[meddlin]

Summary

Adds an Ask button to the top bar that slides out a chat panel on the right. It answers questions about your repositories with a model on an Ollama server, which can be another machine on the network. Ollama is optional: without it the panel says it cannot reach the server and nothing else changes.

The goal is to run local models against the data gitlib already holds, with a single 12 GB GPU in mind. The research behind the design is included in this PR (docs/ollama-integration.md), along with a second note on a scheduled Dependabot watch that is not built here.

What changed

Core (packages/core/src/ai)

• ChatBackend interface and an OllamaBackend over plain fetch: NDJSON streaming, abort, a stall timeout, and readable errors. It is injected into GitLibApp like SearchBackend, so tests use a fake.
• [ollama] config table with GITLIB_OLLAMA_HOST, GITLIB_OLLAMA_MODEL and GITLIB_OLLAMA_NUM_CTX overrides, mirroring [search].
• A token budget that cuts history to fit the context window (Ollama truncates silently and has no tokenizer endpoint), and a system prompt built from a capped repository list plus the current repository's notes and README.
• isGitLibError, a brand check for errors that cross a module-copy boundary (see reviewer notes).

Web

• POST /api/chat streams NDJSON events. server.mjs now aborts the request when the browser disconnects, so closing the panel or pressing Stop stops the model.
• A non-modal Sheet panel in the root layout (conversation survives navigation), with a model picker, Stop, a scope chip, and states for turned off, unreachable and no models installed.
• MarkdownBody gets an untrusted mode used for model output: no images, and link targets are shown. An image URL in text shaped by a hostile README could otherwise carry data out.
• The top bar's link list scrolls sideways so Ask stays visible on a phone.

Docs and config: README "Chat" section, runbook troubleshooting rows, .env.example files and docker-compose.yml.

For reviewers

• Not run against a real model. Everything was exercised against a stand-in Ollama (streaming, Stop, mid-stream errors, unreachable server, phone width, and the production server). Model quality, tool use and VRAM fit are untested.
• @ts-expect-error in routes/api.chat.ts. @tanstack/react-router pins router-core 1.171.38 while Start pins 1.171.39, so Start's type for a route's server option lands on a different copy and is not applied. The build finds server by its shape in the object literal, so it has to stay written out there. When npm ls @tanstack/router-core shows one copy the directive becomes an error, which is the cue to delete it. Realigning those versions is worth a separate change.
• isGitLibError. In the dev server instanceof against core's error classes failed (an expected error came back as a 500). There is a regression test.
• Top bar. It already overflowed at 375px before this change; the new button made it worse, so the nav now has min-w-0 and a scrolling link list.
• apps/web/src/routeTree.gen.ts is regenerated for the new route.
• The branch is 5 commits behind main; a dry-run merge is clean.

Not in this PR

Tool calling and retrieval beyond the capped prompt, a shared work queue for the GPU, and the Dependabot watch (research only, in docs/dependabot-watch.md).

Test plan

• npm run typecheck
• npm test: 278 passing, 126 of them new (config, token budget, prompt building, Ollama client over a real local HTTP server, route handler, protocol parsing)
• npm run build, and no server code in the browser bundle
• Manual, against a fake Ollama: streaming, Stop reaching Ollama (dev and production servers), repo scope, mid-stream error, unreachable then recovered, 375px width
• Against a real Ollama: run the "Verify against your server" commands in docs/ollama-integration.md, ask a question, and confirm ollama ps shows the model fully on the GPU

🤖 Generated with Claude Code