The Mythos Engine
A local-first, provider-agnostic CLI runtime. Mythos goes beyond simple LLM wrappers by enforcing strict file verification, multi-provider routing, skills, receipts, and MCP/external-agent SWD boundaries.
1. Strict Write Discipline (SWD)
Intent Capture
Captured AI output is treated as a Proposed Intent, not a direct command. Each node must contain explicit metadata: Target Path, Operation, and File Content.
Snapshot Sync
The system performs an instantaneous sync with the Host Filesystem. Actual state overrides internal AI assumptions, creating a verified ground truth baseline.
Correction Turn
If an intent references invalid system state, execution is blocked. Real-time filesystem state is re-injected into the AI context, triggering a self-correction loop.
Atomic Verification
Intents are promoted to COMMIT only when verification yields zero divergence. Ensures atomic, validated writes under controlled conditions.
2. Multi-Provider Orchestration
Fallback-Aware Routing
Mythos uses provider health, latency, cost, retry behavior, and fallback settings to choose between configured providers. If the active provider fails with a retryable error or stalls, the orchestrator can move to another configured provider instead of treating one API failure as the whole session.
- Adaptive Performance Scoring
- Exponential Backoff & Circuit Breakers
- 15-second Stream Watchdog Timeout
3. Auto-Healing TDD Loop
Closed-Loop Execution
When the --test-cmd flag is passed, Mythos automatically runs your test suite after applying SWD verified file changes.
Error-Driven Fixes
If tests fail, the stdout/stderr is captured, truncated safely, and fed back into the agent context for an automatic correction turn.
Anti-Thrashing Guards
Safety mechanisms detect identical test outputs across retries or regression in failure counts to prevent runaway token consumption.
Budget Enforced
Every test healing loop respects the global SessionBudget, graceful saving state if token or turn limits are approached.
4. YAML Skills Protocol
Inject targeted operational context dynamically using local SKILL.md files. Skills shape repo behavior, review expectations, provider preferences, and budget settings for the active session.
5. MCP and External-Agent SWD Boundary
Local MCP Tool Server
MCP clients can run mythos mcp and call SWD dry-run/apply, receipts, and skill inspection tools over local stdio.
Automation Output
--json returns machine-readable status, approved and rejected action counts, verification results, and receipt metadata for scripts and CI-style tooling.
Security Defaults
Sensitive paths are blocked by default. Deletes and command-surface files require explicit opt-in.
Receipts
Successful non-dry-run applies write local SWD receipts with external agent metadata, file hashes, rollback status, and drift verification data.
Upgrade Your Workflow
Deploy the Mythos Router in your development environment to transform generative AI into a reliable, verifiable, and cost-efficient engineering partner.