Parizival/xari-bulbul-bench

Xarı Bülbül agent benchmark

How well do models work as the coding agent inside Xarı Bülbül? Each row is one real agent run through the app's engine: the same system prompt, tools, permission checks and provider adapter as the desktop app. Runs are scored by checking the outcome (tests pass, the file is right, the answer is exact), not by grading the text.

The suite, the runner and the method are in bench/.

Results (2026-10-11)

Four free-tier models, 10 tasks: 5 coding, 2 tool use, 3 GitHub MCP.

Model	Pass	Coding	Tools	GitHub MCP	Median time	First token	Output tok/s	Tokens/run	Tool errors	Cost
openrouter:nvidia/nemotron-3.5-lightning	82% (9/11)	4/6	2/2	3/3	25.8 s	0.92 s	30	34625	0%	$0
google:gemini-flash-lite-latest	80% (16/20)	7/10	4/4	5/6	4.4 s	0.96 s	45	22852	4%	free tier
ollama-cloud:gpt-oss:120b	80% (16/20)	9/10	2/4	5/6	4.6 s	0.45 s	122	34065	23%	free tier
groq:openai/gpt-oss-120b	27% (3/11)	2/6	1/2	0/3	108.3 s	18.6 s	4	16408	0%	free tier

Task	google	ollama-cloud	openrouter	groq
Fix an off-by-one bug (fix-off-by-one)	✓ ✗	✓ ✓	✓ ✓	✓ ✓
Implement a function from tests (implement-slugify)	✓ ✓	✓ ✓	✓	✗
Write a parser to a spec (parse-duration)	✓ ✓	✓ ✓	✗	✗
Rename a function across files (rename-across-files)	✓ ✗	✗ ✓	✗	✗
Edit package.json precisely (json-edit)	✗ ✓	✓ ✓	✓	✗
Search the project and report a number (count-todos)	✓ ✓	✗ ✓	✓	✗
Run a script and save its output (run-and-capture)	✓ ✓	✓ ✗	✓	✓
Look up a repo on GitHub (gh-repo-facts)	✓ ✓	✓ ✓	✓	✗
Find code in a GitHub repo (gh-find-file)	✓ ✗	✓ ✗	✓	✗
Read a file from GitHub (gh-read-file)	✓ ✓	✓ ✓	✓	✗

Notes:

• Partial runs: Gemini and Ollama ran every task twice (20 runs each). Nemotron and Groq were slowed by free-tier rate limits, and the run was stopped after they had finished 11 tasks each, most only once. Their percentages come from fewer runs.
• Groq: its free tier allows 8,000 tokens per minute. With the GitHub MCP tools, each request is about 8,800 tokens, so every GitHub task fails with "413 Request too large". Groq also failed some runs with "Parsing failed" errors from its side.
• Gemini: 3 failures are free-tier quota errors (429), not wrong answers.
• Rescoring: one count-todos answer ("There are 5 TODO comments. 5") was rescored after the checker was fixed to accept it.
• Since this run: tools now clamp out-of-range numbers instead of rejecting the call. gpt-oss sent bash timeout_sec: 10000 (it read it as milliseconds) and grep max_results: 10000, which caused most of its tool errors. A rerun of 6 tasks × 3 with the fix: Ollama tool errors 22% → 9%, passes 14/18 → 16/18.

Columns in data/*.jsonl

Column	Meaning
model, providerId, modelId	the model and the provider it was reached through
task, category, rep	the task id, coding / tools / github-mcp, and the repeat number
pass, detail	whether the check passed, and why it failed
reason, error	how the turn ended (done, error, timeout, max_steps) and the provider error, if any
wallMs, ttftMs, modelMs	wall time, time to first token, and wall time minus tool time
requests, toolCalls, toolErrors	model requests, tool calls, and tool calls that returned an error
promptTokens, completionTokens, cost	token use and USD cost (null when the provider doesn't report it)
startedAt	ISO timestamp