Skip to main content

The Intelligence Layer

Bitloops doesn't just record what happened — it builds an intelligence layer on top of your codebase. This is the context that makes your AI agents genuinely useful.

What's in the Intelligence Layer?​

Think of it as everything a senior engineer carries in their head, made persistent and queryable.

Structural Intelligence​

Your codebase parsed into a proper dependency graph using Tree-sitter — deterministic, parser-backed, not grep or heuristics.

  • Every function, class, struct, module, interface, and type — extracted with full definitions
  • Every dependency edge — imports, calls, references, inheritance, implementations
  • Cross-file relationships — how symbols connect across your entire codebase

This is what powers blast radius analysis: "if I change this function, what breaks?" Bitloops traverses the full graph to give you a precise answer.

Semantic Intelligence​

Beyond structure, Bitloops understands what your code means:

  • Purpose summaries — what a symbol does in the context of your system, generated through a smart cascade (docstring → LLM summary → template fallback)
  • Similarity detection — finds functions that do the same thing with different names, spots duplicates, identifies divergent forks
  • Pattern recognition — surfaces the conventions your codebase follows, so agents can match them

The similarity engine combines semantic (embeddings), lexical (naming), and structural (AST shape) signals. Results are explainable — not just "these seem similar" but why they're similar.

Historical Intelligence​

Every captured session becomes part of the intelligence layer. This is where the compounding effect kicks in.

When an agent works on your auth module, it can access:

  • Previous sessions that touched that code — what was changed and why
  • Decisions that were made — and the reasoning behind them
  • Alternatives that were rejected — and why they were rejected
  • Planning discussions — constraints, rules, architectural direction

This context flows across sessions automatically. No files to maintain, no memory to curate.

External Intelligence​

Code tells you what exists. External knowledge tells you why it exists.

Bitloops ingests and links:

  • GitHub issues, PRs, and review discussions
  • Jira tickets and epics
  • Confluence design docs and architectural decisions

Each document is versioned and connected to specific commits and artefacts. The links are append-only — refreshing a document preserves the full history.

Test Intelligence​

The relationship between tests and production code, mapped as verification maps:

  • Which tests cover which artefacts (classified as unit, integration, or E2E based on coverage fan-out, not naming)
  • Branch-level coverage gaps
  • How well-tested a given module actually is

Agents see this before making changes — they avoid breaking tests, identify untested paths, and know when to write new ones.

How It's Served​

All of this intelligence is queryable through DevQL — a graph-navigation language:

# What depends on this function?
bitloops devql query "artefacts(symbol_fqn:'auth::validate') → deps(direction:'in')"

# What was this function like at the last release?
bitloops devql query "asOf(ref:'v1.0') → artefacts(symbol_fqn:'auth::validate')"

Agents call DevQL autonomously. They get precise, high-signal context in milliseconds — no scanning your entire repo, no wasting tokens on files they don't need.

The Dashboard provides the same intelligence visually: browse artefacts, explore dependencies, review session history, check coverage.

The Compounding Effect​

Here's what makes the intelligence layer different from a static index: it gets better every day.

  • Week 1: Structural graph + a few captured sessions
  • Month 1: 30+ sessions with reasoning history, linked GitHub issues, semantic understanding of key modules
  • Month 6: A rich institutional knowledge base — every decision, every rejected alternative, every external context from five months of development

The agent working on your codebase in month 6 doesn't start from scratch. It has the accumulated understanding of every session that came before it.

That's not a file you maintain. That's an intelligence layer that builds itself.