sinter

A local code knowledge graph for coding agents, shipped as one static binary. sinter builds a typed, directed graph of a repository's symbols with tree-sitter and keeps it fresh incrementally. Start in an unfamiliar repository with sinter map: it returns a bounded structural inventory with module node/file counts, dependency hubs measured by graph in-degree, documentation entry points, and graph-health limitations. It does not infer runtime entry points or domain ownership. Then use the graph for focused work:

• Reverse blast radius — what transitively depends on a symbol, cross-file and cross-language (sinter affected). Traversal stops at hubs and names them; test dependents are counted, not listed, unless asked for.
• Paths — how one symbol reaches another (sinter path).
• Diff impact — which symbols and tests a changeset can affect (sinter impact), plus --expect <symbol> for the unfinished-refactor check: direct dependents of a symbol the diff did not touch.
• Bounded text search — a regex search whose corpus is a graph traversal (sinter grep '<regex>' --within 'affected(<sym>)'), replacing the run-affected-then-grep-the-files pipeline.
• Polyglot, zero setup — syntax-only graphs work without an LSP. When the syntax provides the necessary type information, sinter resolves self receivers, locally typed receivers, typed fields, and explicit static/class calls. Complete receiver-call binding often requires a fresh SCIP index; zero setup is not compiler-complete resolution.
• Calibrated lexical navigation — sinter ask ranks content-bearing starting points from names, docs, paths, signatures, string literals, and limited call evidence. Its confidence line reflects the evidence class of the top hit; --explain adds the ranking-margin bucket and that bucket's holdout count with a 95% interval (the holdout is 46 cases, so the interval is wide). The bucket is not a per-result probability; ask is a navigator, not a semantic answer engine.
• Symbol orientation — sinter show turns a selected symbol into a compact card with its definition (attributes included) and a one-line used-by tally; --body adds the source (whole when it fits in 60 lines, else up to the byte budget), --impls the type's impl blocks, and show @file:line names the enclosing symbol. A span over 8 KB or 200 lines is a black hole no body dump can show: the card outlines it instead, listing the nested definitions, literal branches and command literals inside it with their line numbers (--outline forces it).
• Task evidence packet — sinter context "<task>" resolves identifiers in the task against real node names (fuzzy when needed, shown as term ~> symbol) and seeds from them, returning edit candidates, string literals and hand-maintained mirrors that mention the task's terms, and affected tests as runnable commands with the symbol that reached each. It never abstains while it has hits. Add --workspace <manifest> to rank candidates across declared member repositories in one packet.
• Snapshot-scoped assertions — sinter assert no-callers <symbol> checks depth-one call edges in an explicit corpus scope. Its decision is one of violated, holds_for_indexed_snapshot, or not_proven; it never claims runtime exhaustiveness. sinter assert deletable <symbol> tallies every depth-one dependent across all scopes, grouped by scope.
• Citation maintenance — sinter cite <symbol> emits a Markdown location with a stable symbol key. sinter verify-doc <file.md> re-resolves managed citations and fails on moved, missing, invalid, or identity-free references.
• Cross-repo workspaces — federated graphs over many repos for distributed systems: blast radius, paths, and PR impact across service boundaries (--workspace).

The design rule underneath everything: evidence or nothing. An edge exists only when structural, scope, import, or compiler (SCIP) evidence binds a reference to a definition (workspace manifests may add operator-declared cross-repo links, and Rust trait impls add labeled dynamic dispatch fan-out so blast radius survives dyn Trait). Ambiguity resolves to nothing — "unresolved" is a first-class, counted outcome, never a guess. Every edge carries its evidence kind, and every query can filter on it. Be clear about what that buys without a compiler: syntax-only evidence binds names it can prove by scope and imports and leaves the rest (receiver calls, re-exports, macros) unresolved, so the zero-setup graph is a precise subset, not the full graph. sinter scip closes that gap where a compiler index is available.

