A sovereign, deterministic AI coding harness with neural context intelligence, dynamic agent swarms, adversarial verification gates, and a 7-layer provenance memory system — for engineering at scale.
Every module in AETHER is purpose-built. From neural context slicing to adversarial falsification gates — no hollow abstractions, every engine is deterministic and verified.
Personalized PageRank over AST symbol graphs ranks every file by relevance. 0/1 Knapsack allocates tokens — full source, skeleton, or evicted.
src/core/contextEngine.tsTCCI score routes tasks to 1 engineer or a DAG swarm. Exclusive file-write lease locks prevent parallel edit collisions across 117 specialists.
src/core/swarmEngine.tsSix hostile challengers — Architecture, Code Breaker, Security Attacker, Performance Critic, Regression Investigator, Judge — require executable reproducers.
src/core/adversarialEngine.tsProvenance-bound memory from Session through Verified Knowledge to Failure Memory. Auto STALE_REVALIDATION_REQUIRED on file mutation.
src/core/memoryEngine.tsEach engine is a self-contained, testable module — no cross-contamination, no magic. Together they form the complete AETHER reasoning pipeline.
AST-aware symbol graph ranking with Personalized PageRank and token budget knapsack compression.
Dynamic DAG swarm topology with Task Complexity & Coupling Index, exclusive file-lease locks, and conflict resolution.
Deterministic AST call-graph traversal + git co-change coupling with automatic Targeted Verification Plan synthesis.
Six hostile challengers testing architecture, security, performance, concurrency, and regression boundaries with executable reproducers.
Shannon entropy secret scanning, prompt-injection spotlighting, MCP tool manifest guard, egress policy, and integrity-chained audit log.
7-layer store bound to SHA-256 file hashes with automatic staleness downgrade on mutation. Session through Verified Knowledge.
Scientific hypothesis debugging state machine. Hard gate requiring non-zero failing reproducer exit code and Bayesian falsification before minimal fix.
Multi-trial benchmarking with mean, standard deviation, 95% CI, and regression tracking across all 711 components.
Continuous quality scoring across 5 deterministic rubrics. Auto-refines weak specs and re-tests to 99.6+/100 quality gate.
Closed-loop test-score-refine cycle. Evaluates all agents and skills deterministically, auto-refines weak specifications to target score.
Dynamic agent creation, skill composition pipelines, MCP server lifecycle, SemVer plugin resolution with topological sort & cycle detection.
From Principal Systems Architect to Security Attacker — each agent has a strict role, scope, responsibility list, handoff rules, and verification gate.
Skills activate via 28-token AST predicates. Zero tokens consumed until the activation predicate matches the active file graph.
Every skill carries a deterministic activation predicate, required information list, execution steps, useful tools, expected outputs, verification criteria, and failure-case handling — all defined in structured frontmatter. Skills are verified against AETHER's 0-filler quality gate before entering the registry.
Domain packs (core-architecture, distributed-systems, cloud-native, etc.) are loaded on demand. Installing 500 skills adds 0 tokens to the base prompt until an activation predicate matches the active AST/file graph.
DynamicEcosystemRegistry.composeSkills() chains existing skills into higher-order composite pipelines with deterministic scoring, combined trigger keywords, and required permission enforcement.
Each program composes existing AETHER engines rather than reimplementing them — adding the specialized reasoning layer each domain demands.
Research-first TDD gate with deterministic AST skill discovery, Red-Green-Refactor enforcement, and debugging hypothesis ledger.
Isolated, permission-controlled reverse-engineering. Application architecture extraction, dependency discovery, and automated technical reporting.
Goal-to-plan resumable execution layer. Dependency-structured action plans with deterministic verification predicates per step and crash recovery.
Unified browser engine with A11y ref snapshots, deterministic CSS/ARIA DOM locators, and vision coordinate fallback — ranked by determinism.
Perception → Reasoning → Action → Verification loop. OmniParser-style UI element grounding with confidence-floor abort guarantees.
Provider abstraction for Playwright cross-browser, Lightpanda CDP (BSL-1.1, out-of-process only), and bare Chromium backends.
ComfyUI-style node graph (clean-room data model) + Three.js-style 3D scene graph. Topological scheduling with cycle detection.
SHA-256 content-addressed hashing, Merkle tree continuity checkpoints, FNV-1a fast hashing, and integrity-chained audit logs.
Full JSON-RPC 2.0 MCP server engine. 8 AETHER tools, capability negotiation, permission enforcement, and structured error codes.
The selective installer deploys agents, skills, hooks, and workflows to 5 major AI coding harnesses — and the MCP server. Under 650 idle tokens.
Install everything across all supported harnesses with one command.
Every workflow is a 5-stage pipeline with explicit dependencies, failure handling, and completion criteria.
AETHER's harness layer abstracts every major AI provider through a unified interface. Swap models without rewriting agents, skills, or workflows.
Native first-class support via Anthropic SDK. Full extended thinking, tool use, vision, and the 1M context window. Default for AETHER agents and all Fable-tier workloads.
Vertex AI and Generative Language API integrations. Supports multimodal reasoning, tool calling, and function-calling schemas with adapter parity.
OpenAI API adapter for GPT-4o, reasoning models (o3, o4-mini). Structured outputs and function-calling interface through the OpenAI SDK wrapper.
Fully offline via Ollama or vLLM inference servers. Supports Llama 4, DeepSeek, Qwen, Mistral, and any OpenAI-compatible endpoint. No data leaves your network.
Four lifecycle hooks intercept every prompt, tool use, edit, and compaction with non-blocking execution budgets.
Scan untrusted files for indirect prompt injection, strip high-entropy secrets, compute AST PageRank context slice.
Verify exclusive file-write lease held. Block destructive shell AST patterns. Bash, Shell, Edit, Write matchers.
Invalidate stale Layer-3 memory hashes for modified files. Trigger targeted AST-mapped unit test on every edit.
Persist Zero-Loss Continuity Merkle checkpoint with falsified debugging hypotheses and verified DAG task state.
License-audited adapters with strict isolation boundaries. Permissive ecosystems get native in-process adapters; copyleft projects are sandboxed to out-of-process HTTP/CDP only.
Security is structural. From untrusted content spotlighting to license boundary enforcement, AETHER's threat model is explicit.
All repo files, web pages, and MCP payloads pass through the spotlighting gate. Injection directives are quarantined inline — never reach the model as live instructions.
Untrusted tool registration is impossible. Only tools in the local AETHER_MCP_TOOLS manifest are accepted. Permission denials emit structured JSON-RPC -32001 errors.
Permissive (MIT, Apache-2.0) ecosystems get native adapters. Copyleft (AGPL-3.0, GPL-3.0, BSL-1.1) are strictly out-of-process HTTP/CDP only.
Detects high-entropy tokens (Shannon > 4.35) matching credential signatures. Zero hardcoded secrets policy enforced on every commit.
Shell commands are gated by configurable egress rules. Only whitelisted hosts reachable. Enforced at the tool-call boundary.
Every gate decision is recorded with SHA-256 hash chain. Tampering detectable. Complete forensic trail from security scan through adversarial verdict.
Every dependency, inspiration, and ecosystem is credited with SPDX license identifiers. No uncredited borrowings.
Open source. Apache-2.0. 117 agents. 540 skills. Zero filler.