Testing

For the scored benchmark, see benchmark.md.

Unit tests

cd backend
go test ./...

These need no Ollama and no network. context_test.go and summary_test.go cover the context pipeline: the budget, summaries replacing the messages they cover, drop and cut to fit, chunk boundaries, merge candidates and verbatim user notes. memory_test.go covers @memory parsing, per-account scoping, any-word search, the index following edits, and the memory budget. summary_test.go also covers the forced chunk behind @compact.

Rendering

Markdown, math and the no-remote-content rule run in the browser, so there’s no Go test for them. They were checked by rendering these inputs through renderMarkdown on a live page, and confirming no request left localhost: $\pi r^2$, $E=mc^2$ , $$\sum…$$, \[…\], “costs $5 and $10, or $5-$10”, $ inside inline code and a fenced block, broken $\frac{1}{$, ![cat](https://…), raw <img src="https://…"> next to a data: image, and <p style="background:url(…)">, <svg><image href>, <video src>. Recheck these after changing chat.js’s markdown setup.

Live checks

backend/live_test.go is behind the live build tag, so go test ./... never runs it. It needs Ollama running and a copy of a Flint database with at least one conversation:

sqlite3 backend/data/chat.db ".backup /tmp/copy.db"   # .backup, not cp: the DB runs in WAL mode
cd backend
LIVE_DB=/tmp/copy.db LIVE_MODEL=qwen2.5-3b-instruct go test -tags live -run TestLiveSummarizeChunk -v .
LIVE_DB=/tmp/copy.db go test -tags live -run TestLiveDumpHistory -v .

TestLiveSummarizeChunk prints the summary the model writes for the oldest chunk that can be summarized. Add SHOW=1 to also print the transcript it was given. Use it to try out summary prompts on a fixed input, without rerunning a whole chat.
TestLiveDumpHistory prints exactly what the next request would send: role, tool-call count and the start of each message. This is how the tool-call drift in E8 was found.

Web re-ranking

docs/tools/web_rerank.py scores embedding models and prefixes for @web re-ranking on 15 saved, hand-graded Brave searches (web_rerank_results.json), with no Brave calls (E20). Add a setup to CONFIGS, pull its model, and run python3 docs/tools/web_rerank.py.

The long-chat run

docs/tools/long_chat.py drives one 19-turn chat with a folder attached through the real HTTP API and reports tool calls per turn, peak context use and overflow errors. It plants facts, makes the model use tools, and then asks for the facts back.

Run it against a scratch backend and a scratch copy of a folder:

mkdir -p /tmp/flint-run && cp backend/*.go /tmp/flint-run/
cd backend && go build -o /tmp/flint-test .
PORT=8097 DB_PATH=/tmp/flint-run.db ATTACHMENTS_DIR=/tmp/flint-run-att /tmp/flint-test &
python3 ../docs/tools/long_chat.py /tmp/flint-run qwen2.5-3b-instruct http://localhost:8097

The script approves only commands that look read-only and denies the rest. The denials are part of the test, since they’re what pushed the model into confused text-only replies in E8. Because summaries are written in the background, wait a minute after it finishes before looking at them:

sqlite3 /tmp/flint-run.db "select id, level, first_message_id, last_message_id, merged, user_notes, content from summaries"

(The system sqlite3 may be built without FTS5. The summaries table is a plain table, so this works either way.)

To compare against another version, build it from a git worktree and run the same script on a different port against the same Ollama, one run after the other. Runs vary, so compare the tool-call and overflow counts across the whole run, not single replies.

Using your installed models without sudo

The system Ollama service keeps its models in /usr/share/ollama/.ollama/models. That folder is world-readable, so a second Ollama started as your own user can serve those models read-only while the service is stopped:

OLLAMA_NOPRUNE=1 OLLAMA_MODELS=/usr/share/ollama/.ollama/models ollama serve

A plain ollama serve as your own user looks in ~/.ollama/models instead and won’t see them.