Synthesis — Release History¶
Synthesis is a knowledge infrastructure platform for AI-augmented development. It indexes workspaces (code, documentation, PDFs, media, Notion), builds a multi-layer knowledge graph, and exposes everything through a CLI, an MCP server for AI agents, and an LSP server for IDEs.
Repository: github.com/exoreaction/Synthesis
Current Release: v1.29.0 (April 21, 2026)¶
60+ CLI commands · 11 MCP tools · 4,356 tests
The v1.29.0 release adds Notion as a first-class workspace source and git signal analysis. Synthesis can now index Notion workspaces alongside code and documentation — and detect when documentation and code have drifted apart.
What's new in v1.29.0¶
Notion workspace source (6-phase implementation) — Notion pages become part of the Synthesis knowledge graph as virtual filesystem paths. Full and incremental syncs via BFS traversal; block-to-Markdown conversion covering all block types; rate-limited API client (3 req/s, automatic pagination, 429 retry).
Three new health signal types emerged from building this:
- W022 notion-stale — a Notion page references a code entity that no longer exists
- W023 notion-orphan — a Notion page has no corresponding code referent
- W024 notion-conflict — a Notion page and the codebase contradict each other on the same fact
These signals make documentation drift visible before it misleads an agent. See the blog post: When Your Agent Can Finally Read the Room
Git signal analysis — four history-based signals using exponential decay (half-life 180 days):
synthesis hotspots # files by temporally-decayed commit churn
synthesis archaeology --since 90d # knowledge buried in old code paths
Co-change coupling and bus factor analysis also included. Inspired by temporal decay techniques in Repowise; all computation is local.
v1.28.0 — Onboarding Improvements (April 11, 2026)¶
4,300+ tests
Explicit API key guidance when AI features are unavailable. When synthesis detects AI is enabled but no ANTHROPIC_API_KEY is configured, it now prints an actionable message with the exact command to fix it — instead of producing confusing silent failures. The same guidance surfaces during synthesis init guided setup.
v1.27.1 — Memory Maintenance (April 10, 2026)¶
Two new commands for keeping ExoCortex memory topic files healthy:
synthesis topic-health — scans ~/.claude/projects/.../memory/*.md, queries FTS5 session hits per keyword, prints a HOT/WARM/COLD table with delta from baseline. Hotness score weights recency (60%) and age (40%).
synthesis topic-triage — four-dimension scored triage (Recency, Recurrence, Actionability, Staleness) → ARCHIVE / PRUNE / UPDATE / KEEP suggestions for top-5 files. Advisory only; no writes unless --auto is passed.
--auto uses dual-threshold logic (24h + 5 sessions) via ConsolidateState — the same pattern as ReflectState. Nightly cron at 02:45 runs topic-triage --auto after synthesis maintain at 02:30.
Three bundled Claude Code skills added for workshop users who don't have Synthesis source access.
See the blog post: Agent Memory Rots
v1.26.0 — Skills Graph & Subagent Session Linking (March 14, 2026)¶
synthesis skills-graph — a self-contained interactive Cytoscape.js visualization of all ~/.claude/skills/ with three modes:
synthesis skills-graph # skill dependency graph (199 skills, 30 clusters)
synthesis skills-graph --mode workspace # cross-repo dependency view
synthesis skills-graph --mode modules # directory-level module graph
Per-cluster colour palette, collapsible sidebar, search box, hover tooltips, keyboard navigation. FCose layout with CDN fallback for offline use.
Parent-child subagent session linking (V19 Flyway migration) — parent Claude Code sessions that spawn subagents are now linked in the session index, enabling full chain tracing via synthesis sessions.
Reflect quality fixes — noise filter, version batching, TTY detection for scan progress, session freshness checks.
knowledge.yaml — Synthesis now ships its own KCP manifest, making it self-indexing.
v1.24.0 — Skill Reflection (March 9, 2026)¶
4,255 tests
synthesis reflect — closes the dispatch→reflect learning loop. Analyzes recent Claude Code sessions and automatically creates or updates skill YAML files in ~/.claude/skills/.
Signal extraction heuristics look for five patterns in session text: CORRECTION (AI was corrected), EXPLICIT_RULE ("always"/"never" instructions), DOMAIN_TERM (recurring domain vocabulary), TOOL_PATTERN (command sequences), and WORKFLOW_STEP (procedural sequences). Patterns scoring ≥ 5.0 update existing skills; patterns without a match create new reflect-*.yaml files.
synthesis reflect # update skills from recent sessions
synthesis reflect --dry-run # preview changes without writing
synthesis reflect --since 7d # limit to last 7 days of sessions
synthesis reflect --max-new 5 # cap new skill files per run
ReflectState tracks lastReflectedAt in ~/.synthesis/reflect-state.json for fast short-circuit (< 10ms) when skills are already up to date. MCP tool reflect exposes the same logic to AI agents.
v1.23.0 (March 7, 2026)¶
58 CLI commands · 11 MCP tools · 4,230 tests
The v1.23.0 release completes the multi-agent awareness layer by adding synthesis dispatch — a single command that composes skill matching, file search, and team conflict detection into a ready-to-use agent dispatch plan.
