- Route .h files through extract_cpp (was extract_c), fixing missing method nodes in C++ headers
- Extend _get_cpp_func_name to handle field_identifier, destructor_name, operator_name
- Add CPP-specific field_declaration branch in _extract_generic to emit class method/field nodes
- Partition dedup Pass 1 by source_file: only union same-label nodes within the same file;
cross-file matches fall through to Pass 2 fuzzy, preventing generic-label god nodes
- Add _resolve_c_include_path: resolve quoted #include paths to real files on disk so
target node IDs match what _extract_generic creates, fixing dangling include edges
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Common terms like 'error'/'exception' were stealing BFS seed slots from
rare identifiers like 'FooBarService', burning the token budget on noise.
- _compute_idf: weights query terms by inverse document frequency, cached
on G.graph so cost is paid once per graph load not per query
- _score_nodes: multiplies each tier bonus by IDF weight
- _pick_seeds: replaces fixed top-3 with gap-ratio selection — stops adding
seeds when score drops below 20% of the top match
- _subgraph_to_text: truncation hint now tells Claude to narrow with
context_filter or use get_node instead of just saying 'truncated'
Fixes#897
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Clean merge — fixes real token-bloat regression where GRAPH_REPORT.md was loaded on every question. All 10 install surfaces now route to graphify query first. Also fixes hidden bug where re-running graphify install after upgrade didn't refresh stale hook text.
The current install writes "ALWAYS read graphify-out/GRAPH_REPORT.md
before reading any source files, running grep/glob searches, or
answering codebase questions" into CLAUDE.md and equivalents, plus a
PreToolUse hook with the same instruction. On real corpora that report
is 47-91K characters, so Claude Code sessions pay roughly 12-25K tokens
of context up front for every search-able question. Three users on #580
reported this making token usage worse than no install at all.
Reproduced on a 1500-file Go monorepo: a "where is X defined" question
takes 5 tool calls and 34k agent tokens with stock graphify, 4 calls
and 30k tokens with no install, and 1 call and 30k tokens after this
patch. Stock graphify's Read of GRAPH_REPORT.md hit Claude Code's 25k
token cap and failed entirely, then recovered via a partial read plus
graphify explain.
Demote GRAPH_REPORT.md to a fallback for broad architecture review and
route first action to the existing scoped commands: graphify query,
path, explain. The 2k-budget BFS subgraph already exists in serve.py;
the install just wasn't pointing at it.
Updated across all ten install surfaces: _SETTINGS_HOOK,
_CLAUDE_MD_SECTION, _AGENTS_MD_SECTION, _GEMINI_MD_SECTION,
_GEMINI_HOOK, _VSCODE_INSTRUCTIONS_SECTION, _ANTIGRAVITY_RULES,
_KIRO_STEERING, _CURSOR_RULE, _OPENCODE_PLUGIN_JS. Plus the matching
sentence in README.md, which also fixes an inaccuracy about Codex
hooks (Codex's installed hook is intentionally a no-op because Codex
rejects additionalContext, so the guidance there comes from AGENTS.md,
not the hook).
Five installers (claude, agents, vscode, gemini, kiro, cursor) were
also writing their section only when no marker was present, so users
who installed pre-fix kept the old "ALWAYS read" text after upgrading.
Added _replace_or_append_section helper that updates in place when the
graphify marker is found. claude_install also no longer returns before
re-running _install_claude_hook, so stale settings.json hook payloads
get refreshed on upgrade.
Tests:
- tests/test_install_strings.py (3): every install constant still
mentions `graphify query` and matches no banned report-first regex.
- tests/test_install_upgrade.py (7): seeds each platform's instruction
file with pre-fix text, runs install, asserts the on-disk file
reflects the new policy.
- test_claude_md.py idempotency tests still pass.
Fixes#580.
When `root` has at least one direct symlinked child, default to following
symlinks instead of silently dropping their contents. This makes "fake
working dir" patterns (a folder full of symlinks pointing at scattered
source dirs) work transparently — the caller no longer has to know to
pass `follow_symlinks=True`.
Concrete motivating shape:
~/projects/research-corpus/
├── papers -> /external/drive/papers
├── notes -> /external/drive/notes
└── code -> /external/drive/code
Before: `--update` invoked `detect_incremental(root)` without passing
`follow_symlinks=True`, so the scan saw the small set of files actually
inside the root and missed everything reachable only through a symlink.
