Recurse
Paddles
Local-First Coding Agent
Mech suit around a local-first coding agent. A recursive in-context planning harness that makes small models dramatically more capable through bounded iterative evidence gathering.
Highlights
- →Interpretation-first routing — operator memory, guidance graphs, and derived procedures shape every turn before the planner chooses its first action.
- →Recursive work earns better answers — the planner gathers and refines evidence through bounded workspace actions (search, read, inspect, shell, diff, edit) until sufficient for synthesis.
- →Separate planning from synthesis — planner and synthesizer are distinct workloads, each routed to the smallest model capable of that role.
- →Local-first by default — the core loop runs on local Qwen models; heavier planner lanes are opt-in and degrade gracefully.
- →Full-stream visibility — every recursive step, planner action, retrieval result, and synthesis decision is surfaced to the operator via TUI or event stream.
- →Durable trace lineage — stable task/turn/record/branch/checkpoint IDs backed by a TraceRecorder boundary with noop, in-memory, and transit-core adapters.
- →Threaded conversations — model-driven steering-thread decisions with explicit merge-back records and full replay views.
- →Four-tier context model — Inline, Transit, Sift, and Filesystem tiers with typed ContextLocator addressing and lazy cross-tier resolution.