@cheppulabs/sonde

Sonde

A local code-context engine for AI coding agents. Sonde indexes a TypeScript, Python, or Swift repository into a symbol-level graph in SQLite and exposes three MCP tools — find_symbols, query_graph, and get_impact_radius — so an agent can answer who calls this, what breaks if I change it, and which tests relate to it in one call instead of a search loop.

What the benchmark actually shows

The honest claim is not "finds what grep cannot". A competent agentic search loop finds the same structural evidence — we measured it, on a real 19,409-line repository, and it scored 1.000 recall on every task.

The claim is the same answers for a fraction of the cost, inside a budget:

On real production TypeScript	Sonde	Agentic search
Recall on structural tasks	1.00	1.00
Tool calls	1.0	8.0
Context tokens	1,262	3,621
Latency	263 ms	38,602 ms
Runs that blew the token budget	0 of 6	3 of 6

Sonde matches the baseline's answers on every structural task while using ~3× less context, 8× fewer calls, and ~147× less wall-clock time — and it never exceeds the caller's token budget, because the packer truncates to it by construction. That is the trade being offered.

Where it loses. Behavioural queries with no shared vocabulary — "where is the retry backoff decided?" — score 0.00. Local semantic retrieval was built and measured and does not fix it (see the design doc §2.2); the capability is therefore not claimed. If your questions are mostly of that shape, an agentic search loop is the better tool today.

Numbers are reproducible: npm run bench:fixture && npm run bench:large. Full results in BENCHMARK-LARGE.md and BENCHMARK.md.

Install and run

npm install -g @cheppulabs/sonde
cd your-project
sonde init

sonde init indexes the repository and registers sonde as an MCP server in this project's .mcp.json, asking before it writes anything (skip the prompt with sonde init --yes). It never touches an .mcp.json it can't safely merge into — an existing sonde entry that differs from what init would write is left alone and reported, not overwritten.

For Python repositories, use sonde init --resolve; the default tree-sitter tier did not pass the project's placement gate for structural queries.

Equivalent by hand, if you'd rather see every step:

sonde index .
# then add to .mcp.json:
# { "mcpServers": { "sonde": { "command": "sonde", "args": ["mcp", "serve", "."] } } }

No account or hosted service is required.

Architecture documentation

sonde doc writes an ARCHITECTURE.md describing the repository's modules, how they depend on each other, and what each one exposes — generated from the graph, so it reports what the code actually does rather than what someone remembered.

sonde doc            # write ARCHITECTURE.md
sonde doc --stdout   # print it instead
sonde doc --check    # fail if it is out of date (for CI)
sonde doc --module src/store   # symbol-level detail, never committed

It is meant to be committed and regenerated, not hand-edited. Regeneration is byte-identical when nothing changed, so it does not churn your diffs; when two branches both regenerate it, resolve the conflict by running sonde doc again rather than merging by hand. It refuses to overwrite an ARCHITECTURE.md it did not generate.

Three things it deliberately does not do:

• It does not draw a dependency it cannot evidence. Module pairs that merely share symbol names are excluded and counted separately. On this repository two adapters share the filenames symbols.ts, parser.ts and references.ts, which manufactured 62 heuristic "references" between modules that never import each other — once the second-heaviest arrow in the diagram.
• It does not pretend the diagram is complete. The diagram shows the heaviest dependencies and states how many it omitted; the table below it goes further, and --module has the rest. A diagram containing every dependency is unreadable and therefore shows nothing.
• It does not claim to be current when it is not. The header names the commit it describes and warns when files have changed since.

