Changelog

All notable changes to m1nd are documented here. This project uses Semantic Versioning.

[0.6.2] — 2026-03-26

Added

Universal Knowledge Connectome

The ingestion pipeline now supports scholarly and standards documents alongside code:

• Patent adapter (patent): USPTO/EPO XML with claims, citations, and assignee extraction
• JATS adapter (article): PubMed NLM / JATS Z39.96 scientific article XML
• BibTeX adapter (bibtex/bib): BibTeX bibliography file ingestion
• RFC adapter (rfc): IETF RFC XML v3 with section-level granularity
• CrossRef adapter (crossref/doi): CrossRef API JSON for DOI metadata

Document Router (Auto-Detection)

DocumentRouter auto-detects file format by extension and content heuristics. Used via m1nd.ingest(adapter="auto").

Cross-Domain Resolution — 6 Bridge Strategies

CrossDomainResolver merges outputs from multiple adapters and discovers connections:

All bridges are cross-domain only — same namespace never bridges.

Changed

• README hero stats added to all 7 translated READMEs
• .gitignore updated to exclude .DS_Store and *.bak files
• Architecture docs updated with all new adapters, router, and bridges

Removed

• .DS_Store and antibodies.json.bak removed from repository

Stats

• 10 ingestion adapters registered in MCP CLI
• 131+ tests passing
• 3/3 domains validated bridging via shared DOIs (RFC × CrossRef × BibTeX)

[0.6.1] — 2026-03-25

Fixed

Release and Publish Alignment

This patch release aligns the public release surfaces after the v0.6.0 rollout.

• workspace crates now include the missing crates.io metadata needed for clean publish
• internal workspace dependencies now use explicit published-version constraints
• the release workflow now skips crates.io publish cleanly when CARGO_REGISTRY_TOKEN is missing instead of failing the full release

[0.6.0] — 2026-03-25

Added

Guided Proof State Across Core Agent Flows

Several high-value tools now surface proof_state plus explicit handoff guidance so an agent can tell whether it is still triaging, actively proving, or ready to move into edit preparation.

• seek, trace, impact, timeline, hypothesize, validate_plan, and surgical_context_v2 now participate in a shared proof-state model
• guided outputs now include next_suggested_tool, next_suggested_target, and next_step_hint across the main structural triage and edit-prep paths
• trail_resume now behaves more like continuity orchestration than bookmark restore, returning compact resume hints, next-focus guidance, and tool-aware follow-up

apply_batch Progress, Correlation, and Handoff Signals

apply_batch has been upgraded from a “wait until the batch finishes” write surface into an observable execution flow with stable correlation and final handoff data.

• final outputs now expose batch_id for correlating progress and final result
• progress reporting now includes coarse lifecycle fields such as active_phase, completed_phase_count, phase_count, remaining_phase_count, progress_pct, and next_phase
• phases now act as a structured execution timeline across validate, write, reingest, verify, and done
• progress_events now provide a streaming-friendly event log for the same lifecycle
• live apply_batch_progress SSE emission now happens during execution in serve mode
• the final batch_completed event now carries the batch’s proof_state and next-step guidance

Benchmark Harness Expansion

The benchmark system has been extended so progress UX, guidance, and repair loops can be measured as first-class product behavior, not only token proxy.

• benchmark runs now record execution_origin and source_ref
• long-running flows can distinguish live, replay, and snapshot progress delivery
• the harness records progress event counts, proof-state transitions, recovery loops, and guidance-followed behavior
• the current aggregate warm-graph corpus stands at 10518 -> 5182, or 50.73% savings, while also reducing false_starts from 14 to 0

Changed

Help and Docs Are More Agent-Operational

• help entries now include WHEN TO USE, AVOID WHEN, benchmark-aware guidance, composed workflows, and proof-state handoff cues
• common tool failures are now framed as repair loops with hint/example/next-step guidance
• README, examples, benchmark docs, and public landing surfaces now describe the current guided runtime more accurately

[0.5.1] — 2026-03-24

Two-phase transactional editing for LLM agents. Preview before you write.

New Tools

• edit_preview: Build an in-memory preview of a single-file edit without touching disk. Returns preview handle, source snapshot (hash, bytes, line count), unified diff, and validation report (empty content check, noop detection, ready-to-commit flag).
• edit_commit: Commit a previously created preview to disk. Three safety guards:
  • Confirm guard: confirm must be explicitly set to true (default false).
  • TTL: Previews expire after 5 minutes. Garbage-collected on access.
  • Source hash verification: Re-reads the file at commit time; rejects if it changed since preview.