Install

Package managers:

cargo install sinter-io        # or: cargo binstall sinter-io (prebuilt)
uv tool install sinter-io      # or: pipx install sinter-io
brew install shellfu/tap/sinter

All of them put a binary named sinter on PATH. Or the one-liner, no dependencies (Linux and macOS; verifies the release checksum, installs to ~/.local/bin):

curl -fsSL https://raw.githubusercontent.com/shellfu/sinter/main/scripts/install.sh | sh

Windows, in PowerShell (new and less battle-tested; verifies the release checksum, installs to %LOCALAPPDATA%\sinter\bin):

irm https://raw.githubusercontent.com/shellfu/sinter/main/scripts/install.ps1 | iex

To verify a manually downloaded release asset against its GitHub build provenance attestation (requires the gh CLI):

gh attestation verify sinter-<target>.tar.gz --owner shellfu

PyPI wheels carry sigstore provenance via Trusted Publishing. Once installed, sinter update self-updates from GitHub releases.

Or build from source (requires a Rust toolchain):

cargo build --release

Quickstart

Agents that only need a usable local graph should create derived state without installing hooks or editing client configuration:

sinter ensure /path/to/repo

This command only builds or refreshes .sinter/. It is safe to run within a read-oriented coding flow.

Onboard a repository — builds the graph, installs git hooks, registers agent integration (AGENTS.md block, MCP), and finishes with a doctor report:

./target/release/sinter init /path/to/repo

Init prints everything it is about to write, grouped by scope, and asks once before writing any of it (-y skips the prompt; a non-interactive run prints the plan and proceeds). Every write lands inside the repo — nothing under ~/.claude is touched unless --global is passed, which adds the machine-wide skill card and enforcement hooks. Both init and ensure append .sinter/ to the root .gitignore when the repository is a git worktree and no existing line covers it.

Repo-local Claude hooks use bounded strict enforcement: the first broad recursive search in a session is redirected to Sinter, while a retry is allowed with a fallback-search reminder. Machine-wide hooks installed by --global remain advisory.

On a terminal, init asks before running compiler indexers (sinter scip) because those toolchains execute repository build scripts; pass --scip or --no-scip to answer up front (non-interactive init skips them).

After that, every query self-syncs at the query boundary — when nothing changed, the sync is a stat-only walk (no file reads, no write transactions) — and the git hooks refresh eagerly on commit; sinter build stays available for CI and scripting. Commands also work from any subdirectory, discovering the graph root the way git does. The build report distinguishes the heuristic's anchored miss rate from compiler-relative accuracy. The anchored rate is useful without SCIP but is not recall: the heuristic can classify a compiler-resolvable reference as external.

resolution (this pass): ... resolved (scip 0, import 118, scope 189), ... unresolved (105 internal, 3193 external)
anchored miss rate (this pass): 25.5% (heuristic classification, not compiler-relative recall)

When a SCIP index is present, the report also prints the compiler cross-check (what share of internally-bound refs agree with the compiler) and internal recall vs the compiler (how many compiler-bound refs sinter found without SCIP). Files edited since the index was built get no SCIP evidence at all — their references fall back to import and scope resolution rather than being rebound by byte position — so a stale index never attributes a call to a name that no longer sits on that line. Files the grammar could only parse partially are counted in one build line (N parsed partially (M symbols in them; …)); sinter doctor --verbose lists them.

Orient before searching so the next query uses the repository's own module and symbol vocabulary:

sinter map /path/to/repo

The map is the first-pass inventory. Use ask next when the target is still a concept rather than a known symbol.

