Symvanta

Call first, with repository:"owner/name" (your checkout's GitHub remote). Binds the session to the project holding that repository and returns it as the active project with its repositories, index health, and a `usage` field (decision matrix + routing). workspace.attached=false means the checkout is NOT indexed: nothing in the answer describes it; attach it (add_repository, create_project) or work without the graph, do not route through the projects listed. Without repository: the pinned project (ref op:"use_project") or the workspace default.

Search/lookup. Pick a mode, or omit with a bare query to AUTO-ROUTE (text, then semantic on empty; response carries answered_by). Modes: - text (replaces grep): literal identifier/string, one row/occurrence. queries:[...] (<=10 terms) searches several in one call (and implies this mode when mode is omitted). pathsOnly:true = "filePath:lineNumber" only (~5x cheaper). maxPerFile caps rows/file (default 5). Empty carries did_you_mean; do NOT retry reworded, use semantic. - symbol: symbols by name/symbolPath in one repo. kind filters node kind, layer one layer. Tests last unless includeTests:true. - semantic: vector search for symbols similar to a natural-language query (names unknown). - file: fuzzy filename search (basename); filenames only. - config (replaces grep for env/config keys): key/env-var usages with nearest-symbol context; queries:[...] batches several keys. - codebase: cross-repo candidate scan; narrow with repositoryIds.

Graph relationship traversal via one kind (real edges, not text). Kinds (1-10 symbols; heritage/chain one): - callers: direct callers. granularity:'file' (default) one row/file, 'site' one row/caller. - dependencies: outgoing calls/imports. - blast_radius: upstream impact walk (every caller, importer and referencer up to 3 hops, no per-hop sampling) + co-change siblings; risk {level,signals}. includeCrossRepo defaults ON here (off for others). blastRadiusTotalCount is the number of distinct symbols that 3-hop walk reached, not the row count: the blastRadius array is capped by expansionLimit. blastRadiusTruncated:true means the rows are a sample of that total; blastRadiusTotalAtLeast:true means the count stopped at its 10000 bound, so quote it as "N+". The walk stops at 3 hops, so it can miss impact further away; it never overcounts. Never report the number of listed rows as the impact size. - implementers: classes implementing interfaces via heritage edges. - heritage: full hierarchy (ancestors+descendants) of ONE type. maxDepth caps it. - chain: call-chain paths for a node (nodeId OR selectors[0]); chain params tune the walk. - path: shortest connection between selectors[0] (from) and selectors[1] (to). Direction-agnostic hop-labeled path; found:false when none within 6 hops. minimal:true = {id,name,filePath,startLine}-only (~50% cheaper) on callers/dependencies/blast_radius/implementers; granularity on callers/blast_radius. Rows may carry confidence (tiers in init usage); a cross-repo callers/dependencies row joined by a SCIP moniker also carries package (the package name that bridged the repos). 0 results: call locate (mode:text), not Grep; reason:"no_edges_indexed" = no edges.

Raw file/git access at the indexed commit, for cloud agents with no clone; with a clone prefer your own Read/Grep/git. Ops: - read: content of filePath, whole file or slice startLine..endLine. ifHash (a prior read's hash) returns {unchanged:true}, no content. compact:true strips blank/comment-only lines. - list: files+subdirs at path. grep: ripgrep pattern (glob/pathPrefix). tree: layout from path. stat: size/lines/language/binary for filePath. blame: authorship for filePath. diff: fromSha..toSha. Keys: {op:"read",filePath:"src/a.ts"}, {op:"grep",pattern:"foo"}, {op:"list",path:"src"}. filePath is the file, path the directory. Gated by the Source Access add-on; else source_access_required.

Indexed commit history. Ops: - commits: recent commits (metadata + file list); <=200/call, default 50. With a clone prefer git log. - commit: one commit by sha with its diff summary (added/modified/deleted/renamed); null if outside the indexed window. - recently_changed: files ranked by change frequency in the indexed window ("what's hot").

Library catalog lookups. Ops: - packages: list indexed npm packages (specifier, ecosystem, owner/name, indexed-at, sha). Tenant-scoped, no projectId. - version: one package's resolved version for a project (sha, version, dep type, indexed-at, status indexed|indexing|failed|not_found). "indexing" means a job for that revision is in flight, so retrying later converges; "failed" carries a reason and does not change on its own.

