Report export snapshot schema
agentsight report export -o snapshot.json writes a JSON snapshot of the
materialized view. The same shape is returned by GET /api/v1/snapshot.
This document describes schema version 1.
Compatibility
Consumers should check schema_version before reading the rest of the
snapshot and reject versions they do not support. New fields may be added to an
existing version, so consumers should ignore unknown fields. Removing or
renaming a field, changing its JSON type, or changing its meaning
incompatibly requires a schema-version bump.
Every field listed below is emitted. A nullable field is emitted as null
when AgentSight does not have a value; arrays are emitted as empty arrays.
generated_at is an RFC 3339 UTC timestamp.
| Top-level field | Type | May contain captured content? | Meaning |
|---|---|---|---|
schema_version | integer | No | Snapshot schema version; currently 1. |
generated_at | string | No | Time at which AgentSight generated the export. |
summary | object | No | Aggregate counts and the covered time range. |
token_summary | array | Metadata | Token totals grouped by model. |
network_targets | array | Yes | Observed network destinations and request paths. |
process_nodes | array | Yes | Observed process metadata and command lines. |
audit_events | array | Yes | Bounded event history; see summary.audit_limit. |
resource_samples | array | Metadata | Process CPU and memory samples. |
sessions | array | Yes | Agent-session summaries and source attributes. |
tool_calls | array | Yes | Observed tool calls, including inputs and outputs when available. |
The export does not include the full llm_calls or token-usage row sets.
Their totals are available through summary and token_summary.
summary
| Field | Type | Nullable | Meaning |
|---|---|---|---|
source | string | No | View source, such as materialized_view or agent_native_session. |
view_events | integer | No | Total materialized rows counted by the view. |
llm_calls | integer | No | Number of observed LLM calls. |
token_usage_rows | integer | No | Number of token-usage rows used for aggregation. |
audit_events | integer | No | Total audit-event count before export limiting. |
sessions | integer | No | Number of observed sessions. |
input_tokens | integer | No | Aggregated input tokens. |
output_tokens | integer | No | Aggregated output tokens. |
total_tokens | integer | No | Aggregated total tokens. |
start_timestamp_ms | integer | Yes | Earliest covered Unix timestamp in milliseconds. |
end_timestamp_ms | integer | Yes | Latest covered Unix timestamp in milliseconds. |
audit_limit | integer | No | Maximum number of recent audit rows included in audit_events. |
token_summary[]
| Field | Type | Nullable | May contain captured content? |
|---|---|---|---|
group | string | No | Model or grouping identifier; metadata only. |
input_tokens | integer | No | No. |
output_tokens | integer | No | No. |
cache_creation_tokens | integer | No | No. |
cache_read_tokens | integer | No | No. |
total_tokens | integer | No | No. |
calls | integer | No | No. |
sessions | integer | No | No. |
network_targets[]
| Field | Type | Nullable | May contain captured content? |
|---|---|---|---|
pid | integer | Yes | No. |
comm | string | Yes | Process metadata. |
host | string | No | Yes; observed destination host. |
path | string | Yes | Yes. Raw observed request path; not generally normalized or redacted. |
count | integer | No | No. |
error_count | integer | No | No. |
first_timestamp_ms | integer | Yes | No. |
last_timestamp_ms | integer | Yes | No. |
process_nodes[]
| Field | Type | Nullable | May contain captured content? |
|---|---|---|---|
id | string | No | Identifier metadata. |
pid | integer | No | No. |
ppid | integer | Yes | No. |
root_pid | integer | Yes | No. |
start_timestamp_ms | integer | Yes | No. |
end_timestamp_ms | integer | Yes | No. |
comm | string | Yes | Process metadata. |
command | string | Yes | Yes; may contain the executable or command text. |
argv | array of strings | No | Yes; may include user data, paths, tokens, or secrets. |
cwd | string | Yes | Yes; may expose user names and filesystem layout. |
exit_code | integer | Yes | No. |
status | string | Yes | No. |
view_source | string | No | No. |
confidence | number | Yes | No. |
audit_events[]
| Field | Type | Nullable | May contain captured content? |
|---|---|---|---|
id | string | No | Identifier metadata. |
timestamp_ms | integer | No | No. |
audit_type | string | No | No. |
pid | integer | Yes | No. |
comm | string | Yes | Process metadata. |
subject | string | Yes | Yes; source-dependent subject. |
action | string | Yes | Usually categorical metadata. |
target | string | Yes | Yes; commonly a file, process, or network target. |
status | string | Yes | Usually categorical metadata. |
summary | string | Yes | Yes; source-derived event summary. |
details | any JSON value | No | Yes; arbitrary source event details. |
view_source | string | No | No; identifies captured, reconstructed, agent-native, or legacy-unknown provenance. |
confidence | number | Yes | No; source-specific confidence in the row correlation or reconstruction. |
view_source describes the lineage of each row, not the identity of the logical
operation. Its values are view for rows emitted directly from captured events,
sqlite for rows reconstructed from normalized persisted rows,
agent_native_session for rows parsed from native session files, and unknown
for legacy or otherwise unclassified evidence. One logical LLM call can therefore
have a directly captured call row and a reconstructed request row with different
sources. confidence is likewise row-specific: captured LLM rows reflect request/response
correlation confidence, other captured event types carry canonical-event confidence, and
reconstructed rows reflect extraction and lineage confidence. It must not be compared
across sources as a global probability.