Ask a question against the graph (output shown for this repository):

$ sinter ask "where is the trigram search"

1. function Store::search    [doc+name+path+sig 2/2 terms]
   crates/sinter-store/src/search.rs:148
   /// Fuzzy candidates: nodes sharing the most trigrams with the query,
   pub fn search(&self, query: &str, limit: usize) -> Result<Vec<Node>, StoreError>

ask is a calibrated lexical navigator, so treat its hits as places to inspect rather than generated answers. Every hit shows its match provenance and the text card carries one confidence line (confidence: high — verify top hit) rated from the top hit's evidence; a weak top hit is unrated. String literals that match the question are listed after the symbol hits, and fixture, example, and test-local definitions rank below production ones with the same evidence. Agent JSON groups results by topic under one strict result budget and keeps each hit lean: rank, name, kind, file, line, signature, doc, matched terms, channels, and a one-word confidence. --explain (text or JSON) adds the ranking margin, query-term coverage, the named holdout calibration, and per-hit scores. Query terms expand through a small synonym table (cap, limit, budget; size, bytes; caller, dependent) at reduced weight, so a literal match always outranks a synonym. Weak singletons, low term coverage, and undersampled ranking-margin buckets abstain. CLI JSON and MCP structuredContent.data use the same payload.

Every affected, deps, and path response is deliberately bounded. deps defaults to depth 1 (--max-depth widens). affected stops at hubs (fan-in over 100, or a seed with more than 50 direct callers) and says so (stopped at hub Store (fan-in 114); --through-hubs to continue). Both count test-scope rows instead of listing them (--include-tests), and --relations calls,uses drops file-level import rows. path -k N returns up to N node-disjoint routes; a miss prints the closest frontier, the edges the filter excluded, reason: filter_excluded when a filter caused it, and a retry to run. In --json, positive and negative answers carry a coverage summary (completeness, conclusive, snapshot, compiler-index state); --coverage restores the repository-wide block with evidence sources, certain and possible result counts, unresolved-reference counts, and explicit gaps. The universe field names the repository root or every declared workspace member searched. completeness describes only the indexed snapshot and conclusive remains false, so agents do not turn a non-empty syntax-only result into an exhaustive claim.

Agent-facing node id values are stable symbol keys and survive unrelated offset shifts. snapshot_id retains the byte-exact locator for the reported snapshot. Handle-consuming operations accept if_snapshot and return typed stale-snapshot, relocated-handle, or ambiguous-candidate outcomes instead of silently rebinding. A symbol whose bare name is ambiguous can be addressed as Name@file or Name@file:line (run@doctor.rs:175).

Test Sinter on a real repository

Sinter is recruiting ten design partners who use coding agents on repositories where dependency or blast-radius investigations are expensive. Bring one real task. The evaluation will compare the investigation with and without Sinter and record where the graph helped or failed.

Open a design-partner issue to propose a public repository or describe a private repository without naming it. Do not include private source code, credentials, or customer information in the issue.

Commands

MCP registrations use the portable sinter command and start non-required so a missing binary cannot prevent the client from starting. sinter doctor checks that the command resolves to an executable on PATH and performs an MCP handshake.

Reads open the graph with a shared lock; a rebuild holds it exclusively. A query that arrives during a rebuild queues for up to two minutes and prints one waiting for another sinter process notice after a second, so parallel agents see a delay rather than a Database already open error. Leaked MCP servers from finished sessions keep their original binary; when versions differ each one rewrites the graph in its own format, so sinter doctor lists the ones serving this repository with their pids.

affected, deps, and path accept --evidence scip,import,scope,dynamic and --certain to restrict traversal to stronger evidence tiers, and --relations calls,uses,imports,implements,extends,reads,writes,creates,alters,drops to restrict which edge relations are followed (e.g. drop file-level import edges from a blast radius); sinter grep accepts the same traversal filters for its --within bound; their MCP counterparts take the same filters as evidence (array), min_confidence: "certain", and relations (array) parameters.