What's new in v1.23.0¶
synthesis dispatch — given a task description, returns the skills to load, files to pre-read, any team conflicts, and an estimated token cost. Designed for injection into an agent's spawn prompt to skip the 40–60% blind retrieval phase.
synthesis dispatch "fix OAuth2 token refresh" # full table
synthesis dispatch "query" --compact # single line for spawn prompt
synthesis dispatch "query" --json # machine-readable JSON
synthesis dispatch "query" --no-team # skip conflict check
Compact output (for --compact):
skills:kcp-mcp,legacy-modernization | files:TokenRefresher.java,OAuth2Config.java | conflicts:none | ~4200 tokens
The dispatch MCP tool exposes the same logic to AI agents via JSON-RPC. All failures are graceful — missing skills directory, index unavailable, or no active team → empty result, no crash.
v1.22.0 — Multi-Agent Awareness (March 7, 2026)¶
synthesis skills match and synthesis team-context — skill discovery and team briefing commands for Claude Code multi-agent sessions.
synthesis skills match "parse Gerber format" # ranked skills from ~/.claude/skills/
synthesis skills list # all skills with descriptions
synthesis team-context # briefing for active agent team
synthesis team-context --compact # single line for Agent spawn prompt
skills match scores ~/.claude/skills/*.yaml using trigger phrases (5 pts), name/description (2 pts), and instructions (1 pt). team-context enriches each task with related files from the index, skill recommendations, and conflict warnings when two tasks share files. MCP tools: match_skills, team_context. 4,177 tests.
v1.21.0 — Episodic Memory (March 3, 2026)¶
55 CLI commands · 8 MCP tools · 4,170 tests · 18 database migrations
The v1.21.0 release completes the three-layer AI memory model by adding episodic memory — indexed access to Claude Code session history. Synthesis has always covered Layer 3 (workspace knowledge graph). It now covers Layer 2 as well.
| Layer | Type | Source | Synthesis command |
|---|---|---|---|
| 1 | Working memory | Context window | (current conversation) |
| 2 | Episodic memory | Session transcripts | synthesis sessions |
| 3 | Semantic memory | Workspace knowledge graph | synthesis search, relate, impact |
What's new in v1.21.0¶
synthesis sessions — a new subcommand backed by a standalone SQLite module that indexes Claude Code session JSONL files from ~/.claude/projects/. First scan: 2,971 sessions in 109 seconds. Subsequent scans are near-instant (incremental, skips unchanged files).
synthesis sessions scan # index ~/.claude/projects/
synthesis sessions search "authentication" # full-text search across all sessions
synthesis sessions list --since 7d # recent sessions
synthesis sessions get <session-id> # full detail
The sessions MCP tool extends this to AI agents — Claude Code can now search its own conversation history through the same Synthesis MCP server that already provides workspace knowledge. No second MCP server required.
See the blog post: Working Memory, Episodic Memory, Semantic Memory. Your Agent Has One.
v1.18.2 — Session Lifecycle Integration (February 28, 2026)¶
Session lifecycle hooks, synthesis hooks generate for automated hook scaffolding, synthesis session-context for reading active session state, and synthesis claude-md for CLAUDE.md management. 4,107 tests.
v1.13.x — Code Knowledge Graph (February 22, 2026)¶
CKG phases 1–4 landed in v1.12.2, covering class-level code graphs with symbol extraction, import analysis, and cross-file dependency edges. v1.13.0 fixed a rebalance false-positive (issue #209) and integrated .synthesisignore into health checks. 3,893 tests at GA.
v1.12.0 — 9-Phase Maintain Orchestrator (February 21, 2026)¶
synthesis maintain became a rich orchestrator: file tracking, snapshot management, knowledge edge scanning, reconciliation, activity logging, workspace sync, and git integration — all in sequence, with progress reporting. Guided synthesis init added.
v1.11.0 — Workspace Self-Organization (February 20, 2026)¶
synthesis health (structural audit with health score 0–100), synthesis prune (stale file removal), synthesis sweep (orphan cleanup), and the directory identity system that enables workspace self-organization via .synthesis.md companion files.