resource_samples[]
| Field | Type | Nullable | May contain captured content? |
|---|---|---|---|
timestamp_ms | integer | No | No. |
pid | integer | Yes | No. |
comm | string | Yes | Process metadata. |
cpu_percent | number | Yes | No. |
rss_mb | integer | Yes | No. |
sessions[]
| Field | Type | Nullable | May contain captured content? |
|---|---|---|---|
id | string | No | Session identifier metadata. |
agent_type | string | No | No. |
start_timestamp_ms | integer | No | No. |
end_timestamp_ms | integer | Yes | No. |
status | string | No | No. |
model | string | Yes | Model identifier metadata. |
input_tokens | integer | No | No. |
output_tokens | integer | No | No. |
total_tokens | integer | No | No. |
view_source | string | No | No. |
confidence | number | Yes | No. |
attributes | any JSON value | No | Yes; arbitrary source-specific session attributes. |
tool_calls[]
| Field | Type | Nullable | May contain captured content? |
|---|---|---|---|
id | string | No | Identifier metadata. |
session_id | string | Yes | Session identifier metadata. |
conversation_id | string | Yes | Conversation identifier metadata. |
timestamp_ms | integer | No | No. |
tool_name | string | Yes | Tool-name metadata. |
tool_call_id | string | Yes | Tool-call identifier metadata. |
start_timestamp_ms | integer | Yes | No. |
end_timestamp_ms | integer | Yes | No. |
duration_ms | integer | Yes | No. |
status | string | Yes | Usually categorical metadata. |
input | any JSON value | No | Yes; captured tool arguments or request content. |
output | any JSON value | No | Yes; captured tool results or response content. |
related_pid | integer | Yes | No. |
related_event_id | string | Yes | Identifier metadata. |
view_source | string | No | No. |
confidence | number | Yes | No. |
Handling exported snapshots safely
Snapshots are observability artifacts, not sanitized reports. Treat them as sensitive by default and do not publish them without review. In particular:
- redact or remove request paths, command lines, working directories, audit targets/details, session attributes, and tool inputs/outputs;
- review identifiers and destination hosts when they could expose tenant, project, repository, or account names;
- store only the fields needed by the downstream consumer and apply its normal retention and access controls.
The sample in docs/sample-snapshot.json is an
example payload, not an additional compatibility contract. It predates the additive
audit provenance fields, so consumers should treat missing audit view_source and
confidence values as unknown and null, respectively.
继续阅读
返回索引
AgentSight:基于 eBPF 的零侵入 LLM 智能体可观测性工具
 
上一篇 / 上一页
Agent Nebula:目标设计与算法约定
Agent Nebula 回放 Agent session 中可证明的 Git 仓库文件动作时间,而不是 commit 时间。图中只有文件星点;根级 entry 决定色系,目录路径影响力场,但都不产生额外节点、 边界或标签。commit 不会移动文件;某一动作帧覆盖到 commit 时,只让最外框闪烁。
下一篇 / 下一页
使用说明
``sh git clone https://github.com/eunomia-bpf/agentsight.git cd agentsight git submodule update --init --recursive ``
- 最后更新
- 2026年8月24日
- 首次发布
- 2026年8月5日
- 贡献者
- github-actions[bot]
这个页面有帮助吗?