Recursive Planning Harness

Make small local models behave like grounded coding agents.

Paddles is a local-first harness that gives small models operator context, bounded recursive investigation, and a synthesis pass that writes from evidence instead of guesswork.

Read The Story Run A First Turn

Load operator memory and repo docs before the model answers.
Let the planner search, read, and refine instead of improvising from memory.
Keep the same turn trace visible in both the TUI and the web runtime.

Typical First Turn

Why Teams Reach For Paddles

Paddles changes what a small local model can actually do.

The harness is opinionated on purpose. It adds repo-grounded context, recursive evidence gathering, and a visible answer path so local models can behave more like careful operators than autocomplete with better marketing.

Operator Context

Load repo-specific guidance before the first token

Paddles reads AGENTS.md memory, foundational docs, and project conventions into the turn so the model starts from operator reality instead of generic priors.

Read the intro

Recursive Investigation

Search, read, and refine until evidence is sufficient

The planner can branch through the workspace in bounded steps, which lets small local models inspect code instead of bluffing through uncertainty.

See the planner loop

Structured Synthesis

Write the answer from evidence, not memory

A separate synthesis pass turns the gathered evidence bundle into the operator-facing response, keeping the final output grounded and legible.

See the turn loop

The Turn Loop

A single prompt can be vague. A traced turn is inspectable.

Paddles keeps the runtime legible: load guidance, choose the response path, investigate when necessary, and synthesize from evidence before the turn is recorded.

Interpret

Load operator memory

Paddles starts by reading operator guidance, workspace conventions, and runtime hints before it asks the model to act.

Classify

Choose the response path

The runtime decides whether the prompt is casual, direct, deterministic, or worth a planned investigation loop.

Plan

Search, read, refine

Planned turns recurse through bounded evidence gathering instead of spending the entire budget on a single completion.

Synthesize

Write from the evidence bundle

The synthesizer turns the gathered trace into a grounded answer that stays readable for the operator.

The final pass that assembles the response strictly from the evidence bundle accumulated during the turn.

Runtime Surfaces

Context structure and model routing stay visible together.

Paddles exposes both how evidence enters the turn and which model is responsible for each phase, so operators can tune the runtime without losing debuggability.

Context Tiers

Keep inline, transit, and filesystem spans distinct

Paddles separates short prompt context from fetched evidence and workspace reads so operators can inspect where each piece came from.

Read the context model

Model Routing

Route investigation and answer writing independently

A light synthesizer can pair with a heavier planner, which gives you deeper workspace inspection without paying the same cost for every final response.

See routing strategy

Reading Paths

Start with the docs path that matches the question you have.

After the shared intro, the docs branch into setup, loop mechanics, retrieval/context, and reference material so you can go straight to the depth you need.

Start Here

Install the runtime and run a first turn

Use the installation and first-turn guides when you want the shortest path from clone to a live local session.

Open the getting-started path

Turn Mechanics

Understand the planner and synthesis loop

Read how Paddles classifies prompts, runs recursive investigation, and assembles grounded answers.

Open the loop docs

Retrieval And Context

See how evidence is gathered and bounded

The retrieval docs explain query construction, traceability, context tiers, and how pressure is surfaced back to the operator.

Open retrieval docs

Reference Surface

Inspect the foundational contracts

Use the reference pages when you need the owning source documents and transport/runtime constraints in one place.

Open reference docs

Start Here

Read the narrative, install the runtime, and take a traced turn.

Open The Docs Installation Guide