MCP tools share the CLI defaults: scope is the CLI corpus (production,test,docs; ask uses production,docs), deps is depth 1, affected stops at hubs. Arguments are validated against the advertised schema (tools/list carries enums and one-line descriptions), so max_depth: "two" or relations: "calls" is an invalid_arguments result, not a silently ignored filter. Every user-fixable failure — unknown symbol, ambiguous name, bad argument — is an isError tool result whose structuredContent.error carries code, message, and Name@file candidates; outcome.status (complete, partial, not_proven, not_found, error) and outcome.reason (limit_reached, filter_excluded, …) are the one place to branch. Paging covers the whole result: limit: 0 is unlimited, next_cursor is set whenever rows remain, cursor resumes. symbol echoes are trimmed to key, name, file, line; the coverage block is omitted unless include_coverage: true. show, deps, and affected take symbols: [...] and path takes pairs: [[from, to]] for one result per entry. budget_bytes below the smallest answer returns that answer flagged, not an error; next_actions are tool calls.

ask defaults to the production,docs corpus; every other verb to production,test,docs. Use --scope or the MCP scope array to include fixtures, examples, generated files, or vendor code. Exact show remains unfiltered. Repositories can exclude paths in .sinterignore; classify fixture corpora and apply ordered overrides in .sinter.toml (committed) or .sinter/config.toml (local, wins) with gitignore-syntax patterns:

[scope]
fixture = ["worked/**", "tools/*/expected/**"]

[[scope.override]]
pattern = "tools/golden-production/**"
scope = "production"

Without configuration, any path segment named fixture(s), golden, testdata, expected, worked, or snapshot(s) classifies as a fixture and example(s), sample(s), or demo(s) as an example.

Languages

Rust, Go, Python, TypeScript, JavaScript (ESM, CJS, JSX), Java, C#, C, C++ (including Unreal Engine macro conventions), SQL (DDL/DML), Bash, Protobuf, and Markdown (headings become nested section nodes with the opening paragraph as doc, so sinter ask finds prose docs with file:line provenance; [text](https://github.com/shellfu/sinter/blob/main/target) links become uses edges from the linking section to the target file — or, with #fragment, to the section whose heading slugifies to it — when and only when the target resolves to a file in the corpus: dead links stay unresolved, external URLs produce nothing). A language is pure data — a tree-sitter grammar, one .scm capture query (plus an optional secondary inline grammar, spec-declared), and a spec row — consumed by a single engine that never branches on language. Adding a language requires no engine code; if it ever does, the capture contract is wrong, not the language.

Syntax-only binding covers, per language: Python import x as y aliases, function-local from m import f calls, and calls inside nested defs; TypeScript new X() constructor calls, default exports, and export * barrel re-exports; Go interface implementations matched by method set across packages (an affected on an interface method that could not traverse implementations says gap: implementations not traversed). Explicit implementations — Rust impl Trait for T, Java, C# and TypeScript implements, class inheritance — carry a method-level implements edge too, so sinter affected Trait::method reaches the implementing methods and not only the trait method's callers.

An edge keeps every call site it saw, up to eight, plus the true total: a caller that calls a symbol twice renders as path/to/file.rs:12, :48, and --json adds sites and sites_total when there is more than one. Graph schema v16 carries these facts; an existing graph rebuilds once on the first query after upgrading.

SQL graph

For .sql files, Sinter emits table, view, column, and index nodes. It also records the direction and purpose of object references:

• reads: a SELECT, FROM, or JOIN source
• writes: an INSERT, UPDATE, or DELETE target
• creates, alters, drops: schema changes owned by the SQL file
• uses: foreign-key and index dependencies

A view owns the reads in its defining query. Top-level statements belong to their file, including migration files.

SQL embedded in Rust is extracted too, at known query sinks and for literal strings only: the sqlx::query!/query_as!/query_scalar! macros, the sqlx::query* and diesel sql_query functions, and the query*, execute*, batch_execute, and prepare* methods of tokio-postgres, rusqlite, and connection pools. The enclosing Rust function gains the reads/writes edges, so a table's blast radius crosses the language boundary. Table names resolve within a database root first (a directory holding migrations and queries), then fall back to the unique corpus-wide definition.

To find every query, migration, and Rust function that touches a table:

sinter affected users --relations reads,writes,creates,alters,drops

To prove nothing in the indexed snapshot writes, alters, or drops a table:

sinter assert no-writers users --json

To inspect the data and schema dependencies of one migration:

sinter deps migrations/20260901_users.sql \
  --relations reads,writes,uses,creates,alters,drops

sinter doctor folds migrations in filename order and warns when a table is dropped at head but still referenced, or referenced but never created.

