Canadian AI — Research Framework
Grammar Is All You Need
A grammar-first framework for reliable, structure-aware program transformation in AI coding agents. Progressive capability loading, validation-first architecture, and seamless integration with existing tooling.
Design Principles
Grammar-First
AST / LSP — not regex
Type-Safe
Typed inputs & outputs
Reversible
Git checkpoints + patches
Sandboxed
E2B / Modal isolation
Minimal Surface
Six operations, no more
The Token Math
Stop paying the inference bill twice.
Traditional harnesses generate code, then spend a second full inference pass refactoring what they just wrote. GIRL plans the change once — loading capabilities on demand and working on the syntax tree — so one pass does the work of two.
Pipeline comparison — same task
two passes one pass0%
Fewer tokens / task
8.2k → 2.1k end-to-end
0×
Inference pass
down from 2× speculative
Tokens per refactor task
Illustrative per-task totals including generation. GIRL returns compact symbol handles instead of raw file blobs and loads operations only when needed.
0 tok
Capability context
vs 4,500 loaded upfront
The Real Cost of AI Refactoring
When AI writes and then rewrites, you pay the inference provider twice.
Search interest in refactoring has grown over 300% since 2019 — driven largely by teams cleaning up AI-generated code. The current loop: generate with an LLM, then pay a second full inference pass to refactor what the first pass produced. GIRL is designed to collapse that loop into one well-planned, structure-aware transformation.
“Refactoring queries on Google Trends have climbed steadily since 2021, peaking in 2025–2026 as AI code-generation adoption hit mainstream scale.”
Google Trends — 5 year view, 2021 – 2026
2×
Double inference spend
AI generates code, then pays inference again to refactor what it wrote — two full LLM passes for one net change.
96%
Token reduction possible
GIRL on-demand loading cuts baseline capability context from 4,500 tokens to ~200 — before the transformation even starts.
1 pass
Our goal
Structure-aware planning eliminates redundant generation. One well-planned pass replaces two expensive speculative ones.
Core Control Flow
Reversible Pipeline
Every transformation passes through a five-phase pipeline. Each phase produces immutable artifact handles — nothing is discarded, everything is revertible.
Pipeline — Click a node to expand
Minimal Operation Surface
Six Reversible Primitives
Operation
Purpose
Returns
inspect_symbol
Resolve a symbol, its definition, references, and dependency neighbourhood
Symbol summary + handles
query_ast
Ask structural questions without reading whole files
AST-node summaries, selectors
plan_refactor
Produce a change plan with blast radius, strategy, validators
Plan object with checkpoints
apply_refactor
Execute one planned transformation step via AST/LSP/codemod
Patch set + provenance
run_validation
Run type-check, lint, unit/integration suites in sandbox
Validation report + artifacts
rollback
Restore workspace to prior checkpoint or inverse patch
Rollback confirmation + scope
Implementation
Adapter Architecture
Tool
Purpose
Tier
Tree-sitter
Parse & AST traversal
LSP
Symbol lookup & references
Codemod
Transformation execution
MCP
Capability negotiation
Semgrep
Pattern matching & fixes
ast-grep
AST queries & rewrites
Git
Checkpoints & rollback
Comby
Structural search fallback
Empirical Validation — 2024–2026
0%
Token Reduction
Cursor Dynamic Context
0%
Portal Efficiency Gain
Cloudflare Code Mode
0%
MCP Token Savings
agentgateway