YAML files were classified as documents, so code-only / AST-only runs
(graphify update) produced zero nodes from them. Move .yaml/.yml into the
code path like .json: add a PyYAML-based structural extractor (extract_yaml)
registered in _DISPATCH, add the suffixes to CODE_EXTENSIONS, drop them from
DOC_EXTENSIONS and the splittable-text set.
Deterministic ids derived from the file path; top-level and nested keys become
nodes with contains edges, capped at 500 pairs / depth 6 to avoid exploding on
data-heavy YAML.
(cherry picked from commit 2e5565a684)
Adds a regression test that fails on the pre-fix resolver: a caller that
imports a module for a different symbol and calls a same-named function must
keep the edge INFERRED, never EXTRACTED.
(cherry picked from commit 7d41be8891)
When an ambiguous bare call resolves to a candidate solely because the
caller imported the candidate's source module (not the symbol itself), the
edge was promoted to EXTRACTED, falsely certifying a same-named symbol that
was never actually imported. Keep the resolution (for graph connectivity) but
mark it INFERRED. Symbol-level (from X import Y) imports still promote to
EXTRACTED.
(cherry picked from commit e73ca4e4a7)
The stub rewire skipped dotted labels, but a SQL table is schema.name,
so dotted labels from .sql files now count as definitions. SQL stubs
that share a normalized name and bind to nothing fold into one node
(the smallest id, so file order doesn't matter). .sql joins the
language families, so a SQL read no longer binds to a same-named
definition in another language.
(cherry picked from commit 55bc540bbe)
A table no file defines, read from N files, becomes N per-file stub
nodes, and a schema-qualified read never binds to its definition in
another file. A SQL read of `Orders` also binds to a Python class
Orders. Guards cover same-file binding and different schemas.
(cherry picked from commit 1dfc710497)
GO is a client-tool command, not SQL, and tree-sitter-sql has no rule
for it. Before parsing, each GO [n] line found on the masked text
(comments and strings blanked) becomes `;` plus spaces, so byte
offsets and line numbers don't move. A go line after a comma, or
before a comma or `)`, is a column and stays.
AST cache schema 20 -> 21 so cached .sql extractions re-run.
(cherry picked from commit 2293c31d79)
A GO line with no `;` before it is read as an identifier, so the next
statement lands in an ERROR node and its table is lost. Guards cover
`;`-terminated batches, GO inside a block comment, a column named go,
and GO after a keyword line such as SET NOCOUNT ON.
(cherry picked from commit f00ecc42c9)
tree-sitter-sql has no T-SQL bracket identifiers and no create_procedure
node, so bracket names were mangled and procedures only got a node from
the regex recovery, never an edge. Labels and lookup keys now drop the
brackets (ids are unchanged). Each recovered routine's body, up to the
next CREATE/ALTER, a GO line or EOF, is scanned on the masked text:
FROM/JOIN give reads_from, and INSERT/UPDATE/DELETE/MERGE/SELECT INTO
give writes_to. Alias updates resolve to their table; CTEs, variables
and temp tables are skipped.
AST cache schema 20 -> 21 so cached .sql extractions re-run.
(cherry picked from commit 70f521dba9)
Bracket names like [dbo].[Orders] come out as the label `dbo].[Orders`,
stored procedures get no reads_from edges in any dialect, and
INSERT/UPDATE/DELETE/MERGE make no edge. Guards cover aliases, CTEs,
temp tables, variables, comments, strings and body boundaries.
(cherry picked from commit a2a666c6d7)
The landed docstring listed only the get_node/get_neighbors MCP tools,
but the CLI `explain` command also routes node lookup through this
resolver (cli.py), so a substring-only near-miss now refuses there too.
Name explain in the docstring so the shared surface is accurate.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
`get_node` / `get_neighbors` share `_resolve_single_node`, which takes the
first node of the winning tier. The substring tier matches a query anywhere
inside a label, so `get_node(label="MailController")` on a codebase with no
`MailController` returned `EmailController` as if it were the exact symbol,
with no hint that the name does not exist.
