Repository navigation
graph: a code-only, pruned build; communities named by their hub; routing asks with identifiers - #85
Merged
Merged
Conversation
The umbrella's graph answered 1 of 8 symbol questions right first, with 53% of returned nodes noise: `update .` added 3,381 Markdown headings, and 839 library/package nodes with no source file (`marola`, `kyo`, `String`) were the biggest hubs, bridging repos through 1,874 `references` edges. `graph build` now runs `extract . --code-only --no-label`, drops every node without a `source_file`, the edges touching one and self-loops (jq), then `cluster-only . --no-label --graph <graph.json>`, which rewrites graph.json, GRAPH_REPORT.md and graph.html from the pruned graph. It never runs `update`. In 0.9.66, cluster-only calls a labelling model only when neither `--no-label` nor a saved labels file is present; under strace it made no connect() at all. Before/after (graphify 0.9.66): devkit 701 -> 659 nodes, 1,508 -> 1,360 edges; umbrella 3,056 -> 2,218 nodes, 6,331 -> 3,923 edges (6,441 nodes before this change, with `update`). The 8 questions: 2/8 right first, 23% noise; the noise left is tests, JSON and .claude, marola-dev/marola#767. The routing skill now says to query with identifiers and to `git grep` a concept or behaviour word first: graphify matches symbol names, not meanings. CHANGELOG is untouched: release.py writes the entry from commit subjects. Closes #82 Tested: graph --self-test red first (6 failures: build ran update, no prune, no report/HTML), green after; new cases build_is_code_only and build_prunes_library_nodes_and_self_loops, build argv expectation adjusted. shellcheck, docs_lint, tests/self-tests.sh, claude plugin validate all pass. Real graphify 0.9.66 build: devkit 659 nodes/1,360 edges, umbrella 2,218/3,923, 0 nodes without a source file, 0 self-loops, 0 Markdown nodes, "Token cost: 0 input · 0 output", git status --porcelain --ignored empty in both; umbrella 8-question eval 2/8 right first, 23% noise (was 1/8, 53%). Cost: ~$5.87 · diff-size estimate (cost-split --estimate-commit) Co-Authored-By: Claude <noreply@anthropic.com>
`graph extract` and `graph update` passed straight to graphify, and either would overwrite the pruned graph with an unpruned one. Both now exit non-zero pointing at `graph build`; the verbs are build, query, path and explain. With `cluster-only --no-label` every community was "Community N", which left graph.html and GRAPH_REPORT.md unreadable. build now runs cluster-only twice. The first assigns communities. jq then names each one after its highest-degree member (a file node only when nothing else is there; ties to the shortest label, then the first alphabetically) into .graphify_labels.json, with no .sig beside it. In 0.9.66 the second pass keeps a saved label when the community count matches, since clustering is seeded and cids are remapped to the previous run. Otherwise it names the community after graphify's own hub. No model is called in either branch. `export html --labels` would have named the HTML only, not the report. A saved named label makes cluster-only copy the graph to a dated backup dir, so build removes that. Umbrella, under strace -f -e trace=connect: 16 connect() calls, all AF_UNIX (multiprocessing), none AF_INET/AF_INET6; token footer 0/0; all 172 community names in graph.json are ours. The largest: self_test, renderCard, RunHandle, rewrite, WaterSample, GroundTruth, bash, HourlyConditions, run, OpenMeteoForecasts. Part of #82 Tested: graph --self-test red first (10 failures: one cluster-only, no labels, extract/update passed through), green after; new cases refuses_update_and_extract and communities_named_by_hub (incl. a rebuild drops the last labels), build argv and call counts adjusted. shellcheck, docs_lint, tests/self-tests.sh, claude plugin validate pass. Real graphify 0.9.66 build on the umbrella under strace: 0 non-AF_UNIX connects, token footer 0/0, 0 "Community N" placeholders, no dated backup left, git status --porcelain --ignored empty; devkit build likewise clean. Cost: ~$6.15 · diff-size estimate (cost-split --estimate-commit) Co-Authored-By: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary — The umbrella's graph answered 1 of 8 symbol questions right first, with 53% of returned nodes noise:
update .added 3,381 Markdown headings, and 839 library/package nodes with no source file (marola,kyo,String) were the biggest hubs, bridging repos through 1,874referencesedges.Closes #82
$5.87 · diff-size estimate (cost-split --estimate-commit) (
$6.15 · diff-size estimate (cost-split --estimate-commit) (36f5bf4)e223b25)What changed
36f5bf4)e223b25)Real builds (graphify 0.9.66,
unshare -rn): devkit 701→659 nodes; umbrella 3,056→2,218 nodes, 0 without a source file, 0 self-loops, 0 Markdown nodes; report token footer 0/0;straceshows no networkconnect()(only local Unix sockets from Python multiprocessing);git status --porcelain --ignoredempty. With marola-dev/marola#768's ignores the umbrella graph is ~1,600 nodes. 8 symbol questions on the umbrella: 2/8 right first, 23% noise without #768's ignores (1/8, 53% before this change; ~3/8, 0% with both). The remaining misses are concept questions; the routing skill now sends those togit grepand asks the graph with identifiers.Communities are named by their hub symbol (
.graphify_labels.json, no model): the 10 largest are self_test, renderCard, RunHandle, rewrite, WaterSample, GroundTruth, bash, HourlyConditions, run, OpenMeteoForecasts.graph update/extractnow refuse (they would overwrite the pruned graph). Takes effect in consumers with the next release (v0.8.3).