brokk-bifrost-searchtools

Fast, structured code search and analysis for Python, backed by a native Rust extension. This is the Python distribution of bifrost, the Tree-sitter-backed analyzer suite that powers Brokk.

It gives you one in-process client that talks straight to the Rust analyzer (no subprocess, no MCP server) and returns typed results plus ready-to-render text. It understands Java, JavaScript, TypeScript, Rust, Go, Python, C++, C#, PHP, and Scala.

• Install: pip install brokk-bifrost-searchtools
• Import as: bifrost_searchtools

What it does

Index a project once, then ask fast structural questions:

• Symbol search: find classes, functions, and fields by name pattern.
• Locations and sources: jump to a symbol's definition or pull its source.
• Summaries and listings: get signature-level file/class outlines, immediate git-visible directory children, or package types and child packages.
• Usages and call graph: scan references to a symbol, or build the whole-workspace caller/callee graph (feed it to PageRank for a code map).
• Most-relevant files: rank the files most related to one or more seed files.

Quick start

from bifrost_searchtools import SearchToolsClient

with SearchToolsClient("/path/to/project") as client:
    # Signature-level outline of a file
    print(client.get_summaries(["src/main.py"]).render_text())

    # Find symbols by name pattern
    for file in client.search_symbols(["parse_*"], limit=10).files:
        print(file.path)

    # Rank the files most related to a seed file
    print(client.most_relevant_files(["src/main.py"]).render_text())

The client indexes on first use, keeps the index warm for the session, and watches the filesystem so later queries see your edits. Every result has typed fields plus a render_text() helper.

For a runnable end-to-end demo, examples/searchtools_demo.py declares its dependency inline (PEP 723), so uv fetches the package and runs it with no manual install:

uv run examples/searchtools_demo.py --root /path/to/repo Calculator compute

API overview

SearchToolsClient(root, library_path=None, render_line_numbers=True, manual=False) exposes:

Pass render_line_numbers=False to drop line numbers from rendered text while keeping the structured line metadata on the result objects.

The slopcop tools return a CodeQualityReport (.report), and the rest return structured dataclasses from bifrost_searchtools.models. The per-rule tuning knobs on the smell reports are passed through options (keys map 1:1 to the Rust tool arguments).

Location navigation results expose a NavigationOperation. Declarations use DeclarationLookupResult.declarations; definitions use DefinitionLookupResult.definitions. Their statuses distinguish no_declaration, no_definition, and ambiguous. Reference-based definition lookup is unchanged, and there is no declaration-by-reference method.

Exact source positions use 1-based lines and 1-based Unicode code-point columns, with an exclusive end position. Individual usage hits carry line, column, end_line, and end_column; definition, declaration, and nested type candidates carry start_line, start_column, end_line, and end_column. Column fields are absent for aggregate rows or candidates whose exact name token cannot be proven. Byte offsets remain internal.

query_code detail and ranges

query_code is a versioned typed query surface. Omit schema_version for the compatible head, currently v6, or pass schema_version=2 to pin the pre-CFG vocabulary. Pass schema_version=3 to pin CFG without typestate or schema_version=4 to pin typestate without declaration-bounded containment. Pass schema_version=5 to pin containment without value flow. Pass exactly one of pattern, union, intersect, or except_; set operands are complete query-plan dictionaries with compatible terminal domains. A structural pattern or composed set can be followed by ordered steps using enclosing_decl, file_of, imports_of, importers_of, supertypes, subtypes, members, owner, procedure_of, cfg_entry, cfg_exits, cfg_successor_edges, cfg_predecessor_edges, cfg_edge_source, cfg_edge_target, typestate, value_flow, and witness. Import operations traverse one direct project-local edge per step. Hierarchy operations are direct by default and accept a positive depth or transitive: true. Declaration results are limited to declarations indexed by the workspace analyzer, so references into an unindexed library do not manufacture library declarations. Results are tagged as structural matches, declarations, reference sites, call sites, expression sites, receiver analyses, procedures, program points, control edges, typestate findings, typestate witnesses, flow endpoints, flow witnesses, or files. The CFG operations are procedure-local and one-hop. Schema v4 can consume a host-registered typestate protocol/binding pair and project its already-retained bounded witnesses; the Python call sends only protocol_ref. Schema v5 adds inside_decl, which matches containment only until a nested function, method, constructor, or lambda boundary. Schema v6 consumes a host-registered ValueFlowPlan by plan_ref and preserves reachability, exact/may certainty, ambiguity, completion, and budget status as independent diagnostic-neutral fields. It does not classify policy findings. Compact output retains minimal pipeline provenance. Pass result_detail="full" when follow-up tooling needs deterministic IDs and precise ranges.

Omit execution_mode (or pass "results") for the ordinary CodeQueryResult. Pass execution_mode="explain" to receive a CodeQueryExplain containing the normalized query, logical DAG, selected physical plan, and scheduling decision without executing the query. Pass execution_mode="profile" to execute it and receive a CodeQueryProfile; its .result is the ordinary typed result, while .explain, .timings_ns, .work, .cache_layers, .scheduling, and .operators expose structured observations. Explain physical nodes declare any semantic_request, while profile .work.semantic accounts materialization, cache reuse, CFG rows, retained bytes, traversal, and budget exhaustion; .work.semantic.typestate accounts solver reuse, findings, retained witnesses, and termination; .work.semantic.value_flow reports flow solves, meetings, endpoints, witnesses, and termination. Profile timings are elapsed nanoseconds, and temporary_capacity_bytes_lower_bound is deliberately only a lower-bound container-capacity estimate. In the public v2 profile contract, top-level and per-operator .cache_layers are lists of {layer, metrics} records. The nested metrics object has kind="structural_facts" for seed_structural_facts and kind="complete_value" for every other layer. The direct_import_topology layer additionally reports snapshot build files, edges, time, retained bytes, cancellation/unavailability, and request-local fallbacks.

For decorated or annotated declarations, node_range is the matched normalized node's parser-backed range. decorator_ranges are the decorator or annotation role spans extracted by the language adapter. decorated_range is the union of node_range and those decorator ranges. Matching semantics are unchanged by requesting full detail; these fields only make the span policy explicit.

Current structural precision

query_code normalizes common syntax across Python, Java, JavaScript, TypeScript, Go, C/C++, Rust, PHP, Scala, C#, and Ruby, but it is still a syntactic structural query tool. Use these caveats when writing reusable rules or prompts; the public Code Querying guide is the canonical reference:

License

Apache-2.0. See LICENSE.md.