The resolver now splits the two fuzzy tiers by intent. A prefix match still
resolves, because the caller started typing the real name (`email` ->
`EmailController`). A substring-only match is a search hit, not the symbol
the caller named, so it returns `No exact match for '<query>'` followed by
up to five closest names with their source location, plus a `path::symbol`
retry example. `get_neighbors` inherits the same answer through the shared
resolver, so both tools report one lookup the same way (ADR-0001).
Exact, source-exact and prefix resolution, the ambiguity refusals, the
dotted/Python and `path::symbol` forms and scoring are unchanged. The other
near-miss case in the issue (`get_node(label="edit")` with several `.edit()`
methods) already refused as Ambiguous under ADR-0001; tests cover both.
Tests: the issue's MailController repro, prefix shorthand, exact-over-substring
precedence, the five-candidate cap.
(cherry picked from commit d3875f4f1c)
A cross-repo `calls` edge took its `source_file` from the caller node
(an ObjC method node is minted at its `@interface` in the `.h`) while its
`source_location` is the call-site line, recorded in the `.m`. Pairing the
two pointed `source_file:source_location` into the wrong file — often past
the end of a short header.
The park payload now carries the call-site file alongside the line, and
`link_cross_repo_member_calls` prefers it, falling back to the caller
node's file for pre-#3152 graphs. The parked path is relativized in the
same extract-time pass that relativizes node source_files, so it never
leaks the absolute checkout path into graph.json.
Co-Authored-By: Claude Code <noreply@anthropic.com>
(cherry picked from commit 3281d42492)
The #2543 heal re-queues any manifest-stamped code file that has a
registered extractor but zero nodes in graph.json, treating it as a
poisoned stamp. A file the extractor declines by design (data JSON such
as .xcassets Contents.json, #1224/#2879) is exactly that shape — stamped,
zero nodes — yet legitimate, so the heal re-queued it on every incremental
run forever.
extract() now surfaces the declined files (skipped_sources), save_manifest
marks their row ast_empty: true (preserved across unchanged incrementals,
dropped once a file starts producing nodes), and the heal skips rows
carrying the marker. A pre-fix manifest has no marker, so such a file is
re-queued exactly once, re-stamped with the marker, and left alone after.
Co-Authored-By: Claude Code <noreply@anthropic.com>
(cherry picked from commit c930591f08)
`_merge_decl_def_classes` folds an @interface/@implementation pair into one node
only when both live in the SAME directory (the id embeds the directory path,
#1556). A class whose `.h` sits in `include/` and `.m` in `src/` therefore mints
two same-labelled nodes that never merge, so `[w render]` saw two type-def
candidates, tripped the member-call resolver's single-definition god-node guard
(`if len(type_defs) != 1: continue`), and emitted NO edge -- the same corpus
resolved fine when the pair shared a directory. Header/impl directory splits are
common layout, so this hit ordinary code.
Replace the three per-arm god-node bails with a shared `_resolve_member_target`
that unions the candidates' members and binds the selector to the sole candidate
that DEFINES it (an @implementation in `.m`/`.mm`, or a merged node carrying
`definition_file`). A header-only node is a bare declaration and does not count,
so the split pair resolves to its one real definition. Two unrelated same-named
classes that each define the selector still yield two definitions and stay
ambiguous -- the guard holds for the genuinely-ambiguous case it was meant for.
Co-Authored-By: Claude Code <noreply@anthropic.com>
(cherry picked from commit 510934c686)
The review bot flagged a real hazard in the owner-evidence bookkeeping: a
mention whose `references` edge is already in the graph returned early, so
the document never recorded that it cites the target. A later bare method
mention in the same document then lost its owner and was dropped.
`_add` now records the cited target before it decides whether the edge has
to be written, so the evidence survives an edge that is already there.
The pipeline cannot reach that state today (the resolver is handed a fresh
edge list, and the incremental context only carries contains/method/
inherits/imports edges), so this is hardening rather than a fix for a
reachable failure. The regression test drives the resolver directly with a
pre-existing class edge.
(cherry picked from commit 37efe877cf)
`resolve_markdown_mentions` linked any bare, parenthesis-less code span to
the corpus's single `_callable` carrying that label, and methods qualify.
A prose word in a code span (`redis` the server, `data`, `message`) then
became `heading --references--> SomeClass.word()` at INFERRED 0.95, on a
method that has nothing to do with the sentence. On the reporter's corpus
44 of 1,906 mention edges were words of this kind.