Technical Details

• EditPreviewState stored in SessionState with agent isolation (agent_id must match).
• Preview handles are single-use — consumed on successful commit.
• Delegates to existing handle_apply() path for actual disk write + graph re-ingest.
• 7 new integration tests covering: happy path, nonexistent file, commit, TTL expiry, source tampering, confirm guard, invalid handle.
• Help system updated with ToolDoc entries and suggest_next chains.

Stats

• 77 MCP tools total on the current audit branch (was 71 before the audit/session additions).
• 342 tests all passing.

[0.1.0] — Initial Release

The first public release of m1nd: a graph-grounded code intelligence engine with Hebbian plasticity, spreading activation, and 43 MCP tools. Built in Rust.

Core Engine (m1nd-core)

• Compressed Sparse Row (CSR) graph with forward and reverse adjacency
• PageRank computation on ingest
• 4-dimensional spreading activation: structural, semantic, temporal, causal
• Hebbian plasticity: Long-Term Potentiation (LTP), Long-Term Depression (LTD), homeostatic normalization
• XLR differential processing: noise cancellation inspired by balanced audio cables
• Hypothesis engine: claim testing with Bayesian confidence on graph paths
• Counterfactual engine: module removal simulation with cascade analysis
• Structural hole detection: topology-based gap analysis
• Resonance analysis: standing wave computation for structural hub identification
• Fingerprint engine: activation fingerprinting for structural twin detection
• Trail system: investigation state persistence, resume, and multi-trail merge with conflict detection
• Lock system: subgraph pinning with sub-microsecond diff (0.08us)
• Temporal engine: co-change history, velocity scoring, decay functions
• Domain configurations: code, music, memory, generic presets with tuned decay half-lives

Ingest Layer (m1nd-ingest)

• Language extractors: Python, Rust, TypeScript/JavaScript, Go, Java
• Generic fallback extractor: heuristic-based for unsupported languages
• JSON adapter: structured data ingestion
• Memory adapter: text corpus ingestion
• Reference resolver: cross-file import and call resolution
• Incremental ingest: re-process only changed files
• Multi-repo federation: unified graph with automatic cross-repo edge detection

MCP Server (m1nd-mcp)

• 43 MCP tools across 7 layers:
  • Foundation (13): activate, impact, missing, why, learn, drift, health, seek, scan, timeline, diverge, warmup, federate
  • Perspective Navigation (12): start, routes, follow, back, peek, inspect, suggest, affinity, branch, compare, list, close
  • Lock System (5): create, watch, diff, rebase, release
  • Superpowers (13): hypothesize, counterfactual, predict, fingerprint, resonate, trace, validate_plan, differential, trail.save, trail.resume, trail.merge, trail.list, seek
• JSON-RPC over stdio: compatible with MCP protocol version 2024-11-05
• Dual transport: framed (Content-Length headers) and line-delimited JSON-RPC
• Auto-persistence: configurable interval (default: every 50 queries) + on shutdown
• Multi-agent support: agent ID tracking, perspective isolation, shared graph
• Tool name normalization: underscores automatically converted to dots (e.g., m1nd_activate -> m1nd.activate)

Performance (measured on 335-file Python backend, ~52K lines)

• Full ingest: 910ms (9,767 nodes, 26,557 edges)
• Spreading activation: 31-77ms
• Blast radius: 5-52ms
• Counterfactual: 3ms
• Hypothesis testing: 58ms (25,015 paths)
• Lock diff: 0.08us
• Trail merge: 1.2ms
• Memory footprint: ~50MB typical

Known Limitations

• Semantic scoring uses trigram matching, not neural embeddings (planned for v0.2)
• No tree-sitter integration yet (planned for v0.2)
• 6 languages with dedicated extractors; others use generic fallback
• Graph is fully in-memory; very large codebases (400K+ files) need ~80MB
• No dataflow or taint analysis (out of scope; use dedicated SAST tools)

Planned: v0.2.0

• Tree-sitter integration for 64+ language support
• Optional embedding-based semantic scoring
• Graph partitioning for very large codebases
• Community detection algorithms
• Performance optimizations for 100K+ node graphs
• MCP Streamable HTTP transport (in addition to stdio)

Planned: v0.3.0

• Distributed graph (multi-machine federation)
• Real-time file watcher integration
• Plugin system for custom extractors and tools
• Graph visualization export (DOT, D3.js, Mermaid)
• Metrics and observability (Prometheus, OpenTelemetry)