Graphenium

Graphenium v0.3.0

Provenance-aware structural memory for AI coding agents.

v0.3.0 upgrades Graphenium from a code indexer to a trust-aware repository intelligence platform with confidence metadata on every edge, CI extraction, policy gates, architecture drift detection, and self-gated CI verification.

New in v0.3.0

Trust Foundation:

• EvidenceSpan with SHA256 hashing and stale detection on every node and edge
• Claim model: structured, agent-facing interpretations of graph facts with supporting evidence
• Resolution report: import resolution, call resolution, ambiguous edges, and unresolved reference tracking
• Trust harness with check_resolution_quality() for CI gate verification

Repository Verification Graph:

• CI config extraction: parse Cargo.toml, package.json, GitHub Actions workflows, and Makefiles
• Test-to-source mappings: inject test coverage and CI job structure into the graph
• Verification planner: 7-tier prioritized verification plans (evidence -> callers -> dependents -> tests -> entry points -> ambiguous edges -> hot paths)
• Risk gates: auto-trigger security reviews on sensitive changes (auth, security, payment, credentials)

Change Safety and Guardrails:

• Policy engine with 6 configurable rules: MinResolution, MaxAmbiguous, MaxUnresolved, MinCallResolution, MaxStale, MinCoherence
• gm check CLI: run trust quality gates for CI with --min-resolution and --max-ambiguous thresholds
• Watch-mode blast radius: live symbol diff display on file changes
• Confidence-aware pathfinding: safest_path prefers high-trust edges over shortest hops
• Self-gated CI: Graphenium builds its own graph and gates on >=80% resolution

Architecture Intelligence:

• Architecture drift detection: community changes, cross-boundary edges, hub migrations, split/merge events
• gm diff --review-plan: generate prioritized verification plans from snapshot comparisons

MCP Tools (10 new):

• resolution_report, ambiguous_symbols, unresolved_references — trust quality inspection
• safest_path — confidence-aware pathfinding (Dijkstra over edge confidence)
• verification_plan, next_files_to_read — change planning
• blast_radius, diff_graph — impact analysis
• agent_change_gate — policy-based gate checks
• review_plan — full review plan generation

CLI (new commands and flags):

• gm check — trust quality gates for CI
• gm doctor --schema/--resolution/--repository — detailed diagnostics
• gm diff --review-plan — verification planning
• gm query --safe — confidence-aware traversal
• gm watch --impact — blast radius display
• gm graph {schema,build-map,test-map,migrate} — repository metadata
• gm snapshot {create,list} — snapshot management
• gm gate --diff — policy gates with diff analysis

Performance and Testing:

• 339 tests across all modules (338 lib + 1 integration)
• Performance budget test for graph load timing
• All tests pass, cargo fmt --check clean, cargo clippy warning-free

Upgrade notes

The v0.3.0 graph format (schema 0.2.0) is backward compatible with v0.2.0. Existing graphs will work without re-extraction, though v3 features (evidence metadata, resolution status) require rebuilding with v0.3.0.

cargo install --path .
gm run . --no-semantic --no-viz  # rebuild with v3 features
gm doctor --resolution           # check your graph quality
gm check                         # run trust gates