Engineering intelligence runtime for AI coding agents. Structured change classification, risk assessment, intent contracts, and audit trails — packaged as a portable skill bundle that works with Claude Code, Codex CLI, OpenCode, Cursor, Windsurf, GitHub Copilot, Gemini CLI, and Cline.
- Intent Contracts — every change opens with a formal commitment (scope, promises, risk cap) and closes with a post-execution verification. No silent edits.
- 8 active runtimes, one install —
npm i -g synarcwrites the right file format for each editor. No rewrites when you switch tools. - 56 specialized skills — engineering, AI-era, product, design, quality, security, data, ML, leadership, and industry verticals. Loaded by intent, not by file.
- Cognition Mesh — multi-role team collaboration: /mesh triggers a coordinator that picks roles, shares working memory, and orchestrates the conversation between them.
- Zero executables, zero network — the pack is pure Markdown + YAML. Audit trails are local files. No data leaves your machine.
Without Synarc, your agent goes straight to edits:
> add a new API endpoint for /billing/usage
Sure — I'll add a route handler, update the OpenAPI spec, and add a test.
[edits 6 files, no plan, no risk check, no verification]
With Synarc, the same request opens an Intent Contract first:
> add a new API endpoint for /billing/usage
WorkType : FEATURE
Risk : HIGH (billing domain — hard floor)
Scale : SMALL
Scope : [src/api/billing/usage.ts, src/api/billing/usage.test.ts, openapi.yaml]
Promises :
- [x] Idempotent on duplicate POSTs
- [x] Backward-compatible response shape
- [x] Audit log entry per call
Verify : scope-check, contract-check, test-run
Proceed? [Y/n] y
[edits 3 files, runs tests, reports verification verdict]
Same intent. Different discipline. The agent proposes a contract, you accept, it executes, it verifies.
- Getting Started
- Usage
- Documentation
- Architecture
- Development
- Troubleshooting
- Security
- Contributing
- License
Any AI coding runtime: Claude Code, Codex CLI, OpenCode, Cursor, Windsurf, GitHub Copilot, Gemini CLI, or Cline. Node.js 18+ is required.
npm i -g synarcThat's it. The installer auto-detects your editor markers (.cursor/, .claude/, .github/, etc.) and writes the right file for each one. If nothing is detected, it asks which editors to install for.
| Your situation | Command | What happens |
|---|---|---|
| Brand new project, no editor configured | synarc |
Interactive picker asks which editors to install |
| Already have one editor, want to add Synarc | synarc add <editor> |
Installs Synarc for that editor, leaves the rest alone |
| Project with multiple editors | synarc add <editor> |
Installs the missing editor only |
| Want everything everywhere | synarc --target all |
Installs Synarc for all 8 supported editors |
| Remove Synarc from one editor | synarc remove <editor> |
Deletes only that editor's config file |
| Migrate from v5 plugin files | synarc migrate-v5 |
Backs up v5 files, installs v6.6.4 |
| Just check the install | synarc verify |
Reports PASS/FAIL per editor |
| Read the lock file | synarc status |
Prints installed editors, no writes |
| Full diagnostics | synarc doctor |
verify + Node version + git version |
<editor> is one of: claude-code, codex, opencode, cursor, windsurf, copilot, gemini-cli, cline.
synarc verifyExpected output on a full install:
[+] PASS Claude Code .claude-plugin/plugin.json (1317 bytes)
[+] PASS Codex CLI AGENTS.md (10174 bytes)
[+] PASS OpenCode AGENTS.md (project) or ~/.config/opencode/AGENTS.md (global) (10174 bytes)
[+] PASS Cursor .cursor/rules/synarc-core.mdc (1429 bytes)
[+] PASS Windsurf .windsurfrules (1784 bytes)
[+] PASS GitHub Copilot .github/copilot-instructions.md (1823 bytes)
[+] PASS Gemini CLI GEMINI.md (12170 bytes)
[+] PASS Cline .cline/skills/<skill>/SKILL.md (56 skills)
Verification: 8 pass, 0 fail of 8 editors.
Exit code 0 on full pass, 1 if anything is missing.
Start a new session in any editor and ask an engineering question. The Synarc classification headers (WorkType, Risk, Scale) should appear in the response.
