Merged
Reuse Git-backed local project lexical and RAG retrieval (#205) #197
Reuse Git-backed local project lexical and RAG retrieval (#205) #197
Merged
urandonAdd on 23 days ago
urandon
urandonCollaborators
23 days ago

Scope and design

Epic #200 G5; base 6c606783970d3ee34867c19e0bb53d832b91ab80. Target ONLY codex/epic-200-git-backed-harness, never main. Issue contract comment193846897 and reviewed #145 r2 design apply.

A bounded immutable LocalGit snapshot adapts active issues/comments/wiki to the EXISTING repositorydocs Retriever/Indexer/chunker/revision sets/vectors/checkpoints/retention. No new engine, lifecycle protocol, authority DB or journal. Git remains the sole byte authority; derived storage contains only metadata/locators/digests/vectors. Pinned readers cannot inspect a worktree or execute Git; exact blob hashes and line ranges are verified from source bytes. Receipts/counters/tombstones are excluded.

CLI project search/index and local-only MCP equivalents share private IPC/service contracts. Default fulltext requires no cache/provider. Explicit hybrid uses a registered pre-provisioned local-boundary profile and honestly falls back on absent/stale/partial/failed/corrupt derived state. Explicit foreground index requires local write intent, max128 NEW provider chunks, existing10s deadline and resumable vector-only handoffs; replay totals are not confused with new-work limits.

Machine-owned configuration is authoritative in strict-local mode; cwd-owned settings cannot retarget provider/cache. Normal GitCode configuration/catalog behavior is unchanged. Provider transfer has deadlines/body bounds, no redirects/environment proxies/setup commands.

Exact final source and verification

Head 2ce25c5fe00b75c8e82a4dcb0e9aeafaa3864ee6; tree 9756e24ab4a67230ca35146561198ac35bebb1c6.

PASSED on final source:

  • go test ./... (native qualification98.476s; servicectl64.289s).
  • Full nine affected package go test -race -p 4 ... -count=1 (repositorydocs19.915s; service103.673s; servicectl140.771s; main31.894s).
  • Affected vet, product documentation contracts, whitespace checks.
  • Retained independent actual Git/Unix/loopback regressions, plus hash-consistent invalid-vector and real CLI/MCP query-only503 controls.

Review history: 1416a85 had six P2 findings; 41b366b fixed them and exposed one remaining vector-qualification P2. Final source fixes all with retained counterexamples: effective lexical coverage; complete vectors; authoritative lines; overflow-safe counts; model/revision/dimension consistency including per-item fallback; correct checkpoint accounting; stored dtype/SHA256 and finite/nonzero query/stored vectors using existing RAG normalization. Earlier timings/green runs are historical, not final-head evidence.

One preceding allGoattempt failed the UNCHANGED O5 actual stopped-worker retry-conservation assertion. Exact isolated test passed30.934s; private diagnostic-only overlay with identical assertions passed count3/91.865s; final complete gate passed. Root cause remains unestablished; no assertion/deadline/test was weakened or skipped.

Independent final exact-SHA review: SOURCE CLEAR, no P0–P2; all33 files and original raw regression bytes verified. Required head CI38640870904 PASSED, attempt1/job116115724634/all18steps, exact2ce25c5 push. Root independently checked run/jobs in addition to Luna watcher. Refreshed canonical PR: exact reviewed head, unchanged epic base6c606783, main unchangedc7a0af6, zero unresolved discussions, clean checkout. Merged ONLY into the epic branch as 045898578bc55bfd9fdbfd49e6da35084e825e64, expected parents and identical reviewed tree. Resulting epic CI38641719429 PASSED attempt1/job116118598295/all18steps, independently verified and watched by Luna. #205 closed with final evidence; epic checkpoint5/7 is comment193848220. Main unchanged; #206/#207 remain required.

Risk, recovery and UX

Only explicit index/local intent transfers project text to a pre-provisioned local-boundary profile. No request can override paths/endpoints. Derived state is disposable; recovery must preserve Git stores and authoritative receipts/history. Current lexical reads survive damaged derived state. Read-only registrations/MCP policy reject index writes.

Protected Admin cohort #206 and integrated backup/rollout #207 remain REQUIRED before the whole epic enters main. No UI source/assets changed here. No partial main merge, tag or install.

Closes #205

likedislike
Pull Request has been successfully merged by urandon.
(Thank you for urandon's contribution.)
urandonurandonCollaborators
23 days ago Associated issue: [Epic] Git-backed local research harness — thin-service replacement,[LocalGit G4] CLI and MCP usable issue/wiki vertical slice,[LocalGit G5] Reuse lexical/RAG search for issue and wiki Git files,[LocalGit G6] Protected Admin Local projects cohort,[LocalGit G7] Git bundle restore, documentation and integrated acceptance
urandonurandonCollaborators
23 days ago Pushed 1 commits:41b366b2-fix(localgit): verify derived retrieval and model contracts (#205)
urandonurandonCollaborators
23 days ago Modify description of pull request
urandonurandonCollaborators
23 days ago Pushed 1 commits:2ce25c5f-fix(search): qualify stored and query vector contracts (#205)
urandonurandonCollaborators
23 days ago Modify description of pull request
urandonurandonCollaborators
23 days ago Modify description of pull request
urandonurandonCollaborators
23 days ago Modify description of pull request
urandonurandonCollaborators
23 days ago Merged pull request, merge node SHA: 045898578bc55bfd9fdbfd49e6da35084e825e64
urandonurandonCollaborators
23 days ago Modify description of pull request