Result: every legitimate new file was missed, and the manifest paths
were marked as deleted.
After: when `follow_symlinks` is left at its default `None`, `detect()`
runs `_auto_follow_symlinks(root)` (one cheap `iterdir()` + `is_symlink()`
loop) and follows symlinks if any direct child is one. Behaviour is
unchanged for ordinary scans (no direct symlinks → False, as before).
Override is always possible by passing an explicit `follow_symlinks=True`
or `follow_symlinks=False`; existing tests confirming the explicit
behaviour continue to pass unchanged.
Backwards compatibility:
- Type annotation: `bool` -> `bool | None`. Callers passing `True`/`False`
continue to work identically. Callers passing nothing get the new
auto-detect.
- No new dependencies.
- Cheap: one `iterdir()` call once per detect() invocation.
Tests:
- 3 new tests in tests/test_detect.py:
- test_detect_auto_detects_direct_symlink_child
- test_detect_default_does_not_follow_when_no_symlinks
- test_detect_explicit_false_overrides_auto_detect
- Full tests/test_detect.py + tests/test_incremental.py: 49/49 pass.
In Python+TypeScript monorepos, the call and import resolvers match by
label across language boundaries (AuthError -> Member), producing false
positives that dominate 'Surprising Connections' due to cross-dir (+2)
and cross-community (+1) bonuses.
Expand the existing calls guard to also cover uses edges, and zero the
cross-dir and cross-community bonuses for these pairs — not just conf_bonus.
Leaves semantically_similar_to, EXTRACTED, and AMBIGUOUS edges unaffected.
5 new tests: calls suppressed, uses suppressed, semantically_similar_to
preserved, same-language INFERRED preserved, cross-language EXTRACTED preserved.
Alembic/Flask-Migrate revisions, Django migrations, and protobuf/OpenAPI
generated files produce hundreds of degree-1 rationale nodes labeled as
'possible documentation gaps'. Their module docstrings are revision
annotations or boilerplate, not architectural rationale.
- Add _is_autogenerated_python() in extract.py detecting Alembic, Django
migrations, and generic DO-NOT-EDIT markers; skip module docstring only
- Function/class docstrings inside those files still extracted as normal
- report.py: exclude file_type=rationale nodes from isolated-node gaps
section — rationale nodes are degree-1 by construction; flagging them
as missing edges was always wrong
- 5 new tests covering Alembic, Django, protobuf, false-positive guard,
and function-docstring passthrough
Move all intermediate data files and temp step scripts from the
project root into graphify-out/ to match the bash skill behaviour.
Also fix the > redirect on detect.json which wrote UTF-16LE in
PowerShell 5.1 - now uses Out-File -Encoding utf8 consistently.
Files moved: .graphify_detect.json, .graphify_ast.json,
.graphify_extract.json, .graphify_semantic.json,
.graphify_analysis.json, .graphify_labels.json,
.graphify_incremental.json, .graphify_old.json,
all .graphify_step_*.py temp scripts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
antigravity install on Windows now installs skill-windows.md instead
of skill.md by redirecting to a new antigravity-windows platform config
at install time via sys.platform check.
Python detection in Find-GraphifyPython now uses uv tool dir (respects
UV_TOOL_DIR) and pipx environment --value PIPX_LOCAL_VENVS (respects
PIPX_HOME) instead of guessing from the shim location. The graphify.exe
shim and python.exe live in different directories for both uv and pipx
on Windows so the previous Get-Command approach was wrong.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
graphify.exe and python.exe are co-located in the same Scripts\
directory for both uv tool and pipx installs. Use Get-Command graphify
to find that directory and pick up the correct Python instead of
falling through to bare python (which is a different interpreter
that doesn't have graphify).
Also replace all bare `python .graphify_*` invocations throughout
the skill with `& (Get-Content graphify-out\.graphify_python)` so
every step uses the saved interpreter, and fix the bash $(cat ...)
syntax to PowerShell & (Get-Content ...).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
#873: Remove blanket dot-prefix exclusion from detect.py and
extract.py collect_files(). Add framework caches (.next, .nuxt,
.turbo, .angular, .idea, .cache, .parcel-cache, .svelte-kit,
.terraform, .serverless, .graphify) to _SKIP_DIRS so they stay
blocked. Meaningful dot dirs (.github, .claude, etc.) are now
indexed.
