Home
Your agent wastes most of its tokens rediscovering information you already have.
CodeRecon indexes your codebase, gives agents ranked context in one call, runs affected tests and linters automatically as files change, and surfaces coverage gaps and structural risk before commit. All deterministic. Zero added LLM cost.
Works with any MCP-compatible agent · Open source, MIT license
What Changes for the Agent¶
What It Solves¶
Context retrieval¶
One recon(task="...") call returns ranked code spans from across the repo. A quality gate tells the agent whether it found what it needs before it starts editing.
Deterministic refactors¶
refactor_rename and refactor_move compute all edits up front with certainty levels. Cross-file, atomic, import-aware.
Verify before you commit¶
The daemon runs affected tests and linters in the background as you edit. checkpoint reads cached results and commits — one call.
How It Works¶
The core loop: retrieve context, edit, verify. The daemon handles indexing and test runs in the background.
Evaluation¶
Head-to-head e2e eval in progress — same SWE-bench tasks, baseline agent vs agent with CodeRecon.
Resolve rate — does the agent actually solve the task?
| Variant | Tasks | Resolved | Rate |
|---|---|---|---|
| Baseline | TBD | TBD | TBD |
| CodeRecon | TBD | TBD | TBD |
Efficiency — how much does it cost to get there?
| Baseline | CodeRecon | Δ | |
|---|---|---|---|
| Avg tokens | TBD | TBD | TBD |
| Avg tool calls | TBD | TBD | TBD |
| Avg turns | TBD | TBD | TBD |
| Avg wall time | TBD | TBD | TBD |
Statistical significance — McNemar's test on paired resolve outcomes (p-value TBD).