Session-scoped branch / working-tree overlay / project binding. Ops: - use: pin a repository's reads to a branch (latest indexed revision). Branch must be tracked+indexed (open a PR, or add on the dashboard); else "indexing_in_progress". commitSha overrides. - clear: remove the pin, revert to default. - index_working_tree: overlay UNCOMMITTED edits (changedFiles path+content, optional deletedPaths/baseSha) on a short-lived synthetic revision pinned here; find_node/locate/relate reflect them, source and locate mode:semantic do NOT. Cap 100 files/256KB. - use_project: bind this session to projectId (from init.otherProjects / list_projects / create_project) so calls without projectId resolve there. init with repository:"owner/name" sets the same binding from the checkout's remote. - clear_project: drop the binding, back to the workspace default project.

Architecture decision records: the WHY behind code, anchored to a symbol/file/project. Survive reindexes; find_node attaches a node's decisions. Ops: - record: needs title+decision. Anchor with repository+symbolPath or +filePath; omit both for project-wide. Optional context, consequences, status. - list: records for a project, filter by repository/symbolPath/filePath/status. - update: needs id; status:"superseded"+supersededById, or "deprecated".

Impact of a diff between two indexed revisions of one repo (base..head): changed files + symbols, union blast radius (layers + cross-repo fallout), affected endpoints, test suites, and co-change reminders. Replaces per-symbol relate (kind:blast_radius) loops before merging. Defaults: head = session's effective revision, base = previous indexed. For UNCOMMITTED edits, run ref (op:"index_working_tree"), then call with no shas.

Pre-flight estimate before graph traversal: coarse count of files a task likely touches and whether it spans layers (common/browser/node/electron-main/...) or repositories. Decides single-file edit vs consulting the graph. Heuristic, from symbol-name + text matches.

List all projects with indexing summaries. Find a project by name or check which exist.

List repositories in a project. Find repository names to scope a query to one repo.

List symbols in one file (name + kind + line bounds). Step 1 of the two-step read: get line bounds, then pass startLine/endLine to source (op:read) for that slice only (~4-15x fewer tokens). For one symbol + signature, use find_node {symbol, filePath}.

Token-efficient repo (or subtree) skeleton: directory tree with each file's classes (methods in parens), functions, and HTTP endpoints, from the graph. Orient before drilling in. Scope with "path", cap with "maxDepth", density:"full" adds properties. view:"architecture" instead returns a module map (Louvain clusters with hubs, paths and cross-module coupling), plus namedModules/totalModules: a module NAME comes from a generated summary and is provisional, its id (the member-set hash in the header) is the stable key, so pin modules by id and re-read names rather than caching them.

Test suites and cases whose suite name matches a symbol name (case-insensitive substring). Call before adding tests to see existing coverage and mirror its shape.

Fast substring lookup of symbol DEFINITIONS by name: rows {name, kind, filePath, startLine, repositoryName} (displayName only when distinct from name), no ranking, no semantic. Cheaper alternative to locate (mode:symbol) for latency-sensitive pre-fetch (grep-augment hook); interactive search should prefer locate/find_node. Reads the latest indexed revision, no ref pin.

Returns lastIndexedSha, lastIndexedAt, and the current remote HEAD sha for a repository: detect index drift from remote. Reports what Symvanta knows, not local git. currentRemoteSha is null when no GitHub credential is configured. Echoes the active ref pin (pinnedBranch, pinnedSha).