#874: Add _maybe_reload() with mtime+size stat key and threading.Lock
to serve.py. call_tool and read_resource call _maybe_reload() on
every request; the graph reloads automatically when graph.json changes
without restarting the MCP server.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add .sh, .bash, .json to CODE_EXTENSIONS in detect.py so files reach extractor
- Fix test_detect_incremental manifest path collision with new .json extension
- Update test_watch to reflect .json/.sh are now watched extensions
- B-1: only emit source imports for paths that exist on disk
- J-1: replace stat()+read() with bounded read to eliminate TOCTOU
- J-3: move pair_count cap inside loop so it is honoured exactly
- J-4: namespace $ref/extends refs with "ref_" prefix to prevent ID collision
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
In a linked worktree .git is a file not a directory, so root/.git/hooks
fails with NotADirectoryError. Fall back to git rev-parse --path-format=absolute
--git-path hooks which resolves the correct hooks dir from any worktree.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- skill-windows.md: add graphify-out/ prefix to 5 paths in Step B3 that
were being read/written without prefix (lines 244,245,373,387,388,407,413)
causing semantic cache to always appear empty on Windows (#862)
- detect.py: gate converted office/Google Workspace sidecar paths through
_is_ignored() before appending to files list so .graphifyignore rules
apply to graphify-out/converted/ sidecars (#861)
- detect.py: split manifest entries from {mtime,hash} to
{mtime,ast_hash,semantic_hash}; save_manifest() accepts kind="ast"|
"semantic"|"both"; detect_incremental() accepts kind= and checks the
right hash field so graphify update (kind=ast) no longer prevents
graphify extract from re-running semantic extraction (#857)
- watch.py: all three save_manifest() calls use kind="ast" (AST-only rebuild)
- __main__.py: both extract save_manifest() calls use kind="both"
- backwards compatible: legacy {mtime,hash} entries treated as ast_hash only
775 tests passing
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Several host agents (Claude Code in particular) misread the current
phrasing and prompt the user for `ANTHROPIC_API_KEY` when no Gemini
key is set, instead of falling through to subagent dispatch. The
Gemini tip stays as-is; this just adds an explicit one-line callout
that no other provider keys are consulted, with `ANTHROPIC_API_KEY`
called out by name since that is the specific misread we hit in the
wild.
No behavior change — pure documentation.
The rebuild lock file accumulated concatenated PIDs across post-commit
rebuilds without a separator, and was never removed when the rebuild
finished. Two practical consequences for users:
1. Downstream tooling that polls for `.rebuild.lock` to disappear before
doing post-rebuild work (publish scripts copying graph.html to a web
root, etc.) blocked forever / until its own timeout.
2. The accumulated digit string could not be parsed by humans or tooling
to find the owning PID.
The `_rebuild_lock` context manager now:
- Opens the lock file with `a+` so a non-acquiring caller does not
truncate the existing holder's PID.
- After flock acquisition, truncates and writes a single `<pid>\n` line
so external readers can `kill -0 $(cat .rebuild.lock)` to check
liveness.
- Unlinks the lock file in the finally block (only when *we* held the
lock), restoring the "signal-by-absence" convention users rely on.
Four regression tests added under `tests/test_watch.py` covering the
PID-with-newline payload, post-release unlink, no-accumulation across
sequential acquisitions, and the non-blocking-caller-does-not-clobber
invariant.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Astro files have a `---...---` TypeScript frontmatter block at the top
containing nearly all imports, followed by an HTML-with-expressions
template body, and optionally `<script>` blocks for client-side JS.
Feeding the whole file to the JS tree-sitter parser produces a top-level
ERROR node because the template isn't valid JS, so `import_statement`
nodes are never reached.
Mirrors `extract_svelte` (#713) — same regex-rescue strategy:
- Dynamic `import('...')` pass over the full source.
- Static `import X from '...'` pass over the frontmatter block plus any
`<script>` blocks.
Both static and dynamic imports honor tsconfig.json path aliases and
resolve relative paths through `_resolve_js_module_path` so node ids
match the targets that `_extract_generic` produces for sibling files.
Wired in:
- `detect.py:27` — `.astro` added to `CODE_EXTENSIONS`.
- `extract.py:5492` — `".astro": extract_astro` in the dispatch table.
