h-dv/code-index

Watch

Fork 0

feat: explained similarity across active plugin generations #27

New issue

Open

opened 2026-07-20 22:38:21 +02:00 by buildagent · 1 comment

buildagent commented

2026-07-20 22:38:21 +02:00

Member

Rank 6 of the 2026-07-20 index-data brainstorm (_prdoc/records/brainstorm-2026-07-20-index-data-catalog.md).

What: one fingerprint mechanism (ordered ref-sequence per symbol) powering three products: exact clone clusters (consolidation worklist), drift_clones (near-clones that have diverged — "copy 2 already differs, here's the behavioral delta"), and a pre-write find_similar point query ("does this helper already exist?") that prevents future duplication.

Why — it found a real bug: relative_path had drifted between index.rs and watcher.rs (Jaccard 0.82) — the watcher copy lacked ParentDir and single-file-root handling; fixed in v0.5.19 by consolidation. Scale: 58 clone clusters / 193 duplicate symbols / ~1,402 redundant lines (the parse() test helper copied 10x; workspace_root copied 6x despite tests/common/mod.rs existing). setup_logging drift (J=0.93) triaged as intentional (--log-json). The existing outline_hash column finds the m0016–m0019 migration template family for free.

Cost: small for exact clusters + find_similar (pure SQL); medium for drift pairs (Jaccard, name-blocked); no new data.

Runtime-plugin architecture expansion

Fingerprints are generation- and language-profile scoped. The same source parsed by a different package/extractor may legitimately produce different fact sequences; never compare raw ids or silently combine pending/active contributions.

Each result carries:

active generation and package/component provenance;
source language/profile;
features responsible for similarity;
coverage/degradation basis;
whether the match crosses builtin/dynamic implementations.

The full-language migration in #80 can use exact/near similarity diagnostically, but parity is determined by canonical fact comparison, not a similarity score.

Order remains: find_similar point query/engine, exact clusters, then measured drift clones. Consume it through #35/#34 first; add a top-level surface only if #51 demonstrates value.

Rank 6 of the 2026-07-20 index-data brainstorm (`_prdoc/records/brainstorm-2026-07-20-index-data-catalog.md`). **What**: one fingerprint mechanism (ordered ref-sequence per symbol) powering three products: exact clone clusters (consolidation worklist), **drift_clones** (near-clones that have diverged — "copy 2 already differs, here's the behavioral delta"), and a pre-write `find_similar` point query ("does this helper already exist?") that prevents future duplication. **Why — it found a real bug**: `relative_path` had drifted between index.rs and watcher.rs (Jaccard 0.82) — the watcher copy lacked `ParentDir` and single-file-root handling; fixed in v0.5.19 by consolidation. Scale: 58 clone clusters / 193 duplicate symbols / ~1,402 redundant lines (the `parse()` test helper copied 10x; `workspace_root` copied 6x despite `tests/common/mod.rs` existing). `setup_logging` drift (J=0.93) triaged as intentional (`--log-json`). The existing `outline_hash` column finds the m0016–m0019 migration template family for free. **Cost**: small for exact clusters + find_similar (pure SQL); medium for drift pairs (Jaccard, name-blocked); no new data. ## Runtime-plugin architecture expansion Fingerprints are generation- and language-profile scoped. The same source parsed by a different package/extractor may legitimately produce different fact sequences; never compare raw ids or silently combine pending/active contributions. Each result carries: - active generation and package/component provenance; - source language/profile; - features responsible for similarity; - coverage/degradation basis; - whether the match crosses builtin/dynamic implementations. The full-language migration in #80 can use exact/near similarity diagnostically, but parity is determined by canonical fact comparison, not a similarity score. Order remains: find_similar point query/engine, exact clusters, then measured drift clones. Consume it through #35/#34 first; add a top-level surface only if #51 demonstrates value.

buildagent referenced this issue

2026-07-20 22:45:36 +02:00

feat: review_diff v2 — source plus plugin-generation semantic risk #34

~~buildagent referenced this issue 2026-07-20 22:45:36 +02:00~~

feat: context_pack v2 — plugin-aware analogues, cycles and evidence diversity #35

buildagent commented

2026-07-21 14:26:47 +02:00

Author

Member

Triage: P2 — split by agent value

Implement in this order:

find_similar(symbol|span) point query — prevents agents from creating another implementation.
Exact duplicate clusters using existing fingerprints/outline hashes.
Drift-clone analysis only after the first two are measured and trusted.

The first capability should be consumable by context_pack (#35), not required for its v1. Repository-wide clone inventories are secondary to the pre-write question “does this already exist?”. Every similarity result needs the compared evidence/features; avoid an unexplained scalar score.

### Triage: P2 — split by agent value Implement in this order: 1. `find_similar(symbol|span)` point query — prevents agents from creating another implementation. 2. Exact duplicate clusters using existing fingerprints/outline hashes. 3. Drift-clone analysis only after the first two are measured and trusted. The first capability should be consumable by `context_pack` (#35), not required for its v1. Repository-wide clone inventories are secondary to the pre-write question “does this already exist?”. Every similarity result needs the compared evidence/features; avoid an unexplained scalar score.

buildagent referenced this issue

2026-07-21 14:26:47 +02:00

feat: quick wins umbrella — 7 single-query capabilities on existing data #30

buildagent referenced this issue

2026-07-21 14:26:47 +02:00

feat: context_pack v2 — plugin-aware analogues, cycles and evidence diversity #35

buildagent referenced this issue

2026-07-21 15:43:30 +02:00

feat: context_pack v2 — plugin-aware analogues, cycles and evidence diversity #35

buildagent referenced this issue

2026-08-26 13:37:22 +02:00

feat: quick wins umbrella — 7 single-query capabilities on existing data #30

buildagent changed title from ~~feat: duplication family — find_duplicates + drift_clones + find_similar~~ to feat: explained similarity across active plugin generations

2026-08-26 13:41:58 +02:00

buildagent added the

Priority

Low

label

2026-08-26 13:42:32 +02:00

buildagent added a new dependency

2026-08-26 13:44:05 +02:00

#35 feat: context_pack v2 — plugin-aware analogues, cycles and evidence diversity

buildagent added a new dependency

2026-08-26 13:44:13 +02:00

#34 feat: review_diff v2 — source plus plugin-generation semantic risk