Open data is the new open source.
When LLMs drive the logic, traces become the real source: the record of decisions, tool calls, and reasoning behind the outcome.
opentraces lets you parse, sanitise, review, and push those sessions to HuggingFace Hub so you or others can build on real workflows, not synthetic benchmarks.
features ────────────────────────────────
git for traces
init, status, review, push. The workflow you know, applied to agent sessions.
security pipeline
Regex + entropy + optional TruffleHog and LLM review. Stable placeholders like [EMAIL_1] keep traces coherent.
auto or review
Per-project policy: auto-approve safe traces, or gate every session through the TUI or browser inbox.
blame and graph
Every shipped line back to the prompt behind it. ot blame and ot graph map commits to their agent sessions.
schema depth
Steps, tool calls, reasoning, sub-agents, tokens, attribution, outcome in one record.
huggingface native
Sharded JSONL on HF Hub. Load via datasets.load_dataset() or mount it as a virtual filesystem. No lock-in.
quality scoring
Five persona rubrics score every trace. Upload gates enforce minimums. Re-score remotely with ot assess.
content-hash dedup
Reset state, switch machines, re-push safely. murmur3 hashing blocks duplicates on the remote.
agent-native cli
Every command emits structured JSON with next_steps. Built for agents to drive agents.
privacy & trust ─────────────────────────
Layered scanning: 30+ regex patterns, Shannon entropy, optional TruffleHog (800+ detectors, opt-in), and optional local LLM review. Stable placeholders like [EMAIL_1] preserve referential meaning across a trace.
attribution ─────────────────────────────
When agents write most of the code, git blame points at a user, not a session id. ot blame and ot graph resolve every commit back to the traces that produced it — so you can see what worked and what didn't across your agent sessions.
Attribution search can surface as a semantic diff — added functions, modified classes, renamed files — so agents can pull up the session behind any change in a fraction of the tokens a line-level view would cost.
how it works ────────────────────────────
Hugging Face HubFine-tune on real workflows
Alternating role sequences, tool call/observation pairing, reasoning coverage. Validated against 10 quality checks before upload.
Reward from outcomes
Committed patches as reward proxies, per-step token costs for cost-penalized reward, sub-agent hierarchy for credit assignment.
Observability and ground truth
Cache hit rates, per-step token breakdowns, duration timelines, model distribution. Real production inputs with outcome signals become reproducible eval datasets for quality gating — no annotation queue required.
Filter by ecosystem
Language tags, extracted dependencies, VCS context, code snippets with language annotations. Build domain-specific datasets from HF queries.
schema ──────────────────────────────────
One session, one JSONL line. Full schema docs →
{
"schema_version": "0.3.0",
"trace_id": "uuid",
"execution_context": "devtime",
"task": { "description": "Fix the failing test...", "repository": "owner/repo" },
"agent": { "name": "claude-code", "model": "anthropic/claude-sonnet-4" },
"steps": [ // TAO loop
{ "role": "user", "content": "..." },
{ "role": "agent", "tool_calls": [...], "reasoning_content": "..." }
],
"outcome": { "success": true, "committed": true, "patch": "...", "terminal_state": null, "reward": null, "reward_source": null },
"attribution": { "files": [{ "path": "src/parser.ts", "ranges": [...] }] },
"metrics": { "total_steps": 42, "estimated_cost_usd": 2.40 },
"security": { "scanned": true },
"dependencies": ["react", "typescript"]
}get started ─────────────────────────────
Open data is the new open source. Your agent traces are the most valuable dataset nobody is collecting. Start contributing to the commons.