* feat(export): GRAPHIFY_VIZ_NODE_LIMIT env var to override MAX_NODES_FOR_VIZ
Adds opt-in env var so users with large graphs (>5000 nodes) can either
raise the HTML viz limit or disable viz entirely (set to 0) without
patching the package.
- New _viz_node_limit() helper reads GRAPHIFY_VIZ_NODE_LIMIT, falls back
to MAX_NODES_FOR_VIZ on unset/empty/non-integer values.
- to_html() now uses the helper; ValueError message references both
--no-viz and GRAPHIFY_VIZ_NODE_LIMIT for discoverability.
- 7 new tests in test_export.py covering default, raise/lower, zero,
invalid, empty, and end-to-end to_html behavior.
Addresses option 2 in #541 (env var as default behavior, zero-config
for most users, tunable for large repos.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EOF
)
* feat(cluster-only): wire --no-viz flag and tolerate ValueError mid-write
Before this change, `graphify cluster-only <path>` called to_html()
unconditionally, so on graphs >5000 nodes the ValueError fired AFTER
graph.json was written but BEFORE the "Done — ..." print, leaving a
stale graph.html on disk and no clear signal in the CLI output.
This mirrors the watch.py rebuild pattern (already merged): try/except
around to_html, delete the stale graph.html when skipping, and emit a
"Done" line that accurately reflects which artifacts landed.
- Parse "--no-viz" from sys.argv (documented as such in skill.md but
previously never wired into the Python CLI).
- Wrap to_html() in try/except so core outputs (graph.json +
GRAPH_REPORT.md) always land.
- Remove stale graph.html when viz is skipped or fails — avoids the
silent desync where graph.html lags behind graph.json by days.
- Update help text under `cluster-only` to advertise --no-viz.
Addresses option 1 in #541 (--no-viz flag, simplest opt-in).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Tayfun Sert,MBA <tayfunsert@hotmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
On Windows, to_wiki crashes with OSError [Errno 22] when a node label
contains any of < > : " / \ | ? * (Windows reserved chars in
filenames). Repro: an Obsidian note titled "Czy Wii jeszcze ok?.md"
becomes a community label, then write_text fails because Windows
rejects the question mark in the path.
The previous _safe_filename only handled three characters (/ space :)
which is enough on Linux/macOS but not on Windows. Extended the
substitution to cover all Windows-reserved chars plus trailing dots
and spaces (also reserved on Windows), and added a 200-char length
cap to stay well under MAX_PATH-segment limits.
Falls back to 'unnamed' if sanitization produces an empty string,
so we never try to create a file named "" or ".".
Tested on a 1773-file vault with mixed Polish/English content where
~12 community labels and several god node labels contained reserved
characters. Generates 763 wiki articles cleanly.
* Fix#563: prevent rationale-node leakage and preserve calls direction
Two AST-extractor bugs were inflating god-node centrality:
1. Rationale leakage in cross-file resolvers
_resolve_cross_file_imports indexed any node whose label didn't end in
')' or '.py' as an importable entity, so rationale nodes (whose labels
are docstring text) ended up in both stem_to_entities and local_classes.
Result: phantom '<file>_rationale_N --uses--> ImportedThing' edges for
every imported entity in every file with docstrings.
The same leak existed in cross-file call resolution via
global_label_to_nid, where rationale labels could collide with callee
names.
Fix: skip nodes with file_type == 'rationale' in all three index loops.
2. Calls / rationale_for direction lost at export
to_json() called nx.json_graph.node_link_data() on a NetworkX undirected
Graph, which canonicalizes endpoint order. The build path already
stashes _src and _tgt on every edge for exactly this case (with a
comment to that effect at build.py:100), but the exporter never
consulted them, so 'calls' and 'rationale_for' edges were emitted with
arbitrary direction.
Fix: in to_json(), restore link['source'] / link['target'] from
_src / _tgt, then pop those internal fields before writing graph.json.
Repro on a minimal 3-file project (queue.py + contact_form.py + a test
file with multi-paragraph docstrings) showed:
Before fix: 31 edges, 12 incident on SQLiteQueue, 7 phantom uses-edges
from rationale nodes, 1 inverted calls edge.
After fix: 24 edges, 5 incident on SQLiteQueue, 0 phantom edges,
all calls edges go caller -> callee.
Adds tests/test_issue_563.py with 5 regression tests covering both bugs
plus a direction-pin on the existing sample_calls.py fixture. Full suite
passes (446/446).