What it guarantees

• Never returns stale source bytes. Whenever a response includes source, Sonde re-reads and re-hashes the indexed byte range before returning it (spec §8.1, Guarantee A).
• Always reports structural drift, rather than claiming completeness it cannot verify (spec §8.1, Guarantee B). sonde status shows the same drift and tier distribution carried by tool response envelopes.
• Every edge is tier-labelled by how it was found — COMPILER (resolved exactly by a bundled type checker under --resolve: the TypeScript compiler for TypeScript, pyright for Python), LEXICAL (resolved through an import binding or lexical scope), HEURISTIC (member access or another relationship requiring type inference), EXTERNAL (target outside the indexed repository), or UNRESOLVED (genuinely unplaceable, with a reason).
• Never fabricates an edge. An unresolved reference becomes EXTERNAL or UNRESOLVED — never a guessed target and never a silently dropped reference.

Accuracy

Sonde measures its zero-setup tree-sitter path against the TypeScript compiler on a pinned fixture and publishes the result, unflattering numbers included (spec §12). COMPILER edges use that compiler directly, so comparing them back to the same authority would not be an independent accuracy test.

Sonde edge accuracy vs the TypeScript compiler

Generated: 2026-08-23T18:16:41.557Z TypeScript: 5.9.3 (bundled; repository TypeScript is never loaded)

What these numbers cover. The oracle measures the tree-sitter resolution path — the zero-setup default, and the only tier whose accuracy is in question. COMPILER-tier edges come from the TypeScript compiler itself, so scoring them against the same compiler would measure nothing; they are exact by construction and excluded from these figures. Run sonde index --resolve to produce them.

The oracle is filtered to in-repo targets; node_modules and .d.ts declarations are excluded. Type-only references, JSX intrinsics, export =, decorators, and declaration merging are known expected divergences (spec §10). Tier rows compare that tier alone with the complete oracle, making each tier's independent contribution visible; ALL is the combined result.

Why precision below 1.000 is expected here

These divergences are structural, so reading a precision figure as "how often Sonde is wrong" overstates the error rate:

• Ambiguous member calls emit every candidate. For x.foo() with two visible foo declarations, Sonde emits both as confidence-weighted HEURISTIC edges. At most one matches the compiler, so the other counts as a false positive by construction. The alternative is guessing a single target, which invariant 1 forbids — a wrong resolved-looking edge is worse than two honestly heuristic ones. Precision is therefore capped below 1.000 wherever the fixture contains an ambiguous call.
• Constructor calls are ours alone. Sonde emits CALLS for new Foo(); the oracle does not model them, so each one is a false positive against ground truth that omits it.
• Member-level IMPLEMENTS is ours alone. Sonde derives an IMPLEMENTS edge from RegExpRouter.add to Router.add once the class declares it implements the interface. tsc reports heritage clauses at the type level only, so every member-level edge counts as a false positive against ground truth that does not model them. The capability is the reason impact on an interface method works at all, so the precision cost is disclosed rather than removed.

Counts are absolute, not percentages of a large corpus. Fixture edge totals appear below so a single edge's effect on each figure is visible.

tests/fixtures/repos/small

Fixture config SHA-256: e02e2d5003f96d1ad22519f04e10d687fe689cf9298e7fcbc588eab525dce1ad

Oracle edges: 9 · Sonde edges: 7 · one oracle edge moves recall by 11.1%

Edge kind	Tier	Precision	Recall	TP	FP	FN
CALLS	ALL	0.500	1.000	2	2	0
CALLS	LEXICAL	0.500	0.500	1	1	1
CALLS	HEURISTIC	0.500	0.500	1	1	1
IMPLEMENTS	ALL	0.500	1.000	1	1	0
IMPLEMENTS	LEXICAL	0.500	1.000	1	1	0
IMPLEMENTS	HEURISTIC	1.000	0.000	0	0	1
INHERITS	ALL	1.000	1.000	1	0	0
INHERITS	LEXICAL	1.000	1.000	1	0	0
INHERITS	HEURISTIC	1.000	0.000	0	0	1
REFERENCES	ALL	0.571	0.800	4	3	1
REFERENCES	LEXICAL	0.800	0.800	4	1	1
REFERENCES	HEURISTIC	0.000	0.000	0	2	5

Overall: precision 0.571, recall 0.889

Regenerate with npm run bench:oracle.

