Incremental Graph Patching¶

Update an existing CodeGraph in place when code changes, avoiding a full re-index whenever the language backend supports a safe delta. Most languages use an LSP server; C/C++ refreshes clangd .idx data instead.

There are two entry points:

IndexCompiler.update_repo() (recommended) — the compiler drives the patcher for you, alongside every other index, under admission control. See Compiler-Driven Updates.
LSIndexer.graph_patch() — the low-level primitive that patches a single graph directly. The rest of this page documents this layer.

Compiler-Driven Updates (Recommended)¶

IndexCompiler.update_repo() productionizes graph patching: it advances an existing manifest (written by an initial compile_repo() build, see MCP server) to the repo's current HEAD, routing each requested index through its builder's incremental_update() instead of a full build():

from codenib.compiler import IndexCompiler
from codenib.compiler.index_builders import (
    IndexBuilderRegistry,
    register_default_builders,
)

registry = IndexBuilderRegistry()
register_default_builders(registry, languages=["python"])

compiler = IndexCompiler(registry)
compiler.compile_repo("/path/to/repo")   # initial full build

# ... new commits land ...
compiler.update_repo("/path/to/repo")    # incremental advance to HEAD

update_repo() compares the manifest's last_indexed_commit against HEAD and then:

Nothing to do — HEAD unchanged and every requested index fresh: returns the existing manifest untouched.
Retry incomplete builds — HEAD unchanged but some requested index is missing or not fresh (a previous run failed partway): re-runs a full compile_repo() so the failed indexes are retried instead of staying stale.
Full rebuild fallback — no manifest, an unreadable manifest, or an empty last_indexed_commit (no complete baseline was ever established): falls back to compile_repo().
Incremental advance — otherwise, each builder's incremental_update(..., last_commit=<previous>) runs. Builders without a real delta path (BM25, Zoekt) rebuild internally, the vector builder applies a git-diff-driven embedding update, and the symbol-graph builder runs the LSP patcher described below — the result is always correct, only the cost differs.

last_indexed_commit Semantics¶

The manifest claims HEAD as last_indexed_commit only when every requested index built successfully. If any build fails, last_indexed_commit is left at the previous commit (empty when no full build ever succeeded), so the next update_repo() retries the failed indexes rather than reporting "nothing to update" and leaving them stale forever. Failed entries are still recorded in the manifest with status="failed" and the error in their metadata.

Admission Control for the Symbol Graph¶

An incremental symbol-graph result is only admitted when the patched artifact can be taken as equivalent to a fresh rebuild. SymbolGraphBuilder delegates that decision to the UpdateVerifier contract (codenib/compiler/verification.py):

Verifier	Behavior
`NullVerifier` (default)	Proves nothing and says so: `verified=False`, `checked=False`, reason `"no verifier configured"`.
`AlwaysAdmitVerifier`	Admits without checking (`verified=True`, `checked=False`). An explicit opt-in for environments that have accepted the risk deliberately.

With the builder's default require_verification=True, an unadmitted patch is discarded and the graph is fully rebuilt — so out of the box the symbol-graph path behaves exactly like a full rebuild until a real verifier is configured. To accept unverified patches, construct SymbolGraphBuilder with AlwaysAdmitVerifier() (or require_verification=False) and register it in place of the default. Either way the outcome (verified, verification_checked, verification_reason) is written into the manifest entry's metadata, so admission is auditable on disk. A patch that touches no source files of the configured languages is admitted directly.

Independently of verification, the builder falls back to a full rebuild whenever the incremental path cannot run safely: no previously indexed commit, no existing graph.pkl, an unresolvable HEAD, uncommitted changes to tracked files, a missing language server, or any patch failure.

How It Works¶

Given a base commit (matching the current graph) and a target commit:

Detect files and hunks — git diff --name-status identifies modified/added/deleted/renamed files; git diff -U0 provides inclusive, 0-based old/new hunk ranges.
Round 1 (prepare vertices) — all changed files are classified before cross-file edges are queried. Added symbols are created, deleted symbols are removed, and existing shifted/affected symbols are updated in place.
Round 2 (connect edges) — affected symbols rediscover outgoing references; added symbols discover both incoming and outgoing references. Edge writes are batched after all Round 1 vertices exist.
Refresh range indexes — the O(V + E) line-span and call-site indexes are rebuilt so query_range() immediately sees the post-patch graph.

The two-round design ensures a definition added in one changed file exists before another changed file tries to connect a reference to it.

