pngwn/agent-framework-benchmark-pi-traces

Agent framework benchmark traces (Pi + DeepSeek V4.1 Flash, Kimi K3)

Complete, unfiltered traces of coding agents building the same Pokédex web app in 12 JavaScript frameworks, from pngwn/agent-framework-benchmark-pi, a port of Nolan Lawson's agent framework benchmark to the Pi agent and open-weight models on HF Inference Providers (thinking high): deepseek-ai/DeepSeek-V4.1-Flash-1017 pinned to together, and moonshotai/Kimi-K3 pinned to deepinfra.

Code, analyses and judge verdicts: pngwn/agent-framework-benchmark-pi-results.

Three runs are included:

Run	Model	Task
build-deepseek-v4.1-flash	DeepSeek V4.1 Flash	the original task, 5 iterations × 12 frameworks
extend-team-deepseek-v4.1-flash	DeepSeek V4.1 Flash	add a team-builder feature to Nolan's Opus 5.5 apps, 3 × 12
build-kimi-k3	Kimi K3	the original task, 5 × 12

Headline results, including a comparison across the two models and Opus 5.5, are in the results repo's README.

Contents

Path
iterations.jsonl	One row per agent run: framework, model, prompt, pass/fail, timestamps, file paths, metrics and judge_scores (JSON strings)
data/<run>/<framework>-<n>/session.jsonl	The full Pi session (v3): system prompt and tool declarations, every assistant message with thinking blocks and tool calls, tool results, per-message token usage and cost, compactions
data/<run>/<framework>-<n>/events.jsonl	Pi's --mode json event stream, minus per-token deltas: includes tool execution durations and retries
data/<run>/<framework>-<n>/trace.md	Condensed, readable trace (one block per turn), as given to the process judge
data/<run>/<framework>-<n>/judge/<dimension>-<n>.jsonl	Sessions of the LLM judges (configured as GLM-5.3; see the results README on the judge model; read-only Pi) that reviewed this app
data/<run>/<framework>-<n>/lighthouse/	Lighthouse reports (first run per URL, mobile)

Each build agent got upstream's prompt (Implement the app described in SPEC.md using <framework>. Use the latest stable version. Work autonomously until npm test passes.), each extension agent the feature description in FEATURE.md; both had Pi's read, bash, edit, write tools plus Exa-backed web_search / web_fetch, and the framework's official MCP servers and agent skills where they exist.

Reading a session

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("pngwn/agent-framework-benchmark-pi-traces",
                       "data/build-deepseek-v4.1-flash/sveltekit-1/session.jsonl", repo_type="dataset")
for line in open(path):
    entry = json.loads(line)
    if entry["type"] == "message" and entry["message"]["role"] == "assistant":
        for block in entry["message"]["content"]:
            if block["type"] == "toolCall":
                print(block["name"], json.dumps(block["arguments"])[:120])

Or open one in a browser with Pi: pi --export session.jsonl.

Redaction

Exported with scripts/export-results.mjs: API key values and token-shaped strings are replaced with [REDACTED], local machine paths and the account email are removed, and the export is re-scanned before upload. No key value appeared in any trace before redaction. Images the agents viewed (the benchmark's reference screenshots and screenshots of their own apps) are kept as base64 in the sessions.

License

Apache-2.0. Model outputs are from DeepSeek V4.1 Flash (MIT), Kimi K3 and the GLM judge; the task, spec and reference screenshots are from Nolan Lawson's benchmark (Apache-2.0); Pokémon data from PokéAPI.