skein-community/coding-agent-traces

Coding Agent Traces

Content-free, per-turn telemetry from real AI coding sessions — contributed, with consent, through the Skein CLI.

This dataset captures how people drive coding agents (which provider, which model, how much reasoning effort, how many tokens, what the tool-call structure looks like, whether the turn succeeded) without capturing a single word of what was said. No prompts. No completions. No code. By design.

The deal in one line: every row is a label, a coarse bucket, a random surrogate key, or tool-call structure — never conversation content. The row schema is .strict(), so nothing else can ride along.

---

TL;DR

Grain	one row per turn (one assistant response + the request settings that produced it)
Content	❌ none — no prompt/output/system/reasoning text, no tool arguments, no tool results
Identifiers	random surrogate keys only (never real session ids, never hashes)
Metrics	tokens / latency / cost are bucketed, never full-resolution
Timestamps	coarsened to a UTC day
License	ODC-BY 1.0
Consent	explicit opt-in; own-data-only; per-contributor deletion supported
Schema id	skein-trace/1

---

What this dataset is — and is not

• ✅ Structural telemetry only: provider, model, harness, reasoning effort, bucketed token counts / latency / cost, tool-call structure (function names, counts, depth), coarse outcome, and a day-resolution timestamp.
• ❌ No conversation content. No prompts, completions, system prompts, or reasoning text. No tool-call argument values and no tool results. No raw IPs, no precise timestamps, no file or project paths. Session and contributor identifiers are random surrogate keys.

This shape is chosen deliberately. A content-free dataset is not a model-training corpus, so it carries minimal privacy, copyright, and provider-terms-of-service exposure — while still being genuinely useful for studying how coding agents are used in the wild.

Good for

• Benchmarking token / cost / latency distributions across providers, models, and harnesses
• Studying tool-call patterns (how deep, how many, which tools) by model and effort level
• Measuring outcome / error rates across configurations
• Understanding the real-world mix of providers, models, and reasoning-effort settings in agentic coding

Not usable for

• Reconstructing prompts, code, or outputs — that information is never collected
• Training or distilling a language model — there is no content to learn from
• Re-identifying a contributor or a session — keys are random surrogates with no back-reference in the dataset

---

Dataset structure

Rows are stored as JSON Lines at data/traces*.jsonl. The Hugging Face Hub auto-generates the Parquet view that powers the dataset viewer and the datasets loader.

from datasets import load_dataset

ds = load_dataset("skein-community/coding-agent-traces", split="train")
print(ds[0])

Example row

{
  "schema": "skein-trace/1",
  "rowId": "8f3c1d2a-…",
  "sessionKey": "b71e09c4-…",
  "turnIndex": 4,
  "provider": "anthropic",
  "model": "claude-opus-4-8",
  "modelKind": "reasoning",
  "harness": "claude-code",
  "harnessVersion": "1.2.0",
  "effort": "high",
  "tokensInBucket": "4k-16k",
  "tokensOutBucket": "501-1k",
  "latencyBucket": "5-15s",
  "costBucket": "0.01-0.1",
  "toolCalls": [{ "name": "Read", "count": 2 }, { "name": "Edit", "count": 1 }],
  "toolCallDepth": 3,
  "outcome": "ok",
  "errorClass": null,
  "ts": "2026-06-26"
}

Field dictionary

Column	Type	Meaning
schema	string	Always skein-trace/1.
rowId	string	Random surrogate key for the row (not a real id).
sessionKey	string	Random surrogate key for the session (not the real session id, not a hash).
turnIndex	int	Turn position within the session.
provider	string	Harness-observed API provider (e.g. anthropic, openai) or unknown.
model	string	Model identity as reported by the harness, verbatim, or unknown.
modelKind	string	Model type as reported by the harness, or unknown. Skein maintains no model taxonomy.
harness	string	The coding tool (e.g. claude-code, codex, opencode, aider, cursor).
harnessVersion	string?	Harness version, when exposed.
effort	string	Reasoning effort: minimal | low | medium | high | xhigh | unknown.
tokensInBucket	string	Bucketed input token count (see Bucketing).
tokensOutBucket	string	Bucketed output token count.
latencyBucket	string	Bucketed turn latency.
costBucket	string	Bucketed turn cost in USD.
toolCalls	list⟨{name, count}⟩	Tool-call structure only — function names and counts. Never argument values or results.
toolCallDepth	int	Number of tool invocations in the turn.
outcome	string	ok | error | unknown.
errorClass	string?	Coarse error class — no message body.
ts	string	UTC day bucket (e.g. 2026-06-26).

