Cognis — AI Complexity, Optimization & Blast-Radius Guardrails
Cognis equips VS Code, Cursor, and MCP-compatible AI agents (Antigravity, Cursor Agent, GitHub Copilot, Claude Code, Kiro) with real-time algorithmic complexity analysis, optimization prescriptions (OPT-001..007), loop-amplified blast-radius guardrails (CSIA), and graph-diffusion code retrieval (CSAR) — powered by a single local Rust binary.
What You Get in the Editor
- Live Big-O CodeLens — Inline
Cognis: O(n^2) · space O(1) · 1 optimization badges above functions and methods, click-wired to inspect full complexity and optimization prescriptions.
- Real-Time Guardrail Diagnostics — Automatic Problems panel warnings when a saved change introduces quadratic loops (
OPT-001), N+1 sequential await calls (OPT-002), quadratic string copies (OPT-003), loop-invariant .sort()/regex (OPT-004), multi-pass allocations (OPT-005), unmemoized exponential recursion (OPT-006), unbounded queries (OPT-007), or high CSIA caller blast radius.
- One-Click Workspace & MCP Setup — Downloads the checksum-verified
cognis Rust engine + bge-small-en-v1.5 semantic model, builds .cognis/uckg.db, starts incremental indexd sync, and writes workspace MCP config (.cursor/mcp.json, .vscode/mcp.json, .kiro/settings/mcp.json) without clobbering your other MCP servers.
- Interactive MCP Inspector ("Try MCP Tools") — Test all 12 Cognis MCP tools directly inside the Cognis sidebar panel (Advanced view).
Quick Start
- Open the Cognis icon in the Activity Bar.
- Click Install engine (or point
cognis.binaryPath at a local cognis binary built from source).
- Open a repository folder and click Set up workspace (or run Cognis: Set Up Workspace from the Command Palette).
- Reload your MCP host / agent session once so it picks up the 12 Cognis MCP tools.
analyze_patch_impact — Speculative Shadow AST + CSIA blast-radius evaluation on unsaved code (new_content or file_edits), returning Big-O deltas, OPT-001..007 findings, broken call sites, affected entrypoints, and minimal covering tests in milliseconds.inspect_symbol_complexity — Intra- and inter-procedural time/space Big-O, loop/recursion depth, and optimization advisories for any symbol.trace_blast_radius — Loop-amplified reverse Personalized PageRank (CSIA) ranking impacted callers, entrypoints, and tests.suggest_optimal_refactor — Concrete OPT-001..007 optimization prescriptions grounded in existing utilities/helpers in your repository.
diffuse_context — Flagship CSAR forward-push spreading-activation retrieval recovering full on-path call flow in one round trip.retrieve_context_capsule — Token-budgeted Context Capsule for bugfixes, features, and explanations.discover_symbols — Hybrid FTS5 lexical + ONNX semantic search fused via Reciprocal Rank Fusion (RRF).symbol_lookup — Resolve one symbol by ID, qualified name, or fuzzy name.symbol_search — Fast SQLite FTS5 lexical symbol search.semantic_search — Local bge-small-en-v1.5 vector similarity search.resolve_symbols — Batch-hydrate up to 50 symbol IDs into full source records.dependency_trace — Walk callers (in), callees (out), or both across the call graph.
Plus live MCP resource: cognis://live/guardrails.
Supported Languages
- Code (10 languages via Tree-sitter): TypeScript / JavaScript / TSX, Python, Go, C#, Java, Rust, C, C++, Ruby, PHP
- Structured Artifacts (5 formats): YAML, TOML, SQL, HTML, Markdown
Key Extension Settings
| Setting |
Default |
Description |
cognis.guardrails.enabled |
true |
Enable real-time complexity, OPT-001..007, and blast-radius guardrails in the editor |
cognis.guardrails.codeLens |
true |
Show live Big-O complexity and optimization count CodeLens above functions/methods |
cognis.guardrails.analyzeOnSave |
true |
Automatically run analyze_patch_impact when supported source files are saved |
cognis.mcpHost |
"auto" |
Target MCP host config (auto, cursor, vscode, kiro, claude) |
cognis.mcpConfigScope |
"workspace" |
Write workspace-local MCP config (workspace) or user-level config (global) |
cognis.mcpStdioMode |
"proxy" |
Use lightweight thin stdio proxy (proxy) to deduplicate heavy repository daemons |
cognis.mcpWarmSemanticOnStartup |
false |
Lazy-load (false) vs eager-load (true) the local ONNX semantic model |
cognis.binaryPath |
"" |
Optional explicit path to a custom cognis binary |
Commands
| Command |
Purpose |
| Cognis: Set Up Workspace |
Initialize .cognis/, build the index, start live sync, and wire MCP config |
| Cognis: Analyze Current File Complexity & Impact |
Run analyze_patch_impact on the active editor buffer (including unsaved edits) |
| Cognis: Rebuild Index |
Perform a full cold re-index of the workspace |
| Cognis: Troubleshoot & Repair |
Check health, repair MCP wiring, and restart stalled sync daemons |
| Cognis: Show Diagnostics Log |
Open structured extension diagnostics (diagnostics.jsonl) |
| Cognis: Remove Everything (Prepare for Uninstall) |
Stop daemons, clean .cognis/, remove Cognis MCP entries, and delete managed binaries |
Privacy & License
100% local execution — your source code, .cognis/uckg.db graph, and ONNX embeddings never leave your machine. Released under the Apache-2.0 License.
| |