Symbol Classification¶

For modified files, the patcher compares the existing graph with the new documentSymbol snapshot and classifies each symbol:

Class	Meaning	Patch action
`deleted`	Present only in the old snapshot and its old span overlaps a hunk	Remove the vertex
`added`	Present only in the new snapshot	Create the vertex and discover incoming/outgoing references
`affected`	Stable identity, but its new span overlaps a hunk	Keep the vertex; refresh stale outgoing references
`shifted`	Stable identity and body, but its start line moved	Rename/update the vertex and shift its anchored outgoing edges
`unchanged`	Stable identity and line	Leave it untouched
`invisible`	Present in the baseline graph but absent from `documentSymbol`, outside every hunk	Preserve it; this covers definitions the LSP outline cannot represent

An affected symbol takes one of two paths. If every overlapping hunk preserves line count, only outgoing edges anchored in the changed ranges are removed and rediscovered; surviving anchors are shifted uniformly. If a hunk changes line count, all outgoing references for that symbol are cleared and its full body is rescanned. Incoming edges remain attached because the vertex is updated rather than deleted.

When a container overlaps a hunk only because one of its children is affected, the classifier demotes the container to shifted. Its span metadata is still refreshed, but the child owns the reference rescan and avoids duplicate work.

Severed-edge contract¶

A full-file rebuild can temporarily cut references between the rebuilt file and the rest of the graph. Each cut call site is recorded separately as:

(
  source_name,
  target_name,
  source_unified_name,
  target_unified_name,
  anchor_file,
  anchor_line,
)

This is a six-element tuple, not a source/target pair: parallel references between the same symbols remain distinct when their call-site anchors differ. Incoming edges are safe to remap to the rebuilt target. Outgoing edges are remapped only when the source body is still trustworthy; otherwise LSP discovery replaces them. Recorded anchor lines are preserved or rebased through a known file shift before the edge is restored.

LSP Interaction¶

Each language uses a specific language server, started via stdio by LSPClient:

Language	Server	Install
Python	`basedpyright-langserver`	`make python-lsp-tool`
Rust	`rust-analyzer`	`make rust-tool`
TypeScript/JS	`typescript-language-server`	`make typescript-lsp-tool`
Go	`gopls`	`make gopls-tool`
C/C++	`clangd`	`make active-system-deps-ubuntu clangd-tool`

The patcher queries three LSP methods to rebuild the graph:

textDocument/documentSymbol — get the hierarchical symbol tree of a file (used in Round 1 to discover new/changed symbols)
textDocument/references — find all call-sites of a symbol across the project (used in Round 2 to reconnect incoming edges)
textDocument/semanticTokens + textDocument/definition — scan tokens in changed line ranges, then resolve each cross-file token to its definition (used in Round 2 to reconnect outgoing edges)

LSPClient auto-resolves server binaries from PATH / conda / venv / Go, Cargo, npm-global, .NET global tools, and local user bin directories.

C/C++ special case: clangd's background indexer refreshes .idx files for affected translation units. patcher_cpp.py removes the old changed-file subgraphs, reads the refreshed .idx data to restore their vertices and references, then rebuilds the same range indexes as the generic patcher. Header changes can invalidate multiple translation units, so that path may do more work than the source-file list alone suggests. The .idx refresh is prepared before graph mutation; if reindexing, decoding, merging, or range-index construction fails, the patcher restores the original in-memory graph and raises instead of returning a partially updated graph.

Prerequisites¶

An existing CodeGraph (built via LSIndexer.run_pipeline(), see scip_index)
The corresponding language server installed (see table above)
The project must be a git repository with both the base and target commits reachable
The checked-out, tracked working tree must represent target_commit. Symbol and .idx snapshots come from files on disk; the compiler path enforces a clean tree and uses the resolved current HEAD as its target.

Usage¶

from codenib.ls_router import LSIndexer
from codenib.graph.code_graph import CodeGraph

# Load a graph built at commit v1.0
graph = CodeGraph.load_graph("/cache/project/graph.pkl")

# Patch it to HEAD
indexer = LSIndexer("/path/to/project", language="rust")
result = indexer.graph_patch(graph, base_commit="v1.0", target_commit="HEAD")

graph.save_graph("/cache/project/graph.pkl")

base_commit is passed into the patcher as earlier_commit, and target_commit as later_commit. The same pair drives both changed-file detection and every modified file's old/new hunk coordinates. Direct patch_files() callers may omit later_commit; its effective value is then HEAD. The returned commit_earlier and commit_later fields record the effective pair used by the patch.

Supported languages: Python, Rust, TypeScript/JS, Go, C/C++.

Return Value¶

graph_patch returns a stats dict:

{
    "files_modified": 1,
    "vertices_created": 2,
    "vertices_deleted": 0,
    "vertices_shifted": 3,
    "refs_incoming": 5,
    "refs_outgoing": 3,
    "nodes_before": 45,
    "nodes_after": 47,
}

Interactive Demo¶

See the Interactive Demo for a step-by-step walkthrough of classification, in-place vertex preparation, reference discovery, and the final range-index refresh.