* Address Copilot review on #576
- tests: identify rationale nodes by file_type metadata, not id substring
(decouples regression from current `<stem>_rationale_<line>` naming)
- tests: assert contact_form -> SQLiteQueue.enqueue calls edge directly
(previously only inversions were guarded; a dropped enqueue edge would
have passed silently)
- extract.py: clarify cross-file import indexing comment — the filter
intentionally indexes only class-level entities (function/method labels
end in "()" and are excluded by design)
* Fix#563 (HTML export): restore arrow direction in graph.html
to_json was patched to consult _src/_tgt before serializing edge
endpoints, but to_html still read endpoints directly from G.edges(),
so calls and rationale_for arrows in the rendered graph.html still
pointed in canonicalized (often inverted) order.
Apply the same direction restore in to_html when building vis.js
edges. Use data.get (not pop) since G.edges(data=True) yields the
live attribute dict and other exporters may run after to_html.
Adds test_to_html_preserves_calls_and_rationale_for_direction:
parses RAW_EDGES out of the rendered HTML and pins both the
caller->callee assertions from #563 and the rationale->parent
direction. Without the fix, the test reports 11 flipped arrows
on the 3-file repro.
to_wiki() writes a fresh set of community + god-node articles each call but
never deletes old files from previous runs. Since community labels are
LLM-generated and non-deterministic across rebuilds (per skill.md Step 5),
the same conceptual community is often named differently each time, leaving
its previous file as an orphan. After N rebuilds, wiki/ contains roughly N
times the active article count, with index.md only referencing the most
recent run's labels.
Real-world: a knowledge corpus accumulated 822 wiki .md files over 5
rebuilds, of which only 111 were referenced by index.md (710 orphans).
Fix: clear *.md files in the output directory at the start of to_wiki().
This is consistent with its existing fully-regenerative behavior — it
always writes the full set of articles + index, never partial updates.
Subdirectories and non-.md files are preserved (only top-level .md is
touched), so any user-placed auxiliary assets survive.
Tests: two new regression tests cover (1) stale article cleanup across
runs with different labels, and (2) preservation of non-.md user files
and nested subdirectories.
Closes#540.
Production audit on a 10,129-edge graph showed the INFERRED
confidence_score distribution is bimodal, not graded:
| Score bucket | Count | % of INFERRED |
|--------------|-------|---------------|
| <0.4 | 0 | 0% |
| 0.4-0.6 | 5,807 | 57% |
| 0.6-0.8 | 14 | 0.1% |
| 0.8+ | 4,308 | 42% |
Subagents collapse the continuous "0.4-0.9" guidance to a binary:
0.5 for "uncertain", 0.85+ for "confident", almost nothing in between.
Downstream filtering by confidence is therefore an on/off switch, not
the gradient the prompt promises.
Replace continuous ranges with a forced-rank discrete set:
0.95 direct structural evidence
0.85 strong inference
0.75 reasonable inference
0.65 weak inference
0.55 speculative but plausible
Models follow discrete rubrics far better than continuous ranges
(documented in calibration literature; same reason MCQ rubrics
outperform 0-100 scales). The set is anchored at non-round midpoints
to discourage 0.5 as a default.
Applied uniformly across all 10 skill-*.md files:
- 7 long-form (skill.md, skill-codex.md, skill-copilot.md,
skill-droid.md, skill-opencode.md, skill-windows.md, skill-trae.md):
full forced-rank table.
- 3 short-form (skill-claw.md, skill-aider.md, skill-kiro.md):
inline set notation INFERRED ∈ {0.55, 0.65, 0.75, 0.85, 0.95}.
Pure prompt edit — no code changes, no test impact. Effect is
observable only via re-extraction and inspection of the new
confidence_score distribution.
Closes#538.
The full-pipeline path's Step 9 already calls save_manifest, so the
NEXT --update can diff against the full-rebuild's baseline.
But the --update flow itself does NOT save the manifest after merging
the incremental result. Consequence: a file deleted between
--update #1 and --update #2 keeps reappearing in #2's deleted_files
list (since the manifest still records its old mtime), generating a
spurious ghost-node prune attempt every run.
Add save_manifest(incremental['files']) at the end of the --update
merge step in both skill.md and skill-copilot.md (the only two skill
files carrying the merge step verbatim). The 'incremental' dict is
already in scope from the prune block above; its 'files' key is the
full current file list, which is exactly what save_manifest needs.
Pure prompt edit — no code changes, no test impact.
Closes#539.