Tests in `tests/test_astro_extraction.py` cover: extension registration,
frontmatter static imports, dynamic import in frontmatter, client-side
`<script>` block imports, frontmatter-less pages (must not crash), and
tsconfig.json path-alias resolution.
- skill.md + skill-windows.md: add encoding="utf-8" to all read_text()/write_text()
calls and ensure_ascii=False to json.dumps — bare calls defaulted to system
codepage on Chinese-locale Windows, mojibaking non-ASCII content (#832)
- skill.md + skill-windows.md: prefer uv tool install --upgrade graphifyy over
pip in the Step 1 install fallback — pip installs to the wrong env when
graphify was installed via uv tool (#831)
- serve.py + __main__.py: replace flat substring scoring in _score_nodes with
three-tier precedence (exact 1000 / prefix 100 / substring 1); _find_node
returns results ordered exact→prefix→substring; both path CLI and MCP now
emit a clear error when src and tgt resolve to the same node (#828)
- cache.py: normalize path key via .as_posix().lower() in file_hash so Windows
junction/case variants hash identically; mirror abs-path normalization from
save_semantic_cache into check_semantic_cache so relative source_file paths
resolve the same way on both sides (#826)
- __main__.py: add /graphify skill trigger line to _AGENTS_MD_SECTION — affects
all 7 AGENTS.md platforms (OpenCode, Codex, Aider, Trae, Hermes, Claw, Droid)
so typing /graphify actually invokes the skill tool (#827)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add a pre-cluster topology comparison fast path in update rebuilds so unchanged graphs short-circuit before clustering and report generation, preventing residual run-to-run community-count drift.
Co-authored-by: Cursor <cursoragent@cursor.com>
Use safe JSON serialization fallbacks for deterministic sort keys in clustering and graph canonicalization, and skip invalid community IDs with a stderr warning instead of raising during update rebuilds.
Co-authored-by: Cursor <cursoragent@cursor.com>
Make `graphify update` idempotent by skipping output rewrites when graph/report content is unchanged, add `update --no-cluster`, and preserve community IDs across runs via overlap-based remapping with deterministic partition inputs.
Co-authored-by: Cursor <cursoragent@cursor.com>
- Universal -h/--help/-? guard after cmd dispatch: any help flag anywhere in
argv stops execution and prints "Run 'graphify --help'" instead of triggering
the subcommand — cursor/kiro/gemini install --help no longer silently installs;
benchmark --help no longer crashes with FileNotFoundError (#821)
- --version / -v / version subcommand: print graphify {__version__} and exit (#818)
- GRAPHIFY_OLLAMA_NUM_CTX=<invalid> now falls through to auto-derived num_ctx
instead of hardcoding 131072 (the cap that causes OOM on constrained VRAM);
pinned num_ctx < estimated input now triggers an explicit truncation warning
with a suggested --token-budget correction (#820)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- extract/_make_id + build/_normalize_id: use NFKC normalization and casefold
so composed/decomposed Unicode forms produce the same ID; collapse consecutive
underscores; both functions are now byte-for-byte equivalent (#811)
- dedup: use explicit key-presence check instead of `or` for source/from
fallback; pop stale from/to keys so they don't leak into graph.json attrs (#803)
- skill --update: use build_merge() to avoid NetworkX round-trip direction flip;
fix dict merge ordering so explicit source/target win; pull hyperedges from
G.graph (merged) not new_extraction only (#801)
- skill subagents: inject absolute CHUNK_PATH so Write tool doesn't lose chunk
files to undefined cwd (#808)
- __main__: skip skill version check during hook-check (runs on every editor
tool use, must be silent); move warning to stderr
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
#796: add edge_data()/edge_datas() helpers in build.py that tolerate
MultiGraph/MultiDiGraph; replace all G.edges[u,v] 2-tuple call sites in
__main__.py, serve.py, wiki.py, export.py, analyze.py, benchmark.py;
fix same pattern in 10 skill file inline heredocs
#795: all 12 skill files now short-circuit on /graphify --help or -h
and print the Usage block without running any pipeline steps
#792 (hollow response): add _response_is_hollow() predicate in llm.py;
when Ollama (or any backend) returns empty/null/whitespace content or a
parsed result with no nodes/edges, rewrite finish_reason="length" so
_extract_with_adaptive_retry bisects the chunk instead of silently
dropping it; applied to _call_openai_compat, _call_claude, _call_bedrock
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