A bare single-name span is now resolved in three passes. Explicit evidence
(a cited path, a dotted qualifier chain, or a trailing `()`) resolves as
before. Bare mentions that do not name a method keep resolving, and their
targets are recorded per document. A bare mention that names a method is
kept only when its owner class is already cited in the same document, so
`redis` no longer links to `Cache.redis` while `successWithWarnings` next
to `HandlerResult` still does. Ambiguity, built-ins and the case-sensitive
label match are unchanged.
Tests cover the issue repro, the owner-cited case, the `()` case, a free
function, and the cross-document guard.
(cherry picked from commit 6bba91fc65)
#4655 added amp/copilot/kilo/kimi to the hand-maintained --help install
line but still omitted antigravity-windows, so it disagreed with
_print_install_usage. Derive the list from the same
[*_PLATFORM_CONFIG, "gemini", "cursor"] expression the install usage
uses so the two cannot drift, and add test_help_lists_every_install_platform
to catch any future platform that is not advertised.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The top-level help left out the `amp install` / `amp uninstall`
subcommands, and its `install --platform` list omitted copilot, kilo and
kimi, although all of them are accepted (`_PLATFORM_CONFIG`) and
documented in the README platform table. `graphify install` with a bad
platform already prints the full list, derived from `_PLATFORM_CONFIG`;
only the hand-written banner had drifted.
(cherry picked from commit 6480046271)
An ObjC category (`@interface Foo (Cat)` in `Foo+Cat.h`) keys its class
node and methods off the base stem `Foo`, so they fold into `Foo.h`'s
class (#1556). extract() turns an id minted from an absolute input path
into the repo-relative form only when it starts with its own file's stem
prefix (`.../Foo+Cat`), so these ids kept the absolute form. A category
on a class outside the corpus (`NSString (Extra)`) got a node id with
the checkout path in it, and so did the methods of a category on an
in-corpus class. graph.json then carried the path (#502), two checkouts
of one repo gave different graphs, and an incremental rebuild in a moved
or copied checkout added duplicate nodes.
The remap now also registers a category file's base-stem prefix, after
the file's own prefixes. On SDWebImage, AFNetworking and Kingfisher,
334 / 51 / 17 ids carried the checkout path; now 0, two checkouts give
identical graphs, and absolute and relative input agree. Every edge,
compared by label and file, is unchanged. Through the CLI, SDWebImage's
graph drops 125 duplicate category nodes (a leaked copy beside the clean
copy reached through its include/ symlink) and one false reference.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 11fe66a3cf)
`--transport http --host ::1` put the bare literal into the DNS-rebinding
allowlist (`::1`, `::1:8080`), but clients reaching http://[::1]:8080
send `Host: [::1]:8080`, so every request was rejected with 421 and the
server was unusable on an IPv6 loopback bind. Bracket an IPv6 literal
before building the allowlist; foreign Host headers are still rejected.
(cherry picked from commit 831b25703c)
Binary captures pick their filename with lstat so they never replace an
existing directory entry (#4393). Text captures (webpages, tweets, arXiv)
still used Path.exists(), which follows symlinks: a dangling symlink named
like the capture reported "free" and write_text() followed it out of the
corpus directory. The loop also stopped at 999 numbered names and then
overwrote the last one. Share the lstat-based selection between both
paths.
(cherry picked from commit 3e483ca077)
The claude and gemini hooks (since #522/#1987) double-quote a resolved graphify path containing spaces, but _install_codex_hook embedded the resolver output raw. On a checkout path like 'D:/Project V2/.venv/Scripts/graphify.EXE', every shell that runs the hook (Git Bash on Windows to cmd.exe to POSIX sh) word-splits the command at the first space: the guard can never fire, and #2165's subcommand-parsing test splits at the same space and fails.
Mirror _claude_pretooluse_hooks/_gemini_hook with the same two-line guard: bare-name (project-scoped, #3129) installs are unaffected, as is any absolute path without spaces. test_codex_hook_command_is_a_real_cli_subcommand now parses hook commands with shlex.split instead of naive .split(), and a new regression test installs all three hosts under a spaced resolver output, proves argv[0] survives intact, and re-verifies uninstall recognition of the quoted entry.