Limits. The SQL grammar (tree-sitter-sequel) does not parse several PostgreSQL constructs — CREATE FUNCTION/PROCEDURE/TRIGGER/POLICY, DO blocks, row-level security. Files containing them are indexed from partial trees: the statements the grammar dropped are absent from the graph, and doctor reports the gap as one row naming the likely constructs. SQL strings built at runtime, and SQL embedded in Go, Python, or TypeScript, are not extracted. Sinter does not infer column-level lineage, transaction scope, indexes required by a query, or planner behavior; use EXPLAIN against the target database for planner evidence.

Top-level graphify-out/, memory/, and .memory/ are excluded from the semantic corpus and SCIP freshness inventory. These are derived analysis products; indexing them would feed generated summaries back into ask and dependency answers.

If a compiler-produced SCIP index (index.scip) is present at the repo root or at .sinter/index.scip, sinter ingests it as the highest evidence tier. sinter scip discovers configured project roots from build markers and runs only matching indexers that are available on PATH (rust-analyzer, scip-go, scip-typescript for TS and JS, scip-python, scip-clang for C/C++, scip-java, scip-dotnet), merges the results into .sinter/index.scip, and rebuilds. Isolated source files and fixtures without a project marker do not trigger repository-level indexer recommendations. Bash, proto, SQL, and Markdown have no SCIP indexers.

Teams

Graphs are per-machine and rebuild in seconds; the SCIP index is the expensive shared artifact. Build it once in CI (sinter scip), distribute the file, and each teammate's next build ingests it automatically. Recipe, cache-key guidance, and a copy-paste workflow: docs/team.md.

Accuracy and performance are measured, not asserted

• Golden corpus: hand-verified fixtures (82 at time of writing) across all thirteen languages, mined from real-world extraction idioms. Extraction and resolution both gate CI at precision/recall 1.0 for the enumerated contract facts; any change that moves the metric fails with the exact missing/extra tuples printed. Expectations derive from language semantics, never from engine output (harness/golden/).
• Real-repository evaluation: 249 hand-labeled, syntax-only navigation tasks run against pinned releases of ripgrep (Rust), Cobra (Go), Flask (Python), Hono (TypeScript), and Gson (Java): 3 exact lookups, 166 natural-language ask questions labeled by intent with a tuning/holdout split, 36 direct-caller checks, and 44 path checks. The current scorecard reports ask top-1 accuracy 0.705 (holdout 0.696), MRR 0.810, recall@5 0.945, recall@10 0.986, and p95 latency 55 ms. Callers score precision 1.000 and recall 0.523, paths 0.727: syntax-only graphs miss receiver-typed and Java static-class calls that a compiler index binds, and the harness labels those sites anyway. The weekly and manually dispatched workflow uploads the full scorecard and syntax-only build timings (harness/eval/).
• Agent-flow evaluation: eighteen deterministic, network-free coding flows cover multi-step graph use, bounded agent responses, task context, and a labeled design-similarity question that must surface both existing mapping tables before inspection. The current scorecard passes 9/9 flows and 21/21 steps with one correct abstention and zero unsafe-confidence failures. These are observational scenarios, not a claim of general end-to-end coding accuracy.
• Budgets (measured on a ~2M-LOC Go repository, 271k nodes, before the stat-gated scan landed): full build 18s, one-file edit under 1s typical, cold point query under 100ms, ask 66ms end-to-end. A clean sync is now a stat-only walk — no file reads, no write transactions — measured 46→16ms on this repository and 55→10ms on an 80MB/400-file synthetic corpus; the old 73ms 2M-LOC no-op figure predates the stat gate. Cold open and incremental-edit budgets are CI-enforced (crates/sinter-store/tests/cold_start.rs, crates/sinter-cli/tests/incremental_build.rs); the full-build, warm-query, and memory budgets are measured manually per harness/perf/README.md.

Workspace layout