#791: per-repo fcntl flock in _rebuild_code prevents concurrent hook rebuilds
from exhausting memory; changed_paths wired through so only modified files are
re-extracted; stale nodes evicted on deletion; SIGALRM watchdog with
GRAPHIFY_REBUILD_TIMEOUT; Darwin-aware RLIMIT_DATA memory cap
#792: remove hard 8-worker cap (GRAPHIFY_MAX_WORKERS env var); add
--max-workers, --token-budget, --max-concurrency, --api-timeout CLI flags
to graphify extract; fix ollama API key gate for loopback URLs; explicit
timeout on OpenAI client (GRAPHIFY_API_TIMEOUT, default 600s); per-chunk
progress prints during extraction
#453 + #785: rename .agent -> .agents throughout antigravity install/uninstall;
add trigger:always_on YAML frontmatter to _ANTIGRAVITY_RULES so Antigravity
recognises the rules file
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
tree-sitter-pascal is not on PyPI so extract_pascal() fell back to an
empty error result for all users. _extract_pascal_regex() now handles
unit/program/library headers, uses clauses, class/interface declarations
with inheritance, forward method decls, qualified impl headers, balanced
begin/end body extraction, and intra-file calls edges. All 15 previously
skipped pascal_required tests now run unconditionally and pass.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(extract): add Pascal/Delphi language support
Adds full AST extraction for Pascal and Delphi source files using
tree-sitter-pascal (https://github.com/Isopod/tree-sitter-pascal).
Supported file extensions: .pas, .pp, .dpr, .dpk, .inc
Extracted nodes:
- File node (the .pas file itself)
- unit / program / library declarations
- class, interface, and helper type declarations
- procedure and function implementations
Extracted edges:
- file --contains--> module
- module --imports--> dependency (via uses clause, resolved to path-based IDs)
- class --inherits--> base class / interface
- class/module --contains/method--> procedure or function
- procedure --calls--> procedure (in-file call resolution)
Key design: uses clause targets are resolved to path-based node IDs by
scanning all Pascal files under the project root (_pascal_project_root +
_pascal_resolve_unit helpers). This avoids dangling import edges that
result from resolving bare unit names like "SysUtils" to IDs that never
match any file node.
Bare procedure calls (e.g. `Reset;` without parentheses) are detected
by inspecting statement nodes whose sole named child is an identifier,
in addition to the standard exprCall nodes used for calls with arguments.
Requires: pip install tree-sitter-pascal
(https://github.com/Isopod/tree-sitter-pascal)
If not installed, extract_pascal returns {"nodes":[], "edges":[], "error": ...}
so the rest of the pipeline is unaffected.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(extract): add Lazarus .lfm and .lpk file support
Adds two new extractors for Lazarus IDE-specific file formats:
extract_lazarus_form() — .lfm (Lazarus Form files)
.lfm files are text-based UI component trees. The extractor parses
`object Name: TClassName ... end` blocks to build a containment graph
of form components, and captures `OnXxx = HandlerName` event bindings
as `references` edges (context: "event") linking each component to
its handler procedure.
extract_lazarus_package() — .lpk (Lazarus Package files)
.lpk files are XML package definitions. The extractor reads the
package name, required package dependencies (→ imports edges), and
listed unit files (→ contains edges). Unit names are resolved to
path-based node IDs via _pascal_resolve_unit so they connect to the
same nodes produced by extract_pascal on .pas files.
Both extensions added to CODE_EXTENSIONS in detect.py and to _DISPATCH.
13 new tests in test_pascal.py cover both extractors.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(pascal): fix dangling inherits-edge targets and capture all base classes
The declType/typeref handler built inherits edge targets with
_make_id(_read(child)) — just the bare class name. But class nodes
use _make_id(stem, type_name), so targets never matched, making the
entire class hierarchy invisible in the graph.
Add _pascal_class_stem_cache and _pascal_resolve_class(): strips the
conventional T/I prefix, locates the defining file by stem lookup
(same cache mechanism as _pascal_resolve_unit), and returns the
correct _make_id(file_stem, class_name) ID. RTL/unresolvable bases
(e.g. TObject) fall back to _make_id(bare_name) with an explicit
stub node, following the same pattern as the Python extractor.
Also remove the `break` that stopped after the first typeref, so
all parents are captured (e.g. class(TBase, IInterface)).