(cherry picked from commit 87e997d3a0)
#4634 suppressed the module-scope node for a require binding so it no
longer shadows the required class. But its guard used
_find_require_call(value), which matches ANY identifier(...) call, so it
also dropped real `const x = factory()` nodes (e.g. express(), NextAuth(),
wrapper()) and regressed #2604/#2552 (10 test failures). Add
_js_value_is_require, which name-checks the callee is literally `require`
the same way _require_imports_js already does, and gate the binding on it.
Bump _AST_CACHE_SCHEMA 20 -> 21: the require-binding change alters
extraction output (drops the shadow node, redirects the calls edge).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A \const X = require('./m')\ binding is an import, not a module-scope data
binding. The node it produced in the importing file shadowed the class the
required file exports, so \
ew X()\ bound to the local node and the
required class got no incoming edge.
(cherry picked from commit 66bb69114a)
Covers the issue's own repro, a delegating wrapper that no longer
self loops and an unrelated same named sibling that no longer steals
the edge, plus three controls: a bare unqualified call still
resolves by name, this dot calls still bind through the caller's own
class, and a TypeScript typed receiver still resolves through its
declared type. A fourth control checks that a factory object's own
sibling methods still call each other, the one legitimate case that
keeps resolving locally instead of deferring.
(cherry picked from commit 49c645d2ff)
A member call whose receiver is neither this nor self reached the
per file bare name map, the same way an unqualified call does. A
delegating wrapper like history.appendFrame(entry) inside a function
also named appendFrame then bound to itself, an EXTRACTED self loop,
and the same map could just as easily bind it to any other same
named function in the file instead of the receiver's own method.
Every such member call now defers, exactly like it already does for
Python, Java and C#. The deferred call still reaches the receiver
typed resolver, which binds it correctly when the receiver's type is
known, and otherwise leaves it unresolved rather than guessing. A
plain call with no receiver, and this dot calls resolved through the
caller's own class, are both untouched.
One case keeps resolving locally rather than through that resolver,
since the resolver only sees whole file state and could still bind
it to a same named member on an unrelated object: a method assigned
onto a local factory object calling a sibling through that very
object, such as store.get() from inside store.set. The receiver name
there reconstructs the caller's own owner id exactly, so it is read
straight off the owner rather than guessed.
(cherry picked from commit 24b1f28fd4)
_parse_pom collected every <dependency> in the file with
".//dependencies/dependency", including <dependencyManagement> entries,
which only pin versions, and the <dependencies> of build plugins. Each
became a depends_on edge, so a parent POM that manages its modules'
versions depended on every module.
Read <dependencies> under <project> and under <profiles><profile>, where
Maven takes a project's dependencies from.
(cherry picked from commit 534177d1ea)
README, translations, support docs, and the GitHub templates were handing out a coded invite. It still joins the same server. Use the vanity URL.
(cherry picked from commit 7331198fc4)
The partial-read coverage check (#3987) adds a python block between the
Step B3 merge and Part C. Two skill-path tests that landed separately in
the 0.9.9x skill work hard-code the block count in that range (3) and
then execute every block in it. Bump the expected count to 4; the
behavioral assertions (token sums, merged node sets, stale-chunk
clearing) still hold because B3.5 only writes an empty
.graphify_partial.txt on the all-cached and codex-merge paths.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Land #3987 (@nothariharan, reported by @Ryanm218): on the skill
extraction path a long file a subagent only skimmed was cached as
complete and its unread tail silently dropped on the next replace-on-
re-extract. The chunk coverage self-report is now checked against each
file's real line count; a short file is cached partial and left
unstamped so --update retries it, reusing the library's existing
partial_source_files contract (#933, #1950).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Read tool returns at most 2000 lines per call and reports no error when it
truncates, so a subagent that read only the head of a long file still writes a
well-formed chunk. Step B3 treated "chunk exists and parses" as success, and
because build_merge replaces a re-extracted file's nodes, the unread tail was
not merely missed: it was deleted, the run reported success, and the manifest
recorded the file as done so no later --update retried it. On a 4165-line file
that left 103 nodes covering 31% of the document, with 90 of 131 entries gone
and the manifest treating it as finished.
