Graft

Graft is a command-line context layer for coding agents, written in TypeScript under the MIT licence. It reads a repository twice: a deterministic tree-sitter pass that produces a per-symbol graph with cross-file call and import edges, and an optional model pass that summarizes every file and groups those summaries into a few dozen readable markdown nodes joined by typed links. The result is a gitignored folder, and every query re-checks the working tree against the last build’s fingerprint before answering — about three milliseconds when nothing moved — so answers describe uncommitted edits as readily as committed ones. Agents reach it three ways: a marker-fenced block or an owned skill file written by an init command that knows nine agent hosts, six MCP tools the agent calls itself, and Claude Code hooks that pull matching nodes into each prompt and print an edited file’s blast radius. Deep summaries use whichever provider key the user supplies; the structural graph never calls a model at all.

Who built itThe repository belongs to the trailhq organisation, so the name here is the heaviest contributor rather than an owner. Of its 512 commits Shrish Dwivedi wrote 235 and Anirudh Kumar 160, against 25 for Dependabot, 25 for Alex Matthews and 20 for Frankie-Xu; thirty-six accounts appear in the contributor list, and the 111 co-author trailers add about ten more names, most of them once. The project grew out of a NanoNets repository — the npm package is still published under the NanoNets scope, and its own development instructions still clone NanoNets/context-graph-engine.

1. 01

   Thirteen weeks, two people, fourteen tags

   The repository was created on 2026-07-03 and its first commit is Initial commit: Context Graph Engine; by 2026-09-30 it held 512 commits — 243 in July, 213 in August, 56 in September, a first rush that slows once the project must be maintained rather than built. Two people wrote most of it: Shrish Dwivedi at 235 commits and Anirudh Kumar at 160, against 25 for Dependabot, 25 for Alex Matthews and 20 for Frankie-Xu. Thirty-six accounts appear in the contributor list, and the 111 co-author trailers add about ten more names, most of them once: 47 name a Claude model, 20 say Cursor, 25 are the dependency bot. Versions are tags rather than releases — GitHub records no releases at all, while the tags run from v0.7.1 through v0.8.1, v0.8.2, v0.9.0 and v0.12.1 to v0.21.1, with v0.10, v0.11 and v0.17 absent from the record, and a 39,765-byte CHANGELOG.md carrying what release notes would. Upgrades are quiet by design: the CLI checks npm once a day, announces a new version, and the next session refreshes the repository’s own wiring once it is installed. The project also runs on itself, visibly — a bot comments a blast-radius diagram on each pull request and links to a hosted page for it, which is what the blast workflow, its cache workflow, a 15 KB Pages publisher and a 9 KB composite action exist for. Around all of it: 9,431 stars, 865 forks, 188 open issues.

2. 02

   Two tiers, two caches, one three-millisecond stat

   Understanding is built in two tiers with unrelated costs. The first is deterministic tree-sitter: every function, class and call edge, no model and no key, written as graft/.graph/wiring.json and per-file cards. The optional deep pass adds a per-symbol summary and crux, then groups the file summaries into concept nodes joined by typed links. Every pass, the parse included, is cached by content hash, so the README’s own figures for this repository — 124 files — are 0.74 s cold, 0.18 s after one file is edited and 0.18 s with nothing changed; --no-reuse forces the cold path. That cheapness pays for the retrieval design. Rather than a watcher keeping an index warm, every query stats the tree against the previous build’s fingerprint — about 3 ms, structural, no model call — and rebuilds only if bytes moved, which is why ask, grep, callers, skeleton and map describe uncommitted and unstaged edits alike, and why GRAFT_REFRESH=hash exists for anyone who distrusts size and mtime. graft check is the deliberate exception: it never refreshes and exits 1 on drift, because it is the drift report. Injection rides the same artifacts: a marker-fenced block or an owned skill file written by an init command that knows nine hosts, six MCP tools the agent calls itself, and Claude Code hooks that probe each prompt and print an edited file’s blast radius afterwards.

