← All research

Research / AI harness

The code around the intelligence.

Context, tools, approval, and feedback: the machinery that lets an AI assistant work inside a LaTeX project.

Inside the AI harness One illustrative request

YOUR WORKSPACE

Fix the error in this paragraph.

.tex

Active document main.tex · selected paragraph

✓

≡

Allowed context Editor + project file access

✓

✳

AI harness Context · controls · feedback

Local Ollama Cloud

PROJECT TOOLS

get_active_document Read the source

✓

edit_selection Apply an approved edit

✓

compile Return build diagnostics

✓

Ask mode · approval before editing

Tool results return to the model

↳

Available tools follow your access settings.

Source read. The model can inspect the error.

Proposed change: ~~\sectoin~~ \section

Example edit accepted. Compile result: success .

“The command is corrected and the document compiles.”

05 / 05

The model can answer—or continue.

Your request brings the context.
The model asks to read.
An edit goes through approval.
Tools act and return evidence.
The model can answer—or continue.

The harness supplies allowed context, checks tool availability, and returns tool results to the model. This example shows an accepted edit in Ask mode. Actual tool sequences vary; a final answer, error, cancellation, or step limit ends the loop.

01 / Context and capabilities

Start with the access you allow.

Settings control access to chat history, the current editor, project files, the bibliography, and researcher profile. Available tools also depend on the selected workspace tab and whether a document or selection is open.

02 / Model routing

One loop, several runtimes.

Local GGUF and Ollama requests use the desktop bridge. Cloud requests use the configured provider. Local GGUF inference runs through node-llama-cpp; cloud mode supports Anthropic and OpenAI-compatible endpoints. With a cloud provider, supplied context is sent to that provider.

03 / Actions and feedback

Tools return evidence to the conversation.

A tool call becomes a real read, edit, check, or compile action. Ask mode pauses proposed text edits for approval. Each tool result goes back to the model, which can continue or answer. Cancellation and the configured step budget bound the loop.

Inside the implementation

Based on the desktop source reviewed at revision 4059a3d . File paths are relative to the application repository.

src/features/ai/aiAgent.ts Runs the tool loop and stopping rules
src/features/ai/aiTools.ts Defines tools, capabilities, and edit proposals
src/features/ai/useAiAgent.ts Builds context and handles approval
src/features/ai/aiClient.ts Routes local, Ollama, and cloud requests

AI source package and model catalog

Public source

The retired AI subdomain now lives here.

The retired AI source package is now served inside this route, so desktop builds can read the manifest, catalog, source files, scripts, and package metadata from latexdo.org/research/ai/ .

Source root: /research/ai/

Sync manifest

Maps the public AI files into the desktop app before builds.

Open manifest

Model catalog

Defines local GGUF models, cloud providers, and default AI choices.

Open catalog

Package files

Includes the copied README, validation scripts, tests, and implementation files.

Open README