The current prune message in the --update flow is ambiguous:
Pruned 0 ghost nodes from 13 deleted file(s)
A reader can't tell whether (a) 13 files deleted, 0 of them had nodes
worth pruning (benign — graph already clean), or (b) 13 ghost nodes
still exist and only 0 got pruned (bug). The denominator is opaque.
Split into two messages so the semantics are explicit:
- Drift case: "Pruned N ghost node(s) from M deleted file(s) — drift
detected and corrected."
- No-drift case: "M file(s) deleted since last run, but no ghost
nodes were present in the graph — no drift."
Applied to both skill.md and skill-copilot.md (the only two skill
files carrying the --update merge step verbatim).
When a user configures core.hooksPath = ~/gitconfig/hooks in .gitconfig,
_hooks_dir() was constructing Path("~/gitconfig/hooks") without calling
expanduser(), so hooks were installed into <repo>/~/gitconfig/hooks instead
of the intended absolute path.
Add Path.expanduser() call immediately after reading the raw string from
git config, before the is_absolute() / root-relative fallback logic.
Fixes#547
Adds docs/colabfold-macos-install.md, a reproducible recipe for
installing LocalColabFold (ColabFold + AlphaFold2 weights) on Apple
Silicon Macs, so users can predict protein structures locally and feed
the resulting PDBs / confidence scores into a graphify corpus.
Notes upstream LocalColabFold's current top-level installer hardcodes
Linux x86_64 Miniforge and jax[cuda12], which fail on Apple Silicon (no
macOS CUDA wheels, no NVIDIA GPU). Recommends the bundled v1.5.5 M1
installer as the working Apple Silicon path. ColabFold 1.6.1 with
JAX 0.4.23 CPU build, verified by 60s smoke test on a 62-residue FASTA.
Linked from README.md under the existing 'Optional integrations' section
(also added by this PR if not already present from #620).
Co-Authored-By: Oz <oz-agent@warp.dev>
Defines a `test` optional-dependencies group so CI can install dev tooling
via `pip install -e ".[all,test]"` instead of an ad-hoc `pip install pytest`
sidecar. Kept out of `[all]` deliberately — `[all]` is runtime extras only.
graphify per-language AST extractors strip member-expression callees
(`x.foo()`, `obj.bar()`, `Pkg::baz()`) down to the trailing identifier,
queue them in `raw_calls`, and the cross-file resolver matches by bare
lowercase name against a global symbol table. When any file in the
corpus exposes a top-level helper with a common method name (e.g. a
`function log(...)` in a one-off smoke-test script), every unrelated
`Logger.log(...)` / `this.x.find(...)` / `Pkg::baz(...)` call across the
codebase resolves to that single helper, producing phantom god nodes
with hundreds of bogus INFERRED edges.
This patch:
* Adds `_MEMBER_CALL_NODE_TYPES` enumerating member-expression AST
node types across all supported languages.
* Tags every `raw_calls.append(...)` site with `callee_node_type`.
* Skips entries whose `callee_node_type` is in
`_MEMBER_CALL_NODE_TYPES` from the cross-file resolver in
`extract()`. In-file resolution behaviour is intentionally
preserved.
On a real production codebase (~1,668 TS/JS files, NestJS + Nuxt) this
eliminates 1,262 phantom edges and removes a phantom `log()` god node
that previously sat at #2 with 265 inbound edges (255 of them bogus).
Includes 3 regression tests that fail on the original code and pass with
this patch.
AST and semantic entries now write to cache/ast/ and cache/semantic/
respectively. Previously both used the flat cache/ dir causing semantic
results to overwrite AST entries for code files on mixed corpora, making
the shrink guard fire on every subsequent update run.
Migration: load_cached falls back to legacy flat cache/ for AST reads
so existing cache entries are not lost on upgrade.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Grep and Glob tools removed in CC v2.1.117; searches now go through Bash.
Hook now reads stdin tool_input and pattern-matches on search commands.
Uninstall/reinstall handles both old and new matcher for clean upgrades.
Closes#578
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- #550: _file_stem() includes parent dir to prevent node ID collisions for same-named files
- #555: extract() relativizes source_file paths before returning for cross-machine portability
- #562: to_json() returns bool; _rebuild_code() writes report/html only if json succeeded
- #563: skill prompts store rationale as node attribute, not separate node; enforce calls direction
- #566: Show All / Hide All buttons added to HTML community panel
- #575: _import_js() resolves tsconfig.json compilerOptions.paths aliases before external fallback
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Safi