3. 03

   The speed and cost claims, and how each was measured

   The headline figure is the project’s own: up to 4× cheaper and 3× faster, with methods printed beside it. The first ran three variants of a Claude Sonnet 5 agent with the same file tools — cold, a graft ask --source bundle pushed up front, and a pull variant with only the MCP tools — scored by an Opus 4.8 judge against a required-keyword floor, with cache-aware costing at roughly 0.1× for reads and 1.25× for writes, over 162 runs, two repositories and three trials each. It reports cost per task falling from $0.0429 to $0.0292 (+32%), tokens from 8,070 to 4,650 (+42%), tool calls from 4.2 to 2.3 (+46%) and latency from 39.8 s to 15.8 s (+60%), correctness equal at 93%, and 98% for the pull variant that injects nothing. The second is SWE-bench Verified: 50 instances, the same model on both arms, the official 4.1.0 grader — 27 of 50 resolved against 33, 142.0M tokens against 109.4M, $52.34 against $42.43, 1,370 tool calls against 1,031, 13,094 s of wall clock against 8,922 s. The third re-implements five merged PocketBase pull requests from their base commits across 15 tasks: $13.91 down to $11.02, 2,044 s down to 1,762 s, all five reproduced. The table at the top of the README mixes the first two — its correctness row, 54% to 66%, is the SWE-bench result, while the controlled sweep found correctness unchanged. Nothing in the material is a third-party replication of any of it.

4. 04

   What users measured back

   Several of the thirty most recent issues are counter-measurements rather than bug reports, and they are the sharpest material in the history. One user compared graft ask --source on a 291-file Swift and Kotlin application against their own baseline of one ripgrep plus a targeted read: the build took 3.7 s and produced 3,217 nodes and 9,743 edges, the right code was in the pack for all three questions, but the top hit was correct for only one, and the packs were 2.5× to 4× larger than the grep path — a ranking complaint with numbers attached. Another paired 331 prompt-hook outputs with the prompts that produced them and found that most of the injected lines saying there was no strong match came from machine-generated background-task notifications rather than from user prompts. A third documented a plain folder of repositories, never indexed, where every turn marked it dirty and every Stop spawned a build that hit its 120-second timeout: 0.8 of a core and 2.5 GB of RAM held continuously, the stats file reading syncedAt: null with zero nodes. That one also carries the clearest admission in the material — a folder polluted by an earlier version stays adopted, so the new gate still lets it through until the stale cache is deleted by hand.

5. 05

   Generated files that overwrite what people edited

   Two of the loudest bugs share a cause: graft writes into files that are not wholly its own. In one, the session-start wiring refresh read .claude/settings.json through a bare JSON.parse inside a try/catch that fell through to an empty object whenever parsing failed, then wrote the merge back through a truncate. A file left holding merge-conflict markers — or read while another writer had it — therefore came back with the user’s permissions.deny, permissions.allow, outputStyle and hooks gone, and modified enough to commit. The pull request fixing it removes the fall-through, on the principle that a settings file which fails to parse must never be overwritten. The second case is one user filing seven issues in a single sitting about the generated instruction block: regeneration silently reverts hand-written corrections inside the graft marker fences, deleting the generated ignore file was itself reverted within the same session, graft build appends the wiring files to .gitignore while its output mentions only the cache directory, the template points at graft/INDEX.md while the build writes graft/index.md — a dead path on any case-sensitive filesystem — and the ignore file re-admits the card tree through an unanchored pattern. The node files were designed for exactly this pressure: anything a user writes below the generated block survives regeneration.

6. 06

   Adapters, one closed pull request, and a metric checked first

   Much of the pull request traffic is one genre: making a model gateway behave. One reports that OpenAI-compatible gateways answer HTTP 200 with an error object and no choices, which threw a TypeError inside the response parser, and that some models cannot serve a forced tool_choice at all, leaving files and concept batches empty, reproduced against OpenRouter. The follow-up retries a structured call a reasoning model cut off at its length limit, with four times the token allowance, capped at 32,768. A third fixes a crux pass where models echo the whole target row back as the identifier, dropping every summary. The newest parallelizes concept-synthesis batches behind a new --synth-concurrency flag defaulting to four, with a second knob because phase one is hundreds of small calls and phase two a few 48,000-character requests; the graph stays byte-identical to the serial run. Around those sit the maintainer’s habits: a 52-file pull request bundling a SQLite stats store, five host adapters and a raised Node floor was closed unmerged, asking for an issue first and focused changes keeping the Node 20 floor; the telemetry event added for the hosted Trail workflow carries closed sets and buckets only and is pinned by a contract test; and before adding a count of suggestions waiting for review, the author checked the database, found zero such events in a week, and shipped it anyway.