CLI

sonde init [path] [--resolve] [--yes]      # index and register project MCP config
sonde index [path] [--resolve]             # full index; optional compiler pass
sonde update [path] [--resolve]            # update; optional compiler pass
sonde status [path]                        # freshness and tier distribution
sonde search <query> [path]                # find_symbols
sonde query <pattern> <symbol> [path]      # query_graph
sonde impact [path] --symbol <name>        # get_impact_radius by symbol
sonde impact [path] --from-git-diff        # impact from working-tree diff
sonde doctor [path]                        # parser/database/tsc health
sonde clean [path]                         # remove the cached index
sonde mcp serve [path]                     # MCP server over stdio

impact also accepts repeatable --symbol options and --token-budget <n>. The init, index, update, status, search, query, impact, doctor, and clean commands accept --json.

Known limitations

• Node 22+ is required. better-sqlite3 needs it; installing on an older Node prints an EBADENGINE warning but still completes. If sonde then fails to run, this is why — upgrade Node rather than ignore the warning.
• Python needs --resolve, and its edges are not oracle-verified. Without it, Python indexes at the tree-sitter tier only, which measured 62.81% unresolved on a real 56-file project and 57.39% on pydantic — far past the 30% ceiling this project requires, so it is not fit for structural queries. With --resolve, a bundled pyright drives a COMPILER tier and the same corpora measure 27.00% and 17.42% unresolved. That gate measured placement — whether a reference found a target — not whether the target was correct; TypeScript's edges are scored against tsc in ORACLE.md and Python has no equivalent oracle. On the worse corpus the margin is thin: 0.28 points once a known upward bias is reversed. See probes/python-placement/FINDINGS.md.
• TypeScript, Python and Swift only; no other language adapter.
• Swift resolution is heuristic, not compiler-backed, and one narrowing rule is unvalidated. References are narrowed by file visibility and explicit local type annotations, not full type inference. On a real 376-file Swift application this reached 74.84% placed / 25.16% unresolved — see probes/swift-narrowing/FINDINGS.md for the full measurement. The third narrowing rule (SwiftPM target boundaries) has never been tested: the validating corpus was an Xcode project, which has no Package.swift layout to supply that signal. The curated Swift SDK symbol table is also deliberately conservative — several ambiguous names were dropped rather than guessed, so some legitimate SDK references may still show as UNRESOLVED rather than EXTERNAL.
• Compiler resolution for TypeScript (--resolve) is opt-in because it is materially slower and uses more transient memory. On the 19,409-line Hono fixture, default indexing took 3.58 s; --resolve took 13.70 s, added 10,329 exact placements/promotions, and changed callers_of Hono.route from zero graph results to eight compiler callers. The Program is discarded after indexing; inline refresh stays compiler-free and explicitly downgrades affected evidence until a full sonde update --resolve.
• Compiler resolution uses bundled TypeScript 5.9.3, not the repository's own compiler, so version skew is possible and disclosed in doctor and response envelopes.
• A file with a parse error still contributes whatever tree-sitter recovered from it — this is deliberate (see the design spec §8). Its parse_state is partial when real declarations were recovered despite the error, or failed only when nothing usable came out of it at all.
• TESTS edges indicate structural relatedness only; they never prove coverage.
• Type-only references, JSX intrinsics, export =/import =, decorators, and declaration merging are known gaps in the tree-sitter extraction path.
• No rename inference. Renaming a file changes every stable key derived from its path; anything holding an old key (a saved query, an agent's prior turn) will silently stop resolving rather than following the rename.
• Semantic/behavioural search is not available. It was built and measured — two embedding models, four document configurations — and none of them beat lexical/structural retrieval on the task class it was meant to help with. See design spec §2.2 for the numbers. It is not wired into find_symbols.

See docs/superpowers/specs/2026-08-16-sonde-design.md for the full design.