Repository navigation
Conversation
skills/claude/references/query.md was 103 lines — missing the constrained vocabulary expansion step (Step 0) that all other platform skill installs (vscode, codex, copilot, opencode, etc.) already ship at 249 lines. Without Step 0, the graphify query CLI latches onto wrong start nodes for natural-language questions (e.g. 'how does auth work' starts BFS from a Session test fixture instead of auth.ts) because the binary does no stemming or synonym matching — it relies on the agent to expand the question against the actual graph vocabulary first. Fix: copy skills/vscode/references/query.md → skills/claude/references/query.md so both are identical. Closes Graphify-Labs#1255
uditstocks
left a comment
There was a problem hiding this comment.
Spotted a minor inconsistency between Step 0 and Step 1.
inline comment added on line 49.
Step 0 extracts vocab tokens with if 3 <= len(t) <= 30, so 3-char tokens (api, use, get, tok, auth) are deliberately included. Step 1 was filtering with len(t) > 3 (minimum 4 chars), silently dropping those same tokens before BFS/DFS. Change to len(t) >= 3 to align both steps. Addresses review comment by @uditstocks on PR Graphify-Labs#1256.
Same >= 3 fix applied to amp, claw, codex, copilot, droid, kilo, kiro, opencode, pi, trae, vscode, windows - all had the same len(t) > 3 bug.
|
Also addressed the reviewer's observation cross-platform in f485fc2. The All 13 platforms now use |
|
As far as I can tell, none (or very few, I haven't looked at all of them) of the shipped skills include this step 0. Looks like skillgen wasn't done correctly? Edit: Nevermind, I figured out that none of the other platforms have it because they use the cli-inline query variant. Not sure why or why that one doesn't include this step 0 (which seems pretty important to get good results), but that's a different issue. |
What
graphify/skills/claude/references/query.mdwas 103 lines — missing the Step 0 Constrained query expansion block that every other platform skill (vscode, codex, copilot, opencode, claw, kiro, amp, pi, trae, droid, kilo, windows) already ships at 249 lines.Additionally, a cross-platform bug was found (via reviewer @uditstocks) and fixed across all 13 platform skills: Step 1's length filter used
len(t) > 3(min 4 chars) while Step 0 extracts tokens withlen(t) >= 3(min 3 chars), silently dropping valid 3-char tokens likeapi,use,get,tok,authbefore BFS/DFS.Root cause
The
skills/claude/directory contains a stale copy ofreferences/query.mdthat predates the vocabulary-expansion rewrite. All other platform directories received the update;claude/did not.The
len(t) > 3vs>= 3mismatch existed in the source file and was copied to all platforms — making it cross-platform from the start.Fix
Commit
a1ac4cb— syncskills/claude/references/query.md(103 lines → 249 lines):Commit
5643c7f— align Step 1 filter forclaude/(> 3→>= 3on line 49)Commit
f485fc2— same>= 3fix applied to all 12 remaining platforms:amp, claw, codex, copilot, droid, kilo, kiro, opencode, pi, trae, vscode, windowsSee: #1256 (comment)
Why it matters
Without Step 0, the
graphify queryCLI latches onto wrong start nodes for natural-language questions. The binary does no stemming or synonym matching — it relies entirely on the agent first expanding the question against actual graph vocabulary tokens.Without the
>= 3fix, tokens that Step 0 deliberately selected from graph vocab are silently discarded before traversal even begins.Verified on a 757-file, 11 634-node TypeScript/Next.js repo: after applying this fix, natural-language query quality matches VS Code Copilot results from the identical graph.
Closes #1255