fix(explore): three retrieval regressions found against v1.6.0 - #2057
Merged
Merged
Conversation
The pre-release agent A/B (main vs v1.6.0, 7 README repos, 3 runs each) had main returning 15% less source on vscode for the same queries, with Reads in 2 of 3 runs where v1.6.0 had none. Replaying every query both builds' agents issued traced it to three causes: 1. Named gaps (#1711) displaced source. Every cluster fit measured the gap names, which on a long path run to hundreds of chars, so a class shrunk into its file's room overran it by the names alone and was dropped whole: rpcProtocol.ts emitted 222 of 5,496 funded chars (v1.6.0: 5,194), and the test fixture came back as an empty fence. Fits now measure bare gaps; names are added at assembly from what the file's budget has left. 2. Interface members (#1638) corroborated English words. The NL-stopword guard lets a bare word seed a definition when another query word names a symbol in the same file; `readonly host` in an options interface made "main" seed `main()` in agentHostServerMain.ts, which took the named-first tier from the answer files. Members declared inside an interface no longer count as corroboration. 3. The interface-dispatch note announced dispatch through any common base, declared or not: "`extension` -> runtime dispatch to 2706 types implementing Disposable" opened 9 of 31 vscode answers. A supertype now has to declare the member (it, an ancestor, or a same-named Swift extension). When absence can't be judged, because a protocol's requirements aren't indexed, the note stays as before. Replayed queries, source lines vs v1.6.0 (main -> fixed): vscode 10,744 -> 12,577 (12,617), excalidraw 13,065 -> 13,698 (13,764), django 8,420 -> 8,803 (8,954), tokio 13,618 -> 14,151 (14,600), okhttp 6,511 -> 6,883 (7,090), alamofire 7,019 -> 7,249 (7,120); no query below 60% of v1.6.0. Real dispatch notes (Alamofire adapt/retry/asURLRequest, django, okhttp) are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner
Author
|
Confirming A/B (this PR + #2055 vs v1.6.0, same 7 README repos, 3-turn sessions, 3 runs per arm, Sonnet
main before these fixes had Reads in 2 of 3 vscode runs (5 and 2); with them it has 0 of 3. Wall-clock from this pass is not usable: the machine was shared with another workload (load average ~20), and the slow arms' API time ballooned. Alamofire run 1 took 513s against 1,801s of API time, at the same cost as its 68s run. Explore latency measured directly on the same index is identical between main and this build (vscode median ~4.1s, alamofire ~110ms, interleaved runs), so this change adds no codegraph time. The first pass, on a quieter machine, had main 5% faster than v1.6.0 overall. |
bompus
added a commit
to bompus/codegraph
that referenced
this pull request
Sep 28, 2026
The fork keeps its kernel-only extraction and resolution, its synthesis refresh on incremental sync, and its migrations v10-16. Upstream's TS resolver files (name-matcher, js-builtins, cpp-constructor, cpp-macro-visibility, synthesis-stage) and the Scala wasm grammar stay deleted. The kernel walkers carry the ported upstream kernel changes: NodeIdAllocator column suffixes, Python body docstrings, Scala objects as modules, the C++ constructor and macro-name helpers, and the TS/JS expression-receiver and curried-wrapper cases. Taken from upstream: - read-only CodeGraph.open/openSync (colbymchenry#1963), the shared project lease for MCP engines, and the query-pool and daemon fixes - sync throws LockUnavailableError when the file lock is held (colbymchenry#2014); open() destroys the instance if its sync fails - watcher lock-contention re-arm, explore miss diagnostics, the Scala grammar bump (colbymchenry#1823), and the vitest run-TMPDIR setup (colbymchenry#1879) - pending-reference recovery counts on SyncResult and in `codegraph sync` output (colbymchenry#1360), ported onto the fork's orphan sweep - `this.#field.method()` call edges (colbymchenry#1987), ported into the kernel's field-type matcher and method-call receiver pattern Fork-side adjustments: - edges keep the fork's `ORDER BY kind, id`. Upstream's `kind, target, line, col` tie-break (and a position-first variant) dropped CG-36's answer cluster and colbymchenry#1711's named gaps; the db-perf ordering test now checks kind precedence and insertion-order ties - a gap marker that can't afford its full name list names fewer symbols plus `+N more` instead of going bare (colbymchenry#2057 budgeting kept) - the lease catch-up sync defers synthesis, as the fork's engine did - enableWatcherAfterHandover nudges the existing lease instead of acquiring a second one - dropped idx_edges_synthesis_site from the bulk-load index list (the fork schema has no such index) - Python and Go function refs stay bare-named in both the kernel and the generic extractor until colbymchenry#1820 is ported to both - tests: the colbymchenry#1963 daemon test waits for attach before connecting; the Windows zip-archive cases run only where `zip` exists; the watcher and generated-flag tests keep their fork fixtures - golden dumps re-baselined; torture-multilang now covers 42 files Port follow-ups (tests skipped with a "kernel port pending" note): - colbymchenry#1820 receiver-path function refs (Python, Go) - colbymchenry#1838/colbymchenry#1839/colbymchenry#2035 C++ macro names and constructor calls - colbymchenry#1932/colbymchenry#2034 alias-binding object member boundaries - Scala companion resolution; colbymchenry#1857 Go bare calls; colbymchenry#1986 ABI evidence; expression receivers; HAR excluded imports; value+interface pairs - colbymchenry#2033 synthesis_inputs convergence model (7 tests) - colbymchenry#2043 stale-refusal banner policy needs a decision
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.
Problem
The pre-release agent A/B (main vs v1.6.0, 7 README repos, 3-turn sessions, 3 runs per arm, Sonnet
--effort high) had main slower on vscode (+9%) and django (+20%), with Reads in 2 of 3 vscode runs where v1.6.0 had none. Replaying every query the agents issued through both builds showed main returning 15% less source on vscode (and 1–8% less elsewhere) for identical queries. There were three causes:rpcProtocol.tsemitted 222 of 5,496 funded chars (v1.6.0: 5,194), and the new fixture came back as an empty code fence. Fits now measure bare gaps; names are added at assembly from what the file's budget has left.method_signature/property_signaturemissing from the extractor) — every call through a.d.tsplatform API is invisible #1638) corroborated English words. The NL-stopword guard lets a bare word seed a definition when another query word names a symbol in the same file.readonly hostin an options interface made "main" seedmain()inagentHostServerMain.ts, which took the named-first tier from the real answer files (9 symbols found instead of 32). Members declared inside an interface no longer count.extension→ runtime dispatch to 2706 types implementingDisposable" opened 9 of 31 vscode answers, becauseDisposablenever declaresextension. A supertype now has to declare the member: itself, an ancestor, or a same-named Swift extension. When that can't be judged (Swift protocol requirements aren't indexed), the note stays as before.Validation
Replayed agent queries, total source lines (main → this PR, v1.6.0 in parentheses):
No query is below 60% of v1.6.0 (main had 2). Real dispatch notes are unchanged: Alamofire
adapt/retry→RequestInterceptorandasURLRequest→URLRequestConvertible, django, okhttp. vscode's bogusDisposablenotes are replaced by real declarers (send→IProtocolTransport,call→IServerChannel).Tests: new
explore-named-gap-budget.test.ts, plus cases indynamic-boundaries.test.tsandexplore-nl-stopword-collision.test.ts; 5 of the new cases fail onmain. All 27 explore-related suites pass (348 tests). A confirming agent A/B of this PR together with #2055 against v1.6.0 is running; results will be posted here.🤖 Generated with Claude Code