Reference
Design Boundaries
Explicit limitations and non-goals. These are not bugs — they are deliberate scope decisions.
The System Does Not Understand Code
The Knowledge Engine retrieves structural facts. It does not:
- Understand semantics or intent within a function body
- Reason about algorithmic correctness
- Detect business logic errors
- Understand comments or documentation prose
Reasoning happens in the LLM consuming the ECO. The knowledge layer provides the evidence that constrains that reasoning.
Known Gaps
| Gap | Impact | Mitigation |
|---|---|---|
| Dynamic imports not tracked | Some dependency edges missing | Graph confidence penalty |
| Monorepo cross-package types without LSP | Symbol resolution incomplete | −25 confidence, escalate |
| Generated files (gql, proto) indexed as-is | Schema drift not detected | Freshness penalty covers this |
| Test files excluded from module graph | Test coverage not structural | Gate results used instead |
| Binary and asset files not indexed | No structural data for non-TS files | Summary-only metadata |
Explicit Non-Goals
- No autonomous commits — the system never commits to the main branch without human approval
- No internet access — all retrieval is local; no external API calls in the critical path
- No multi-repo awareness — each repo has its own index; cross-repo analysis is out of scope
- No historical blame analysis —
git blamepatterns are not used for evidence scoring - No AI-generated tests — the system validates against existing tests; it does not write them
When to Distrust the System
Trust the ECO less when:
- Coverage dimension is at
escalateorblock - The codebase has heavy use of dynamic patterns (
require(variable), factory functions) - The task involves a language without tree-sitter support
- The working tree is dirty