Mid-session drift refresh gate + embeddings cap#4
Merged
Conversation
`--embeddings` with no arg keeps GitNexus's default embedding limit; `0` removes it so the full symbol set gets vectors. Applied to all four analyze scripts (refresh / full / pdg / full-pdg) — the ones agent-refresh and the pre-commit hook run. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… index Staleness was purely commit-based (indexedCommit === HEAD → "fresh"), so the uncommitted edits an agent makes mid-session never marked the index stale. Agents only refreshed at trigger events (session start / commit), never during routine work — so graph queries silently ran against a graph that ignored the edits. - check-staleness now reports `driftingFiles` = git-dirty SOURCE files modified since the index build (mtime > indexedAt). Cheap (stats only the dirty files) and RESETS on refresh because indexedAt advances — a plain `git status` count would nag forever. - new shared `classifyMcpDrift`: graph QUERY tools (query/context/cypher/impact/pdg_query/ trace/explain/api_impact/route_map/shape_check) hard-block past a drift threshold with a nudge to the FAST incremental `gitnexus:refresh`. Non-query tools (detect_changes, rename, check, tool_map…) pass; edits/reads/greps are untouched, so editing flow isn't interrupted. - wired into the Claude + Cursor MCP guards, AFTER the must_refresh check (so it only applies when the index is otherwise commit-fresh). - `driftRefreshThreshold` config (default 3; 0 disables). Surfaced in `gitnexus:status` + a `driftRefreshBlocks` scorecard counter. Contract now teaches proactive incremental refresh. Tests: 58/58 — classifyMcpDrift logic + real check-staleness drift/reset in a temp git repo. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…m cost Self-review (microscope waves, 5 independent lenses) of the drift gate — fixes: - HIGH: the drift gate fired in classical_fallback too (the else-of-must_refresh covers it), overriding the classical-fallback escape hatch and telling the agent to run the refresh that just FAILED. classifyMcpDrift now enforces phase === "fresh" itself; both guards pass ctx.phase. - MEDIUM: countDrift used a 3rd hand-rolled source-ext regex diverging from the kit's DEFAULT_SOURCE_EXT_RE (silently no-drift for .cts/.ex/.clj/.lua/…, and ignored the sourceExts config override). Now uses loadHookConfig(root).sourceExtRe. - MEDIUM: countDrift's `git status` ran on EVERY tool call (grep/read/edit/shell/stale) though driftingFiles is consumed only by the MCP guards on the fresh path. Moved it into the commit-fresh branch + skip it entirely when the gate is disabled (threshold <= 0). - LOW: non-ASCII paths undercounted (git octal-escaping) -> `-c core.quotePath=false`. - LOW: Cursor never bumped the driftRefreshBlocks scorecard counter -> added. - LOW: tempered the "fast" overclaim (uncapped embeddings make large batches / cold model slower). Verified-correct by the review (no change): --embeddings 0 = "disable the cap" (traced in the gitnexus CLI source; also a latent-bug fix vs bare --embeddings capping at 50k nodes), and the core premise holds (incremental analyze hashes on-disk content -> reindexes uncommitted edits + advances indexedAt). 58/58 (+ phase-gate regression test). Co-Authored-By: Claude Opus 4.8 (1M context) <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.
Two agent-effectiveness fixes for the GitNexus enforcement layer.
1. Lift the embeddings cap on refresh (
--embeddings 0)Bare
--embeddingsapplies GitNexus's default 50k-node cap, so on large repos embeddings get skipped → the kit's ownmissing_embeddingscheck then wedges a permanent stale loop.--embeddings 0disables the cap (verified in the GitNexus CLI help: "pass 0 to disable the cap entirely"). Applied to all four analyze scripts.2. Mid-session drift gate
Staleness was purely commit-based (
indexedCommit === HEAD → fresh), so the agent's own uncommitted mid-session edits never marked the index stale — agents only refreshed at trigger events (session start / commit), never during routine work, and graph queries silently ran against a graph that ignored the edits.check-stalenessreportsdriftingFiles= git-dirty source files modified since the index build (mtime > indexedAt); cheap (stats only dirty files) and resets on refresh.classifyMcpDriftblocks graph query tools (query/context/cypher/impact/pdg_query/trace/…) pastdriftRefreshThreshold(default 3,0disables) with a nudge to the fast incrementalgitnexus:refresh. Non-query tools (detect_changes,rename, …) + edits/reads/greps are untouched.freshphase only. Surfaced ingitnexus:status+ adriftRefreshBlocksscorecard counter.Microscope self-review (pass #1)
5 independent lenses. Fixed: drift wrongly firing in
classical_fallback(overriding the escape hatch); a divergent source-ext regex (silently no-drift for.cts/.ex/.clj/…); the driftgit statusrunning on the whole hot path; non-ASCII paths, Cursor telemetry, "fast" overclaim. Verified against GitNexus source:--embeddings 0= disable-cap, and incremental analyze re-hashes on-disk content (so it reindexes uncommitted edits + advancesindexedAt).Tests: 58/58.
🤖 Generated with Claude Code