feat(temporal): LLM cleaning pass — analyze kind=temporal_verify - #96
Merged
Conversation
avfirsov
force-pushed
the
pr/temporal-verify
branch
from
June 15, 2026 07:00
e3a8695 to
91f986a
Compare
The precision backstop for the aggressive AST recall layers: an LLM, grounded in real caller+candidate source, verifies each low-confidence Temporal edge and the pass promotes the confirmed ones, suppresses the rejected ones, and leaves the uncertain ones — never touching register-confirmed 0.9 edges. - resolver.VerifyTemporalEdges: deterministic core over edges at confidence ≤0.65 (confirmed→0.85 visible, rejected→0.1 hidden, uncertain→unchanged), driven by injected TemporalVerifier + TemporalSourceProvider interfaces (fully unit-tested with fakes). - analyzer adapter: llm.Provider-backed verifier (strict-JSON verdict, tolerant parse), file-backed source provider, and a disk verdict cache keyed by model+name+caller-src+target-src for reproducible / CI-safe re-runs. - CLI: `gortex analyze --kind temporal_verify` builds the provider in-process from llm.* config (daemonless); cache in git-ignored .gortex/. (Also carries the allow-list wiring for the analyze path.) - AGENTS.md: corporate-agent instructions for maintaining the allow-list and running the clean pass, with the anonymization protocol. MCP `analyze kind=temporal_verify` handler is a deferred follow-up (corp flow is the daemonless CLI). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The temporal LLM cleaning pass (resolver.VerifyTemporalEdges) had no entrypoint after the CLI analyze host was removed and analyze became MCP-only. Wire it into the analyze dispatcher. - analyzer: add NewLLMTemporalVerifierFromProvider so the handler reuses the MCP server's live llm.Provider instead of constructing a second one from raw config. - mcp: handleAnalyzeTemporalVerify builds the verifier from s.llmService.Provider(), runs VerifyTemporalEdges over the active graph, and returns analyzer.VerifyReportToMap(report). With no LLM configured it returns a clear "LLM provider" error result rather than no-op'ing or panicking. A serverSourceProvider reads node source through the server's own resolveNodePath so paths are correct in single-repo, multi-repo, and worktree layouts. - Register the kind in the dispatch switch + the kind description strings. - Tests: assert routing, the no-LLM-configured clear error, unknown-kind fallthrough, and the source-provider read seam. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
avfirsov
force-pushed
the
pr/temporal-verify
branch
from
June 15, 2026 07:06
91f986a to
f53809c
Compare
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Make the temporal_verify LLM pass actually act, not just report: - VerifyTemporalEdges now PERSISTS promote/suppress verdicts via ReindexEdges, so it works on the disk-backed store (where EdgesByKind hands back copies), not only the in-memory graph - runs the LLM verification OUTSIDE the resolve mutex: snapshot candidates under the lock, verify unlocked, re-acquire only to apply + persist — so a multi-edge pass no longer stalls concurrent resolution/edits - the MCP handler wraps the verifier in the disk-backed CachingVerifier (keyed on provider + source) and Flushes it, so re-runs (CI) skip the LLM - persistence regression test (re-fetch from the store, not the pointer)
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
2 participants
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.
What
Adds an LLM "cleaning pass" for Temporal dispatch edges, invocable via
analyze kind=temporal_verify: it re-checks low-confidence temporal edges against the actual source via an LLM and returns a promote / suppress / keep verdict per edge (it reports verdicts; it does not mutate the stored graph unless wired to).internal/resolver/temporal_verify.go— deterministic coreVerifyTemporalEdgesover the graph, with injectedTemporalVerifier/TemporalSourceProviderinterfaces (no I/O in the core).internal/analyzer/temporal_verify.go— anllm.Provider-backed verifier + file source provider + disk verdict cache +VerifyReportToMap.internal/mcp/tools_analyze_temporal_verify.go— theanalyze kind=temporal_verifyMCP handler: obtains the server's live LLM provider, runs the pass over the active graph, returns the report. With no LLM configured it returns a clear error ("set llm.provider …"), never panics.Result keys:
checked, confirmed, rejected, uncertain, errors, details[], totals. Supportscompact,format,max_bytes.Why
Synthesized temporal edges (wrapper-following, convention, env-default, etc.) land at inferred/speculative tiers — some are wrong. This gives a graph-grounded, opt-in LLM verification to promote the right ones and flag the noise, surfaced through the existing
analyzesurface alongsidesynthesizers/resolution_outcomes.Notes
main.analyzeCLI that no longer exists onmain; this PR wires the feature to the MCPanalyzesurface instead.Tests
Core promote/suppress/keep + error-untouched + skip-placeholder (resolver); report-map, verdict-parse, cache hit/persist (analyzer); handler routing + no-LLM clear-error + source-read-via-resolveNodePath (mcp).
go build ./...,go vet, resolver/analyzer/mcp suites pass.