v2.7.0 — Session Persistence

Session Persistence Tools

Four new MCP tools for preserving agent context across sessions:

New Tools

• save_session_bookmark / restore_session_bookmark — Capture and recover orchestrator state (working set, summary, findings, next steps). SessionBookmark nodes link to code graph nodes via [:REFERENCES].
• save_session_note / recall_session_notes — Durable observations with vector embeddings and semantic recall. Supports categories (architectural, bug, insight, decision, risk, todo), severity levels, and auto-expiry. Notes link to code nodes via [:ABOUT] and to bookmarks via [:HAS_NOTE].

Swarm Enhancements

• New session_context pheromone type (8hr TTL) for marking session working sets
• sessionId field added to swarm_pheromone and swarm_sense for session-scoped coordination

Implementation Details

• Semantic vector search on session notes (text-embedding-3-large, 3072 dimensions) with fetch multiplier for multi-project isolation
• Lazy vector index creation (IF NOT EXISTS) — existing installations work without re-parsing
• Filter-based recall by category, severity, session, or agent when no semantic query provided

v2.6.2

Changes

• fix(parse): Default async parameter to true for background parsing by default
• fix(parse): Use z.coerce.boolean() to handle string "true"/"false" values from MCP clients

Full Changelog

v2.6.1...v2.6.2

v2.6.1

Fix

• MCP protocol corruption fix: Redirected all console.log calls to console.error (stderr) to prevent diagnostic logging from polluting the MCP JSON-RPC stdout channel
• Added quiet: true to dotenv config to suppress warnings when .env file is missing

v2.6.0

What's Changed

Swarm Coordination Improvements

• Atomic claim_and_start: New default action eliminates race window between claim and start operations
• Retry logic: Automatic retry with exponential backoff when losing race for task claims
• Recovery actions: New abandon and force_start actions for handling stuck tasks
• Better error messages: Task state diagnostics help debug coordination issues

Graph-Based Task Targeting

• targetNodeIds fallback: Automatically resolves node IDs from file paths when not provided
• Flexible path matching: Uses ENDS WITH to handle absolute vs relative path mismatches
• Resolved targets in response: Claim response now includes resolved graph node targets
• Updated worker instructions: Workers now prefer nodeId-based graph traversal

Performance & Context Optimization

• Reduced debug logging: Removed verbose intermediate logs, kept only startup/stop/error
• Slimmer tool descriptions: Reduced context window usage with concise descriptions
• Leaner response payloads: Essential fields only in swarm tool responses

Upgrade Notes

No breaking changes. The claim_and_start action is now the default for swarm_claim_task.

v2.5.4

Changes

• fix(natural-language_query): bump to gpt-4o
• feat(swarm): add task cleanup and orchestrator responsibilities docs

v2.5.3

Fixes

Worker Instructions Improvements

• Add graph tools (traverse_from_node, search_codebase) to worker workflow for understanding code context
• Add pheromone usage (swarm_sense for conflict detection, swarm_pheromone for modifying/completed states)
• Change run_in_background from true to false (required for MCP tools)
• Generate unique agent IDs for each worker
• Include full step-by-step blackboard protocol

Task Claiming Fix

• CLAIM_NEXT_TASK_QUERY now considers 'blocked' tasks (not just 'available')
• Tasks stuck as 'blocked' can now be auto-claimed if their dependencies are complete

v2.5.2

Bug Fixes

Critical Swarm Coordination Fixes

• Race condition prevention: Added APOC locking to swarm_claim_task to prevent multiple workers from claiming the same task simultaneously
• Task completion fix: swarm_complete_task now works when called directly after claim (without requiring start action first)
• Query result fix: Fixed COMPLETE_TASK_QUERY and APPROVE_TASK_QUERY returning empty results when no dependent tasks need unblocking

Previous fixes in 2.5.1

• Fix wrong vector index name
• Fix pheromone timestamp field consistency
• Add missing :DEPENDS_ON and :TARGETS relationships
• Fix task status not being set to blocked when dependencies exist

v2.5.1

Bug Fixes

• swarm-orchestrate: Fix wrong vector index name (code_embeddings -> embedded_nodes_idx)
• swarm-orchestrate: Fix pheromone timestamp field inconsistency (use p.timestamp consistently)
• swarm-orchestrate: Add missing p.halfLife to pheromone creation for proper decay
• swarm-orchestrate: Add missing :DEPENDS_ON and :TARGETS relationships in task creation
• swarm-post-task: Set status to blocked when dependencies exist (was always available)
• swarm-claim-task: Properly require all dependencies complete before claiming
• swarm-get-tasks: Fix worker activity query to use p.timestamp instead of p.updatedAt

v2.5.0 - Swarm Task Queue

What's New

Swarm Task Queue System

Added explicit task management to complement pheromone-based coordination. Agents can now post, claim, and complete tasks through a shared blackboard.

New Tools:

• swarm_orchestrate - Decompose complex tasks and spawn worker agents automatically
• swarm_post_task - Post tasks to the blackboard with dependencies
• swarm_claim_task - Claim/start/release tasks with atomic updates
• swarm_complete_task - Complete/fail/review workflow with auto-unblocking
• swarm_get_tasks - Query tasks with filters, stats, and progress tracking

Features:

• Dependency management - tasks auto-unblock when dependencies complete
• Priority-based selection - agents claim highest priority available task
• Race condition protection - atomic Neo4j updates prevent conflicts
• Progress tracking - monitor completion percentage and active workers
• Review workflow - tasks can be submitted for review before completion

Documentation

• Updated README with config-driven extensibility emphasis
• Added swarm skill file (skills/swarm/SKILL.md) for optimal agent execution
• Documented custom framework schema support

Install

npm install -g code-graph-context@2.5.0

v2.4.5

Changes

• Smaller package size: 785KB → 656KB unpacked (-16%)

  • Removed orphaned build artifacts (graph-v2.js, typescript-parser-v2.js)
  • Added prepublishOnly script to prevent future orphans
  • Trimmed README from 45KB → 29KB

• Documentation fixes:

  • Updated Claude Code config path to ~/.claude.json with --scope user
  • Fixed stale typescript-parser-v2.ts reference
  • Updated response format examples to match actual code
  • Consolidated duplicate troubleshooting sections