Per-project index health. pendingLibraryVersions: cross-repo specifiers still indexing (an empty result may mean "not indexed yet", not "not in code"); call when a query is surprisingly empty. A row with status "skipped" is never going to arrive (a dev or optional dependency, which the catalog does not index) and status "failed" means discovery found no source, so only "indexing" is worth waiting on. architecture: per-repo Louvain module count + modularity Q (low Q = tangled), plus circular-dependency and mutually-recursive-symbol counts (inspect via map view:"architecture"). versionDrift: joined cross-repo SCIP edges whose consumer-pinned package version differs from the sibling's current version (package, pinned vs current, edge count). externalsCoverage: per-repo count of external references joined onto a sibling definition vs left as materialized stubs. unindexableRepositories: repos whose last index produced no graph and left a status (no_default_branch = the remote has no resolvable default branch, e.g. an empty repo; awaiting_upload = an upload-provider repo with no zip yet; remote_head_unresolvable = the remote could not be reached to resolve its head, so this is a network or credential problem, not a repository one), with the worker's log line as indexStatusDetail; explains a silently-empty repo. degradedRepositories: repos that DID index but lost a derived artifact (community_detection_skipped = Louvain clustering did not run, so map view:"architecture" has no modules; raise COMMUNITY_EDGE_CEILING or set COMMUNITY_DETECTION=true, then reindex). Both carry a human-readable reason.

Resolve an HTTP route by path + optional method to endpoint + handler (filePath + line bounds). Use instead of grep/locate for URL paths: follows framework router metadata (Express, Fastify, Hono, Laravel, FastAPI, etc.) and resolves dynamic segments, prefix groups, and middleware-mounted routers text search misses.

Resolve 1-10 symbols to filePath + line bounds + signature. selectors: array keyed by symbol (bare name) / symbolPath / filePath / nodeId; a flat query/name/single field also works. includeSource=true adds the excerpt, includeSummaries=true AI summaries. A node from a SCIP-indexed package also carries scip_package_name (and scip_package_version / scip_manager when known). For a class/interface/struct/protocol it carries a compact heritage field (<=5 ancestors, <=20 descendants, omitted when empty); includeHeritage:false skips it, includeCrossRepo:true walks cross-repo heritage. {resolved:false, candidates} = no confident match.

Synthesis Q&A for behavior/architecture questions: answer + citations in one call. HEAVY (~2-4 s, ~3-5 k tokens): prefer context, find_node, or locate when they suffice. scope:"repo" (default) = ONE repo; scope:"all" = cross-repo (repositoryIds narrows, omit spans all; rejects repository). If sufficient_to_answer=true, answer without further tool calls. commitSha pins a revision.

Attach a GitHub repository to a project (owner + repo_name; clone URL derived; projectId defaults to the bound / default project). Public repos need nothing else. Private repos need installation_id from list_installations (a GitHub App installation that covers the repo) or a workspace PAT saved on the dashboard; without either the attach is refused (private_repository_needs_credential) instead of failing on the queued clone. Idempotent. First attach dispatches an indexing job; response carries visibility, credential ("installation" | "pat" | null), indexing_status ("queued" | "already_indexed") and indexing_job_id to poll. Needs mcp:admin.

Create a project in the workspace (name; slug derived and made unique). Needs mcp:admin; plan project caps apply (plan_limit_exceeded). Then attach repositories with add_repository { projectId, owner, repo_name } and bind the session with ref op:"use_project" (or init with the checkout's repository once attached).

GitHub App installations connected to this workspace and the repositories each can reach (private included). Use before add_repository for a private repository: pass the installation's id as installation_id. pat_configured says whether a workspace PAT exists as the fallback credential. Empty = connect the GitHub App on the dashboard (Project > Repositories). An installation whose GitHub call failed carries error and an empty list.

Dispatch a reindex job for an already-attached repository (repositoryId from list_repositories / add_repository). Optional commit_sha pin. force:true re-applies the current parser even when unchanged at HEAD: it bypasses the per-file hash-skip AND re-runs the SCIP indexer subprocesses (no same-sha .scip cache serve), so a parser fix inside a SCIP runner, an external-moniker re-pin, or a toolchain upgrade takes effect without a version bump.

First-touch task orientation. Pass a natural-language task; returns the top 5 relevant files with their symbols and best-matching snippets in one call (~1-1.5k tokens, ~500 ms, no LLM). Replaces chaining locate (text) + locate (semantic) + list_file_symbols. Known symbol: find_node/locate. Synthesized explanation: ask_codebase.

Run up to 10 independent tool calls in parallel in one round-trip. Calls share the bundle-level project/repository scope unless a call sets its own args.repository, which scopes that call alone. Supported: locate (args.mode text|symbol|semantic|file|config, not codebase), relate (kind:callers only), find_node, list_file_symbols, find_http_route. Returns results in input order; unsupported tool names return {error}.