Datasets:

kelikelibababian
/

mimir-agent-traces

language:
  - en
  - zh
license: mit
task_categories:
  - text-generation
tags:
  - agent
  - tool-calls
  - trajectories
  - audit
  - llm-agents
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.parquet

MimirAether Agent Traces

7,124 unique real agent sessions · 35,000+ tool calls from a multi-agent system in daily production — four cooperating agents (orchestrator, executor, red-team reviewer, verifier) running real workloads since May 2026.

What makes this corpus different: it is not single-agent coding transcripts. It captures inter-agent collaboration in production — task dispatch, cross-audits, failure & recovery (real incidents: gateway freezes, API outages, deadlocks — with their debug sessions included), and self-correction loops. If you train or evaluate agents on cooperation, failure recovery, or long-horizon operations, this is the rare real-world source.

🔗 This dataset is the runtime trace of MimirAether — an open-source, self-hosted agent runtime under active development (code, docs, roadmap & CI on GitHub). Issues and feedback on data quality are welcome on either repo.

🔄 Continuously updated — new sessions are appended in small batches (~every 100 sessions) as the runtime keeps operating.

Loading

from datasets import load_dataset

ds = load_dataset("kelikelibababian/mimir-agent-traces", split="train")
print(ds[0])          # one row = one complete agent session

The viewer table is served as Parquet (data/train.parquet) — a unified-schema view over the append-only JSONL source (data/mimir-agent-traces.jsonl, which is kept as-is and never rewritten). It covers the curated main file (4,168 sessions); the larger pre-filter archive is kept separately as data/mimir-agent-traces-legacy-full.jsonl and is not part of this table. Rows uploaded before schema v2 carry null in the seven v2-only columns, so the curated corpus loads as a single table.

What's inside

Each row = one complete agent session:

Schema note: v1 rows have task_desc_sanitized capped at 200 chars; from schema v2 (2026-09-13) the cap is 300. v1 rows (uploaded before 2026-09-13) lack the schema_version field — absence of the field = v1. The taxonomy is "9 categories designed, 7 reachable" via the tool-sequence heuristic — see the v2 section below.

Field	Type	Description
`session_id`	string	unique session identifier (salt-free SHA-1 prefix)
`n_tool_calls`	int64	number of tool calls in session
`duration_seconds`	float64?	wall-clock duration
`exit_reason`	string?	how the run ended (`natural` / `max_turns` / `interrupt` / `circuit_breaker` / `tool_storm` / `api_failure`); `null`/`unknown` for pre-2026-08-29 sessions
`tool_sequence`	list	ordered sequence of tool names (capped at 80)
`has_exit_reason`	bool	whether telemetry recorded an exit reason
`task_desc_sanitized`	string	task description, truncated (200 chars in v1, 300 in v2)
`schema_version`	string?	`"2.0"` for v2 rows; `null` for v1 rows (absence = v1)
`agent_id`	string?	which agent ran it — `mimir` / `hermes` / `openclaw` / `loki`
`model`	string?	batch-level model name (see Model field caliber)
`trace_id_prefix`	string?	first 8 hex of the trace id, for joining the audit stream
`trigger_source`	string?	`feishu` / `api` / `buzz-watcher` / `watchdog` / `cron`
`task_category`	string?	heuristic task class (weak signal — see below)
`task_subcategory`	string?	finer task class

v1 rows carry null in the seven v2-only columns — the Parquet view pads them so the whole corpus loads as one table.

Why this matters

Real-world agent behavior data is scarce; real-world multi-agent failure data is nearly nonexistent. Most published agent benchmarks are synthetic. This dataset captures what agents actually do in production — including genuine incident-response sessions (e.g., a cron job freezing the event loop, diagnosed via py-spy and fixed live; API-provider outages with retry/backoff traces), tool-usage patterns, and self-audit loops where an agent finds its own bug and files a correction. Failure trajectories are included deliberately: they are the hardest data to synthesize and the most valuable for robustness training.

Dataset statistics (full corpus)

Metric	Value
Sessions	7,028 unique
Total tool calls	36,995+
Duration	p50 = 0s (many short automated wakeups) · p90 = 132s
`exit_reason`	`natural` 247 · `max_turns` 18 · `empty_response` 14 · `api_failure` 8 · `interrupt` 2 · `unknown` 6,739
`tool_sequence` non-empty	6,936

exit_reason caveat: 96% of v1 rows are unknown — the recorder only started writing session_end telemetry partway through the corpus, and pre-2026-08-29 session_end events lack the exit_reason key. Even in schema v2 batches, exit_reason can be null when the underlying telemetry omits it (observed: ~22% in the first v2 preview window, which spans legacy sessions). The v2 gate guarantees the session ended; it cannot backfill a missing reason.

Useful for: agent behavior modeling, tool-use pattern mining, failure analysis, run-boundary research, agent memory studies.

Sanitization

All identifiers are hashed; user names, paths, emails and internal infrastructure names are redacted. Task descriptions truncated to 200 chars. Tool arguments/outputs NOT included (privacy) — only the tool-name sequence and metadata. Each batch is scanned for personal identifiers before upload; any row that fails the scan is dropped.

Source

Generated by MimirAether — an open-source, auditable agent runtime.

Schema v2.0（2026-09-13 起）

自 2026-09-13 起新增批次采用 schema v2，v1 行保留（混合读取安全）：

v2 新增字段

字段	类型	说明
`schema_version`	`"2.0"`	行级版本标识（前向兼容）
`agent_id`	enum	`mimir` / `hermes` / `openclaw` / `loki`
`model`	string	运行模型名
`trace_id_prefix`	string	trace id 的 hex 段前 8 位（类型前缀 `run_`/`tr_` 已剥除，join 审计流用；32-bit 熵）。仅 2026-09-13 X2-a 注入体系上线后的会话有值，历史会话为 null
`trigger_source`	enum	`feishu` / `api` / `buzz-watcher` / `watchdog` / `cron`（同上，仅新会话有值）
`task_category`	enum	9 类任务分类设计，实际 7 类可达（Communication/Creative 极少；见下）；由工具序列启发式投票推断，非原生标注；零工具会话归 Other
`task_subcategory`	string?	细分（可空）

任务分类法（9 类设计 · 7 类实际可达）

Terminal&Coding · Browser · Agent Tools · Retrieval · Memory · Communication · Analysis · Creative · Other

Heuristic disclosure: categories are assigned by a deterministic tool-name voting table (no LLM). In the full corpus, Communication and Creative are structurally rare; ~31% of a recent window lands in Other (mix of zero-tool sessions and tied votes). Treat task_category as a weak signal for stratification — not ground truth.

Model field caliber

model is a batch-level value (trajectories carry no per-session model). Resolution order: runtime telemetry (last_context_usage.json, what the gateway actually used) → config.yaml model.default → legacy .env MIMIR_MODEL. A model switch mid-window makes the batch value approximate for sessions before the switch; exact per-session models land once the runtime writes model into session_start (planned).

Known limitations

Session identity: session_id is a salt-free SHA-1 prefix of Mimir's raw session id. A session continued across midnight (≈4% of sessions) keeps the same id; the dataset keeps the first uploaded version of such a session and later continuation files are skipped (no upsert). Rows are always duplicate-free and never wrong — at worst the tail of a continued session is missing. This will disappear naturally once the runtime emits unique ids per session.
Batch-level model: see Model field caliber above.

Provenance

每批上传记录于 manifest.jsonl（批次/日期/会话数/备注）。原始会话文件永不上传——只有聚合统计行。