Once installed, Synarc activates by intent. You can also issue these session commands at any time:
| Command | What Synarc returns |
|---|---|
| what did we change? | Full session ledger (file, delta, risk, scope status) |
| summarize this session | Cognitive summary of the work, decisions, and open questions |
| is this safe to deploy? | Risk delta + explicit YES/NO with reasoning |
| what tests are missing? | All unfilled test gaps for the current WorkType |
full handoff |
Agent handoff block + brain updates for the next session |
For multi-role collaboration, use /mesh (or any phrase like "as a team", "have someone review this", or a 3+ intent signal in one request). The Cognition Mesh coordinator picks roles, defines shared working memory, and orchestrates the conversation between them.
| Doc | Purpose |
|---|---|
| Installation Guide | One-command install + per-editor deep dive + 6 scenarios |
| Architecture | Universal agent skill architecture, 7-layer design |
| Usage | Skill activation, writing new skills, fallback tier usage |
| Compatibility | Capability – runtime matrix for the 8 active editors |
| Migration | Migrating from v5 (per-editor plugins) to v6.6.4 (universal) |
| Enterprise Deployment | Org-scale install, CI/CD, compliance, multi-team distribution |
| Security | OWASP LLM mapping, prompt-injection matrix, capability boundaries |
| CHANGELOG | Release history |
Synarc runs a 7-step pipeline on every interaction:
Classify — Inject — Execute — Log — Aggregate — Checkpoint — Emit
The pipeline is deterministic and runtime-agnostic — the same pipeline operates identically across all 8 supported editors.
Cache architecture (5 tiers, lazy-loaded):
| Tier | Contents | Scope |
|---|---|---|
| 0 | Pack header (AGENTS.md, manifest.yaml) | Session |
| 1 | Core reasoning (synarc-core, cognition-layer, schemas) | Session (~60 KB) |
| 2 | Active domain skill (one of 56) | Task (~10 KB each) |
| 3 | Skill references | Lazy-loaded |
| 4 | Dynamic context (project files, tool outputs) | Never cached |
Scale adaptation (auto-detected, zero configuration):
| Scale | Threshold | Tracking |
|---|---|---|
| NANO | Single file, single purpose | WorkType + risk only |
| MICRO | 2–10 files | CURRENT_STATE.md |
| SMALL | <5k LOC, 1–5 modules | Full brain directory |
| MEDIUM | 5k–50k LOC, team | Full ledger |
| LARGE | 50k–500k LOC, multi-service | Service-boundary tracking |
| ENTERPRISE | >500k LOC, regulated | Compliance audit trail |
| Command | What it does |
|---|---|
npm run sync |
Sync v6 source tree into the universal pack |
npm run validate |
Validate all 56 skills (frontmatter, links, size caps) |
npm run test-fallbacks |
Test all 4 fallback tiers |
npm run measure |
Measure skill sizes against the cache budget |
npm run lint |
Vendor-lockin check + reference integrity check |
npm test |
validate + test-fallbacks |
npm run build |
sync + validate |
npm run precommit |
lint + test |
All scripts run from inside synarc-universal/. PowerShell 5.1+ is required.
| Symptom | Cause | Fix |
|---|---|---|
| Classification headers not appearing in agent responses | Editor can't find the rule / AGENTS.md | Run synarc verify — the FAIL row tells you which file is missing or too small |
| Mesh not triggering | No multi-intent phrase, no /mesh command | Try /mesh explicitly, or write a 3+ intent prompt (e.g., "build a checkout flow" matches PM, designer, frontend, backend, accessibility, performance, SDET, release, security) |
| "Source not found" error | Synarc package not installed or corrupted | Run npm i -g synarc to reinstall |
| Want a clean re-install for one editor | Stale config or version drift | Delete the editor's file (e.g. rm .windsurfrules) and run synarc add <editor> again |
synarc opens an interactive picker in CI |
Stdin is a TTY in the CI runner | Add --yes (or -y) to skip prompts and use the AGENTS.md fallback |
For the full failure-mode matrix, see Installation Guide — Troubleshooting.
| Guard | Status |
|---|---|
| Sandboxed execution | Enabled |
| No network access | Verified |
| No filesystem write outside project | Enforced |
| Deterministic activation | Validated |
| Safe fallbacks on protocol error | Configured |
| Hash-verified integrity (SHA-256) | Active (every skill in .claude-plugin/marketplace.json) |
| Tamper protection | Enabled |
Risk hard floors (no agent may classify a change below this risk, regardless of declared intent):
| Domain | Minimum Risk | Reasoning |
|---|---|---|
| Auth, billing, payments, security | HIGH | Revenue, access, or trust impact |
| Schema change (remove/rename) | CRITICAL | Data integrity + migration complexity |
PRs welcome. See CONTRIBUTING.md for the dev loop, code layout, and how to add a new skill to the pack.
MIT — UpFlame Labs.