Skip to content
MCP AI Agent Guidelines

Changelog

All notable changes to mcp-ai-agent-guidelines are documented here. This project follows Keep a Changelog and Semantic Versioning.

[Unreleased] — Serena integration + memory backend rewrite

Section titled “[Unreleased] — Serena integration + memory backend rewrite”

Breaking changes. The MCP server now treats Serena as the canonical cross-session memory + LSP-grade symbol backend. The custom TOON memory subsystem, the onboarding/orchestration CLI surface, and the per-tool-call disk writes that motivated them are removed.

Removed (MCP tool surface)

Section titled “Removed (MCP tool surface)”
  • agent-memory-* (read/write/fetch/delete) — replaced by Serena memory (mcp__serena__write_memory / read_memory / list_memories), surfaced via the new 🧭 enrichment footer on every instruction response.
  • agent-session-* (read/write/fetch/delete) — workspace state lives in Serena now.
  • agent-snapshot-* (read/write/fetch/compare/delete) — drift tracking moved to Serena’s LSP symbol surface.
  • orchestration-config — orchestration.toml is no longer auto-bootstrapped; defaults stay in memory unless explicitly persisted via model-discover save.
  • project-onboard — wizard-driven onboarding deleted; per-IDE skill hooks now live in the slim mcp-cli onboard skills command.
  • agent-workspace persist / fetch / compare and scope=artifact — reduced to list / read on source files only.

Removed (CLI surface)

Section titled “Removed (CLI surface)”
  • mcp-cli onboard init / status / reset
  • mcp-cli orchestration edit / run-pattern
  • mcp-cli memory list / show / sessions
  • mcp-cli report / export / analytics / docs
  • mcp-cli status / info / dev test utilities

The remaining commands are hooks setup / print / remind-* and onboard skills — see CLI Reference.

Removed (disk writes)

Section titled “Removed (disk writes)”
  • ~/.mcp-ai-agent-guidelines/ — never written.
  • ~/.cache/mcp-ai-agent-guidelines/sessions/ — Hebbian session warmup removed.
  • .mcp-ai-agent-guidelines/config/orchestration.toml auto-bootstrap on missing file.
  • Per-tool-call session-context + memory-artifact writes (now gated behind MCP_LOCAL_MEMORY=true).
  • Workflow checkpoint files.

Added

Section titled “Added”
  • src/serena/client.ts — SerenaClient interface with AdvisorySerenaClient (default, emits structured hints) and ChildSerenaClient (opt-in, spawns Serena via MCP_SERENA_COMMAND).
  • 🧭 Serena enrichment footer on every instruction tool response.
  • New env vars: MCP_SERENA_COMMAND, MCP_SERENA_ARGS, MCP_SERENA_CWD, MCP_LOCAL_MEMORY.
  • npm run test:mcp:serena — opt-in e2e test (MCP_SERENA_E2E=1) that spawns uvx serena and verifies the full code-review → ChildSerenaClient → footer round-trip.
  • New concept page: Serena Integration.

Migration

Section titled “Migration”

If you depended on agent-memory-* / agent-session-* / agent-snapshot-*: install Serena (most MCP hosts already do) and let the host model follow the 🧭 footer. For automated end-to-end runs without a host-level Serena, set MCP_SERENA_COMMAND=uvx and MCP_SERENA_ARGS="--from git+https://github.com/oraios/serena serena start-mcp-server --project <path>".

[0.1.0-alpha.1] — 2026-04-08

Section titled “[0.1.0-alpha.1] — 2026-04-08”

First pre-release of the MCP AI Agent Guidelines server.

Section titled “First pre-release of the MCP AI Agent Guidelines server.”

This release ships a fully functional TypeScript ESM MCP server exposing 101 AI agent skills as callable tools across 18 domain prefixes, with a complete model-routing layer, governance surface, and observability pipeline.

Added

Section titled “Added”
  • 102 skills across 18 domain prefixes (req-, orch-, doc-, qual-, synth-, flow-, eval-, debug-, strat-, arch-, prompt-, adapt-, bench-, lead-, resil-, gov-, qm-, gr-)
  • 21 mission-driven instruction files covering bootstrap, implement, refactor, debug, test, design, review, research, orchestrate, adapt, resilience, evaluate, prompt-engineering, plan, document, govern, enterprise, physics-analysis, meta-routing, onboard_project, initial_instructions
  • Model router with role-name architecture (free_primary, free_secondary, cheap_primary, cheap_secondary, strong_primary, strong_secondary, reviewer_primary) backed by orchestration.toml
  • orchestration-model-discover MCP tool — self-reporting LLM discovery that writes physical model IDs into orchestration.toml by role
  • orchestration-config-read / orchestration-config-write MCP tools for non-interactive config management
  • LlmProvider type extended to cover "xai" and "mistral" in addition to "openai", "anthropic", "google", "other"
  • Observability pipeline with structured JSON logging and ObservabilityOrchestrator
  • 5 evaluation patterns (parallel critique → synthesis, draft → review, majority vote, cascade with fallback, free triple parallel + synthesis)
  • PID homeostatic controller for SLO-driven model tier selection
  • Ant Colony Optimization router (adv-aco-router) with pheromone decay
  • Hebbian router (adv-hebbian-router) for agent pairing reinforcement
  • Physarum router (adv-physarum-router) for self-pruning topology
  • Redundant voter (adv-redundant-voter) with NMR fault tolerance
  • Membrane orchestrator (adv-membrane-orchestrator) for data-boundary enforcement
  • Replay consolidator (adv-replay-consolidator) for orchestrator meta-learning
  • Quorum coordinator (adv-quorum-coordinator) for decentralised task assignment
  • Clonal mutation (adv-clone-mutate) for self-healing prompt recovery
  • Simulated annealing optimizer (adv-annealing-optimizer) for workflow topology search
  • All 30 physics skills (qm-*, gr-*) gated behind physics-analysis instruction and guarded against default invocation
  • scripts/verify_matrix.py — zero-orphan skill coverage enforcement
  • scripts/generate-tool-definitions.py — auto-regenerates src/generated/ from canonical registries
  • Biome for lint + format; lefthook for pre-commit quality gates
  • Coverage reporting with thresholds: statements 83%, lines 84%, functions 87%, branches 75%