Three changes, in the fragments, so all thirteen disk-dispatching hosts pick
them up from one edit:
The extraction spec (verbose and compact) now states the cap and forbids
emitting nodes for a partly-read file, and asks the subagent to report a
top-level "coverage" map of file -> lines read. A short count is recoverable; a
missing entry is indistinguishable from a complete read, which is why the
machine check needs the prompt to ask.
Step B1 bounds a chunk by total lines as well as file count (20-25 files AND
~6000 lines), and gives any file past 2000 lines a chunk to itself. File count
alone was only half the mechanism: a subagent holding 25 files also dilutes
attention, which is how a 590-line plan came back with zero nodes.
Step B3.5 compares each reported coverage against the file's real length, names
what was short, and hands the list to the two places that decide whether a file
is ever looked at again: save_semantic_cache(partial_source_files=...) stamps
those entries partial, which load_cached treats as a miss, and
_stamped_manifest_files(partial_source_files=...) leaves them unstamped so
detect_incremental re-queues them. Both are the existing contract the library
path already uses for a truncated LLM response (#933, #1950), so the skill and
library paths now recover the same way. Nodes already found for those files are
kept; the next run's replace-on-re-extract swaps them for a complete set.
The check is fail-safe in the direction that loses nothing: no coverage reported
means no file is flagged, and a flagged file is re-read rather than trusted.
Codex collects subagent results in memory with no per-chunk files on disk, so it
gets the prompt-layer fix but not the Step B3.5 gate; its inline extraction has
no chunk file to compare against.
Tests execute the emitted python instead of matching its source (the #2126/#2641
convention), including a round trip through the real semantic cache: the
partly-read file must come back as uncached, while the fully-read file still
hits. Reverting the fix makes that round trip return an empty uncached list -
the file served from cache forever, which is the bug.
Fixes#2225
(cherry picked from commit 0b6a1868af)
The ambiguous-name refusal added in #4517 routed through _resolve_single_node,
whose winning tier can be a prefix or substring match, so a bare search keyword
that merely substring-matched several labels was wrongly refused: `query "client"`
returned EXIT 3 because it is a substring of `AlphaClient`/`BetaClient`. That
narrowed query's keyword-search semantics.
Adds _query_name_refusal, which refuses only when the term EXACTLY names several
nodes (the source_exact/exact tiers) -- #4511's `.nonce()` case -- and otherwise
falls through to scoring. The `query` and `--seed` paths use it; explain/path/
get_node keep _resolve_single_node, whose prefix/substring tiers are correct for a
direct node lookup. Adds a regression test that a substring keyword still searches.
Co-Authored-By: shivampal46011 <shivampal46011@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
query guessed a node when the same name existed in several files, and an ambiguous --seed still exited 0. explain and path already refuse that lookup. query now returns the same candidate list and exits 3, and a --seed that matches nothing exits 4.
#4561 changes extract_markdown's output (vault-relative wikilinks now resolve to
the real page), and markdown extraction is AST-cached, so a .md file cached under
schema 19 would serve stale unresolved links on re-extract until the file itself
changed. Bump the schema so the first build after upgrading re-resolves markdown
links. (The other batch-3 fixes -- cli, serve, PDF-warning, shebang detection --
do not change cached per-file AST output.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
_shebang_interpreter returned the interpreter exactly as written, and both
detect's _SHEBANG_CODE_INTERPRETERS and extract's _SHEBANG_DISPATCH list
only the generic names. An extensionless CLI starting with
#!/usr/bin/env python3.12, #!/usr/bin/python2.7 or #!/usr/bin/lua5.4 was
classified as None and silently left out of the scan. Map a version
suffix on a known interpreter back to that interpreter (python3.12 ->
python3, lua5.4 -> lua) in the shared parser, so detection and extractor
dispatch both see it. Unknown names are left as they are.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
extract_pdf_text wrapped the whole PdfReader loop in except Exception:
return "", so a malformed or encrypted PDF looked exactly like a PDF with
no text and dropped out of the graph with no trace. Keep returning ""
(callers rely on it), but print a warning naming the file and the error
first. Warn once per path, since detect() and the semantic pass both read
the same file.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nothariharan