Summary
The code_only check in the --update runbook hard-codes a set of code extensions that has drifted far from graphify.detect.CODE_EXTENSIONS. On 0.9.22 the literal covers 31 of 94 extensions. The consequence is a silent token cost: an incremental update that touched only code is classified code_only: False and takes the full semantic path, dispatching LLM subagents over a changed-file set that has nothing semantic in it.
Where
Source of truth is the shared fragment:
tools/skillgen/fragments/references/shared/update.md
which propagates to every host variant (graphify/skills/{pi,copilot,opencode,kilo,...}/references/update.md) and to the tools/skillgen/expected/* fixtures — 36 code-search hits for the literal in total.
code_exts = {'.py','.ts','.js','.go','.rs','.java','.cpp','.c','.rb','.swift','.kt','.cs','.scala','.php','.cc','.cxx','.hpp','.h','.kts','.lua','.toc','.f','.F','.f90','.F90','.f95','.F95','.f03','.F03','.f08','.F08'}
...
code_only = all(Path(f).suffix.lower() in code_exts for f in all_changed)What is missing
The literal is a strict subset of CODE_EXTENSIONS — nothing in it is stale, it simply never grew. The 68 missing entries include extensions that do have an AST extractor shipped in graphify/extractors/:
.sql .ps1 .psm1 .psd1 .jl .zig .tf .tfvars .hcl .pas .dpr .ex .exs .dart .groovy .cls .trigger .sv .v .m .mm .tsx .jsx .mjs .cjs .mts .cts .vue .svelte .astro .razor .cshtml .sh .bash .json .sln …
Reproduction
--update on a repo where the only changed files are three .sql files:
code_only: False # expected True
.sql is in CODE_EXTENSIONS, is dispatched to extractors/sql.py, and extracted deterministically by Step 3A — so the semantic pass that this misclassification triggers has nothing to do. On my run the AST step produced 12 nodes / 9 edges for those files; the semantic dispatch would have added zero and cost a subagent batch.
Suggested fix
Read the list from the installed package instead of restating it:
try:
from graphify.detect import CODE_EXTENSIONS as code_exts
except ImportError:
code_exts = {...} # existing literal as fallbackThen regenerate tools/skillgen/expected/*. I am running this patched locally against 0.9.22 and the three cases that flip (.sql-only, .ps1-only, .R-only) all behave correctly; mixed code+doc and doc-only are unchanged.
One caveat worth deciding on
CODE_EXTENSIONS is not exactly the right predicate — it is broader than "has an extractor". .r is the clearest case: it is in CODE_EXTENSIONS but has no AST dispatch (#1689), so after this fix an R-only update becomes code_only: True and correctly skips the semantic pass, but then produces zero nodes on the AST path. The net outcome for R is unchanged (it yields nothing either way, since code-typed files are never queued for semantic extraction), so this fix does not regress anything — but it does mean the truly correct predicate is "extensions with a dispatchable extractor", i.e. gated on _get_extractor(...) rather than on the extension table. That distinction is the same one #1689 is about, and it might be worth exposing as a single source of truth both places can import.
Summary
The
code_onlycheck in the--updaterunbook hard-codes a set of code extensions that has drifted far fromgraphify.detect.CODE_EXTENSIONS. On 0.9.22 the literal covers 31 of 94 extensions. The consequence is a silent token cost: an incremental update that touched only code is classifiedcode_only: Falseand takes the full semantic path, dispatching LLM subagents over a changed-file set that has nothing semantic in it.Where
Source of truth is the shared fragment:
tools/skillgen/fragments/references/shared/update.mdwhich propagates to every host variant (
graphify/skills/{pi,copilot,opencode,kilo,...}/references/update.md) and to thetools/skillgen/expected/*fixtures — 36 code-search hits for the literal in total.What is missing
The literal is a strict subset of
CODE_EXTENSIONS— nothing in it is stale, it simply never grew. The 68 missing entries include extensions that do have an AST extractor shipped ingraphify/extractors/:.sql.ps1.psm1.psd1.jl.zig.tf.tfvars.hcl.pas.dpr.ex.exs.dart.groovy.cls.trigger.sv.v.m.mm.tsx.jsx.mjs.cjs.mts.cts.vue.svelte.astro.razor.cshtml.sh.bash.json.sln…Reproduction
--updateon a repo where the only changed files are three.sqlfiles:.sqlis inCODE_EXTENSIONS, is dispatched toextractors/sql.py, and extracted deterministically by Step 3A — so the semantic pass that this misclassification triggers has nothing to do. On my run the AST step produced 12 nodes / 9 edges for those files; the semantic dispatch would have added zero and cost a subagent batch.Suggested fix
Read the list from the installed package instead of restating it:
Then regenerate
tools/skillgen/expected/*. I am running this patched locally against 0.9.22 and the three cases that flip (.sql-only,.ps1-only,.R-only) all behave correctly; mixed code+doc and doc-only are unchanged.One caveat worth deciding on
CODE_EXTENSIONSis not exactly the right predicate — it is broader than "has an extractor"..ris the clearest case: it is inCODE_EXTENSIONSbut has no AST dispatch (#1689), so after this fix an R-only update becomescode_only: Trueand correctly skips the semantic pass, but then produces zero nodes on the AST path. The net outcome for R is unchanged (it yields nothing either way, since code-typed files are never queued for semantic extraction), so this fix does not regress anything — but it does mean the truly correct predicate is "extensions with a dispatchable extractor", i.e. gated on_get_extractor(...)rather than on the extension table. That distinction is the same one #1689 is about, and it might be worth exposing as a single source of truth both places can import.