Bucketing methodology

Token counts, latency, and cost are proven side channels, so full-resolution values are never published — only bucket labels.

Metric	Buckets
Tokens (tokensInBucket, tokensOutBucket)	0 · 1-100 · 101-500 · 501-1k · 1k-4k · 4k-16k · 16k-64k · 64k+ · unknown
Latency (latencyBucket)	<1s · 1-5s · 5-15s · 15-60s · 60s+ · unknown
Cost USD (costBucket)	0 · <0.01 · 0.01-0.1 · 0.1-1 · 1-10 · 10+ · unknown
Timestamp (ts)	UTC day, e.g. 2026-06-26 · unknown

Dropped entirely — never emitted

All prompt / output / system / reasoning text · tool-call argument values and tool results · raw IP · precise timestamps · project and file paths · free-text notes. The extractor is allowlist / drop-by-default: a field that is not explicitly allowed is dropped, and a no-content-leak property test enforces that known secret/PII/code marker strings in a source session never appear in any emitted row.

---

How to contribute

Traces come from your own captured sessions, and nothing is published until you opt in.

# 1. Preview exactly what would be sent — content-free rows, allowlisted columns only
skein trace export --dry-run

# 2. Record your consent (own-data-only, ODC-BY, deletion supported)
skein trace consent --accept

# 3. Publish your content-free rows to this dataset
skein trace push

skein trace push is dry-run by default and refuses to send anything without (a) recorded consent and (b) an explicit confirmation flag. Each contributor keeps a local surrogate-key map (~/.skein/trace/keymap.jsonl) mapping their published sessionKeys back to their own sessions — this is what makes later deletion requests possible. The map stays on your machine; it is never uploaded.

---

Personal and sensitive information

The dataset is engineered to be content-free, and the residual-risk surface is deliberately small:

• No content means no names, secrets, code, or free text by construction.
• Surrogate keys, not hashes — a hash of a real id is still pseudonymous (still personal data); random surrogate keys carry no back-reference.
• Bucketed metrics and day-resolution timestamps blunt token/latency/cost and timing side channels.

Residual side-channel risk (e.g. an unusually rare combination of buckets) is low but non-zero; this is noted in the spirit of honest disclosure rather than because any specific re-identification path is known.

---

License

Open Data Commons Attribution License (ODC-BY) v1.0.

The artifact is a database of factual, content-free telemetry rows, so a database-appropriate, attribution-only license is the right fit — the same choice made by comparable public trace/conversation datasets (e.g. WildChat). A work license such as CC-BY is a poor fit because the rows are facts, not authored expression. No license clears third-party privacy or copyright in source material — which is exactly why this dataset publishes no content.

Takedown & data removal

The dataset is designed to be content-free. If you nonetheless believe it contains something that should be removed:

• Contributors: identify the affected sessionKeys from your local key map (~/.skein/trace/keymap.jsonl) and open a removal request referencing them.
• Third parties: open an issue describing the concern — you do not need to identify specific rows.

Contact: open an issue at https://github.com/FrancoisChastel/skein/issues, or email francois@chastel.co.

We acknowledge requests within a reasonable window and remove identified rows from the canonical dataset revision. Third-party mirrors and caches are outside our control. This policy complements, and does not replace, the platform's own reporting and takedown channels.

Citation

@misc{skein_coding_agent_traces,
  title        = {Coding Agent Traces},
  author       = {Skein community},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/skein-community/coding-agent-traces}},
  note         = {Content-free, per-turn telemetry from AI coding sessions. Schema skein-trace/1. ODC-BY 1.0.}
}

---

Collected and contributed via Skein — an open-source tool for migrating configuration, skills, memory, and now content-free traces between AI coding tools.