Extend test_pascal_no_dangling_edges to also assert that within-file
edge targets (contains, method, inherits, calls) resolve to real nodes.
* feat(extract): add Delphi .dfm form file support
Adds extract_delphi_form() for Delphi Form files (.dfm), which use the
same `object Name: TClassName ... end` text syntax as Lazarus .lfm files.
Binary .dfm files (FF 0A magic header) are skipped gracefully with an
informative error message so the pipeline is unaffected. Text .dfm files
are parsed identically to .lfm: component containment (`contains` edges)
and event handler references (`references`, context "event").
Adds .dfm to _DISPATCH and CODE_EXTENSIONS.
10 new tests in test_pascal.py, including a regression test for the
binary-format detection.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Add .lpr extension and fix removesuffix in Pascal extractor
Address review feedback from @safishamsi:
- Add .lpr (Lazarus program file, identical syntax to .dpr) to _DISPATCH
in extract.py and CODE_EXTENSIONS in detect.py so Lazarus project entry
points are indexed. Completes the promised Lazarus IDE support.
- Replace rstrip("()") with removesuffix("()") in the call-resolution
dict comprehension for precise suffix removal (rstrip strips individual
characters, not the literal string "()").
- Add .lpr assertions to test_pascal_dispatch_registered and
test_pascal_detect_extensions_registered.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Simeon Bodurov <simeon.bodurov@speedy.bg>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Three independent Windows compatibility fixes shipped together because they
all surface during the same first /graphify run on Windows.
graphify/benchmark.py
print_benchmark() unconditionally printed U+2500 (box-drawing) and U+2192
(rightwards arrow), which UnicodeEncodeError'd on stdouts that can't encode
them — most notably the legacy Windows console at cp1252. New _safe()
helper falls back to ASCII when the active stdout encoding can't carry the
glyph; _hr() uses it. Two regression tests cover both paths and prove
print_benchmark survives a cp1252-strict stream.
graphify/extract.py
ProcessPoolExecutor on Windows uses spawn, so worker subprocesses
re-import the calling __main__. When the caller is `python -c "..."` or a
script without an `if __name__ == "__main__":` guard, the workers
recursively spawn themselves and the pool dies. The user-visible failure
was a 290-line traceback ending in BrokenProcessPool, hiding the actual
cause. _extract_parallel now catches BrokenProcessPool, prints a one-line
warning that names the __main__-guard idiom, and returns False so the
public extract() routes to the existing _extract_sequential fallback. Two
tests cover the parallel-returns-False contract and the sequential
fallback wiring.
graphify/skill-windows.md
Every `python -c "..."` block (30 in total) is replaced with a
Write+run+delete pattern using PowerShell's literal here-string @'...'@.
The old form was a quote-escaping minefield: any double-quote inside the
Python source had to be backslash-escaped for the shell, and PowerShell's
parser ate them inconsistently — failing on f-strings like
`f'AST: {len(result["nodes"])} nodes'`. The new form passes Python source
to disk literally, so what the model writes is what Python sees. The AST
step's script template now includes an explicit `if __name__ == "__main__":`
guard so multi-core extraction works even before the runtime fallback above
kicks in. All 31 resulting heredoc blocks parse cleanly under
`ast.parse`.
Co-authored-by: Nauman Hameed <Nauman.Hameed@enghouse.com>
Adaptive retry only recovered from `finish_reason="length"` (output
truncation). It did not handle the other shape of overflow: the API
rejecting the prompt outright with a 400 because the input plus
`max_completion_tokens` doesn't fit in the model's context window.
This shows up immediately on local OpenAI-compatible servers (LM
Studio, llama.cpp, vLLM) where the default context is small (4K-32K)
and a 60K-token chunk packed for cloud Kimi/Claude blows past it.
Without retry the whole chunk fails with no output, even though the
two halves would each fit cleanly.
Catch a heuristic set of context-overflow exception messages,
classify them as the same kind of recoverable failure as
`finish_reason="length"`, and split-recurse on the same path. Single-
file overflow returns an empty fragment so the rest of the corpus
keeps running. Unrelated errors (rate limit, auth, etc.) still
propagate.
Tested with qwen3.5-9b on LM Studio (32K ctx) against a 215-file
corpus where chunks 4-12 of 12 previously failed; with this change
the overflowing chunks self-heal by splitting in half.
Safi