Sverklo

Permanently delete a memory after recall returned a stale or wrong entry. Prefer remember (with the new content) over forget+remember when superseding a decision — supersession preserves the audit trail via valid_until_sha + superseded_by; forget loses it. Get IDs from recall or memories.

[Zilliz claude-context compat] Alias for sverklo's `status` tool. Returns current indexing progress and statistics. Provided for drop-in compatibility with Zilliz claude-context.

Filter the result blocks of a prior sverklo tool call by a regex pattern — operates on the cached text so no second retrieval happens. Pass the response_id returned on any search/refs/impact call.

Keep only the top N result blocks of a prior sverklo tool call. Cheap way to shrink a chatty response when you just need the top hit.

Refactor blast-radius: callers of a symbol with confidence scoring. Run before editing. Use cross_repo:true to see impact across linked projects in a workspace.

[Zilliz claude-context compat] Index (or re-scan) the current codebase. Sverklo indexes automatically on startup and via file watcher; calling this triggers a manual rescan. Equivalent to sverklo's built-in indexing.

Single-call research primitive: fans out to BM25, embeddings, symbol lookup, and reference-expansion in parallel; RRF-fuses the candidates; returns one ranked bundle with per-hit provenance (which retriever(s) found it). Cheaper than running search + refs + lookup back-to-back. WORKS WELL for open-ended questions where you don't yet know whether the answer lives in code, callers, or documentation — the `found_by` tags tell you which signal agreed. Use search instead when you already know you want pure text/semantic retrieval.

List all indexed repositories in the global registry. Use this to discover available repos when running in global (multi-repo) mode. The repo names returned here can be passed as the `repo` parameter to any other sverklo tool.

Look up a specific symbol (function, class, type, variable) by name. Returns its full definition, signature, and location.

List all memories for the current project, or surface conflict candidates. Default mode shows memory health (staleness, confidence, access frequency). Pass `mode: "conflicts"` to surface pairs of active memories that share a pin and may contradict — the bi-temporal model preserves both, so this is a review prompt for the agent or human, not an auto-resolution.

Get a structural map of the codebase. Shows the most important files and their key symbols ranked by dependency importance (PageRank). Use this FIRST when starting work on an unfamiliar codebase or directory.

Query the LLM-derived design-pattern annotations on indexed symbols. Pass a `pattern` from the closed taxonomy (observer, repository, validator, ...) to list every symbol tagged with it; pass no args to see the taxonomy + counts. Requires `sverklo enrich-patterns` to have been run.

Pin a memory to a specific file or symbol. Pinned memories surface automatically when recalling by that file path or symbol name, without needing semantic search.

Promote a memory to the core tier. Core memories are auto-injected into every session via sverklo://context resource — use for project invariants that should always be in the AI's context (style rules, framework conventions, 'never do X' rules).

Search memories semantically. Finds past decisions, preferences, and patterns relevant to a query. Supports two specialized modes: `mode=core` returns only the always-on project invariants (fast, no query needed — use at session start); `mode=archival` searches the full archive with semantic ranking; `mode=all` (default) searches both. Use core for 'what are the project-wide rules I must not violate' and archival for 'what did we decide about X on this codebase'.

Find all references to a symbol across the codebase. Shows where a function, class, or type is imported, called, or used. Matches on identifier word boundaries by default — `embed` does NOT match `embeddingStore`. Pass `exact: false` to opt into substring matching.

Save a persistent memory tied to git state. Auto-invalidates conflicting prior memories.

Diff-aware context bundler for code review. Takes a git ref or range and returns: changed files, semantic delta (added/removed/modified symbols), dangling references for removed symbols, impact set for modified symbols, and similar-symbol detection for added ones. Replaces 10-20 grep+read calls with one structured response. Use this FIRST when reviewing an MR/PR — it surfaces blast radius and convention violations grep cannot see.

Hybrid semantic + text search with PageRank ranking. WORKS WELL for: exploratory questions where you don't know the exact symbol ('how does auth work', 'find anything related to billing', 'where's the retry logic'), anti-pattern discovery ('swallowed exceptions', 'silent null returns'), and cross-file semantic matches. STRUGGLES WITH: framework registration and wiring questions ('how is X registered as a bean', 'where is this interceptor configured'). For those, grep the specific annotation (@Component, @Configuration, etc.) directly. Response includes a confidence signal and a fallback hint w…

[Zilliz claude-context compat] Alias for sverklo's `search` tool. Semantic + text hybrid code search using embeddings and PageRank. Provided for drop-in compatibility with the Zilliz claude-context MCP server.

Wider-pool, iterative-friendly variant of search. Returns a ctx:// handle to the top-200 candidate pool plus refinement hints (co-occurring symbols, dominant directories, concept overlap). Use ctx_grep / ctx_slice to refine without firing another retrieval. Worth the extra latency (~50ms) on hard multi-hop questions; for single-shot lookups use search.

Project state + tool usage guide. Returns index health (files, chunks, languages), memory summary, and specific tool recommendations tailored to this codebase. Call this first when starting a new session to understand what sverklo knows about the project.

Map a git diff to its test coverage. Given a ref/range, lists which tests likely cover each changed source file (via name heuristics + import graph), flags changed source files with NO matching tests, and shows which test files were modified in the diff. Use this during MR/PR review to answer 'what tests should I run?' and 'is this change tested?' without grepping the whole repo. Coverage is heuristic — sverklo doesn't run code, so treat results as candidates, not ground truth.

Remove a pin from a memory.

Check whether one or more evidence ids (from a prior search-family tool's ```evidence block) still point to the same code they did at retrieval time. Returns unchanged / moved / modified / deleted / file_missing per id. Use this to prevent hallucinated citations after code changes.

500-token codebase summary suitable for pasting into a system prompt when the agent has no MCP access. Use this when integrating sverklo intelligence into Cursor, Codex, or a CI bot that can't run the MCP server. For MCP-connected agents, prefer status and the richer per-tool surfaces — they return more and stay in budget.