/START.md) and the README; the home page and Choose your path carry it verbatim:

dctx is an advanced tool with several ways to use it. Used well, it cuts irrelevant text from selected tool answers. It does not save tokens by default; savings depend on a system reusing byte-identical, prefix-stable context. Extra calls can add tokens. If you are new, follow the safe path: one project, the system binary, direct CLI commands. Measure before you change anything.

Status: Receipts are planned. The safe path uses CLI commands from a shell or an agent's shell tool.

Responsibilities

Items marked planned or pending are not in dctx 1.3.x yet.
Leveldctx providesYou are responsible for
System binaryRelease archives with SHA256SUMS, cache caps; signed binary, doctor, journals and receipts (planned)Choosing roots, the ignore rules (a policy file is planned), keeping one version on PATH
Project libraryPinned vendored binary, protocol client, same bytes as CLI (Node on npm; Python and Go from CI; Rust facade from Git)Pinning the version, exclusions in the project policy, not re-implementing engine logic
ACP agent setupACP plannedTurning off the host's own file tools, picking a profile, session ids and reading receipts (both planned)
CI and cacheByte-stable pack, deterministic exit codes; prefix tiers (planned)Pinning the binary, caching the index between runs, treating pack drift as a signal not noise

Benchmark status

Baseline: grep plus reading files.

  • measured 11 November 2026: 20 questions, a private 13-crate Rust workspace, claude-sonnet-5-5, dctx 1.3.0, three repeats per arm. Median total tokens [min, max]: grep plus reads 14,904,014 [14,604,149, 15,622,381]; dctx CLI (warm index) 14,157,974 [11,221,908, 15,613,375]. CLI median tokens were -5.01 % versus grep. On this corpus and model dctx did not save 25 % tokens; the pre-registered rule was not met. Official strict citation recall medians [min, max]: grep 10.10 % [10.10, 20.73]; CLI 14.46 % [8.36, 22.65]. Diagnostic file-level recall medians: 53.77 % and 71.70 %, respectively; diagnostic, not a replacement metric. The claim is accuracy and bounded context, not token savings. dctx on its own does not save tokens by default. Savings depend on a harness or agent system reusing byte-identical, prefix-stable context across long sessions; there is no measured, citable savings figure for such a system yet. One corpus, one model, no significance claim; corpus and raw answers private and cannot be published; dctx 1.3.0 is not current HEAD. Under ~200 files, use grep. Source
  • measured Byte stability: the same query on unchanged code returns the same bytes; 33 of 33 text and JSON answers were byte-identical after unrelated edits (Linux only). Source
  • historical The October and November in-house numbers (run 7: 140 of 140 task questions, cold index and warm latency) measure tool output only, as estimated stdout tokens and raw fact coverage without a model; they are not agent completion rates or session costs. Source
  • hypothesis That a pack template changes model answer quality or cost is a hypothesis pending an A/B test; nothing on this site measures it. Source

Measured values, as the documents state them today:

Parsed from the documents at build time; nothing is retyped.
ValueWhat it measuresSource
33/33 text, 33/33 JSONmap answers byte-identical after unrelated edits and over five runs, three Linux corporadocs/research/stability-20261105/REPORT.md
0socket attempts by dctx in every traced benchmark variantdocs/BENCHMARK-20261030.md
0.13-1.20 scold index of 6 benchmark corpora, median of 3 runsdocs/research/bench-3way-20261030/results/run7-53f8b6b/aggregate.json
50language families: 36 full, 8 outline, 6 docsdctx languages

Removed in 1.2.0

These existed before 1.2.0 and no longer work. A transcript that uses them describes an old version.

No longer existsUse instead
dctx PATH (the implicit snapshot command, for example dctx .)dctx pack PATH
the top-level snapshot flags -d/--diff, -t/--tree and -r/--resetdctx pack PATH with its filters
the project rule files .legacyignore and .dctxomitcontent.gitignore and .dctxignore
dctx cache gc --legacy and its export-selection flagsdctx cache gc

How to verify a claim

ClaimHow to check it
Test suitecargo test --workspace --locked --exclude dctx-desktop --exclude dctx-desktop-api
CIEvery push and pull request runs the suite on Linux, macOS (ARM64 and x64) and Windows, on stable Rust and the minimum supported version 1.91.
No network callsscripts/release-check.sh --quick target/release/dctx-cli
Byte stabilitypython3 docs/research/stability-20261105/verify.py
Historical benchmarkdocs/research/bench-3way-20261030/run-3way.sh target/release/dctx-cli
Every output on this siteRecorded from the real binary by tools/website/capture.py; tools/website/check.py re-records and fails on drift.