v1.10.x — Knowledge Integrity and Discovery (February 19–20, 2026)¶
synthesis discover (surface unknown structure), synthesis validate (drift detection between index and filesystem), and unified knowledge graph with confidence metadata. Gap detection flags areas where the knowledge graph is sparse.
v1.9.x — Test Expansion and Operational Maturity (February 18–19, 2026)¶
Test count grew from 1,054 to 2,325 during this series (v1.9.1–v1.9.5). Staging pipeline improvements, project-level skill generation, MCP schema fixes, and synthesis enrich targeting. The architectural security report landed in v1.9.4.
v1.8.x — Staging Pipeline (February 17–18, 2026)¶
A multi-phase staging pipeline for file ingestion and classification: staging ingest, staging route, staging rename. V8 Flyway migration for report cache. Vision resize for large images.
v1.7.x — Dashboard, Research, Reports (February 16–17, 2026)¶
synthesis dashboard, synthesis research (deep multi-source investigation), synthesis report (executive report generation), synthesis org enrich, and git integration in maintain. Credential store for API key management.
v1.6.x — Executive Summaries (February 16, 2026)¶
synthesis summary with eight role-specific perspectives (engineering manager, architect, executive, DevOps, product manager, security, developer, general). Client-to-codebase auto-discovery. V5 Flyway migration for summary cache.
v1.5.x — Sub-Workspaces and Bundled Skills (February 16, 2026)¶
Sub-workspace support for multi-project configurations. Bundled Claude Code skills installable via synthesis export-skills. Smart exclusion defaults. V4 migration.
v1.4.0 — File Tracking and Change Reporting (February 16, 2026)¶
synthesis track (hash-based file movement detection with 7-day safety period), synthesis changelog (cross-workspace change reports with smart filtering). V3 migration: 6 new tables.
v1.1.0 — Protocol Servers (February 15, 2026)¶
MCP server (JSON-RPC 2.0 over stdio, 7 tools), LSP server (LSP 3.17, document links, hover, diagnostics, workspace symbols), air-gapped edition mode, self-update mechanism. Three deployable JARs: CLI, MCP, LSP.
v1.0.0 — Genesis (February 14, 2026)¶
Core indexing engine (Apache Lucene 10.1.0), 6 file analyzers (code, markdown, YAML, PDF, image, video), organization registry, dependency graphs, and 282 tests. Built to solve the lib-pcb problem: 8,934 files generated in 11 days of AI-assisted development with no tool to navigate the output.
Version Timeline¶
| Version | Date | Highlights |
|---|---|---|
| v1.0.0 | Feb 14 | Core indexing, search, CLI, org intelligence, 282 tests |
| v1.0.1 | Feb 14 | Distribution, skill generation, install scripts |
| v1.0.2 | Feb 14 | Media support (image, video, PDF), directed synthesis |
| v1.0.3 | Feb 14 | Bundled ffprobe binaries |
| v1.1.0 | Feb 15 | MCP server, LSP server, update system, air-gapped mode |
| v1.2.0 | Feb 15 | AI features (enrich, explain, perspectives), local media enrichment |
| v1.2.1–v1.2.3 | Feb 15 | Unified workspace system, status aggregates |
| v1.4.0–v1.4.1 | Feb 16 | File tracking, change reporting, V3 migration |
| v1.5.0–v1.5.3 | Feb 16 | Sub-workspaces, bundled skills, visual workspace tree |
| v1.6.0–v1.6.1 | Feb 16 | Executive summaries (8 perspectives), client auto-discovery |
| v1.7.0–v1.7.8 | Feb 16–17 | Dashboard, research engine, report engine, git integration |
| v1.8.0–v1.8.4 | Feb 17–18 | Staging pipeline, vision resize, report fixes |
| v1.9.0–v1.9.12 | Feb 18–19 | Report output config, 2,325 tests, staging integration |
| v1.10.0–v1.10.6 | Feb 19–20 | Knowledge integrity, discover, validate, unified knowledge graph |
| v1.11.0–v1.11.3 | Feb 20 | Health command, prune, sweep, directory identity system |
| v1.12.0–v1.12.2 | Feb 21–22 | 9-phase maintain, guided init, Code Knowledge Graph |
| v1.13.0–v1.13.1 | Feb 22 | CKG dogfooding fixes, 3,893 tests |
| v1.18.2 | Feb 28 | Session lifecycle hooks, session-context, claude-md, 4,107 tests |
| v1.21.0 | Mar 3 | Episodic memory: Claude sessions module, FTS5 search, 4,170 tests |
| v1.22.0 | Mar 7 | Multi-agent awareness: skills match, team-context, match_skills + team_context MCP tools, 4,177 tests |
| v1.23.0 | Mar 7 | Agent dispatch planner: synthesis dispatch, dispatch MCP tool, 4,230 tests |