Export schema

S3 log export writes gzip-compressed NDJSON parts and a manifest for each time window. Read the manifest first and consume only the parts it lists.

File layout

Every export writes the same batched, gzip-compressed NDJSON-per-window layout, uniform across all three log types. Three concepts are enough to consume any of them:

Windows: each scheduled run exports one time window per enabled log type.
Part files: The window’s records are written as gzip-compressed NDJSON (one JSON object per line), rolled into one or more part files of up to ~128 MB compressed each.
Manifest: a single .manifest.json per window lists the authoritative set of part files. It is written last, as the atomic commit point.

Object key layout

All three log types share the same key structure. Each window emits part files and a sibling manifest:

{key_prefix}{workspace_type}/{workspace_id}/{type}/{YYYY}/{MM}/{DD}/{type}_{window_start}_{window_end}.part-NNNN.ndjson.gz
{key_prefix}{workspace_type}/{workspace_id}/{type}/{YYYY}/{MM}/{DD}/{type}_{window_start}_{window_end}.manifest.json

Segment	Description
`{key_prefix}`	Your optional key prefix (empty if unset; ends in `/` if set).
`{workspace_type}/{workspace_id}`	Scopes every object to your workspace, so multiple workspaces can safely share one bucket and prefix.
`{type}`	One of `check-run-logs`, `code-review`, or `activity-logs`.
`{YYYY}/{MM}/{DD}`	Date partition, derived from the window’s start time (UTC).
`{type}_{window_start}_{window_end}`	Filename stem: the log type token, then the window bounds in compact ISO 8601 UTC, e.g. `check-run-logs_20260716T180000Z_20260716T190000Z`.
`.part-NNNN.ndjson.gz`	A gzipped-NDJSON part file. `NNNN` is a 1-based, zero-padded ordinal (`0001`, `0002`, …).
`.manifest.json`	The window manifest (see below).

Records roll into a new part file once the current part reaches roughly 128 MB compressed, so a large window is split across parts instead of buffered whole. A window always produces exactly one manifest: even a window with zero records, which writes an empty manifest.

The window manifest

Each window’s .manifest.json is the authoritative commit point. It has the same schema for all three log types:

Field	Type	Description
`schema_version`	string	Manifest schema version, currently `"1"` (distinct from the record `schema_version`).
`type`	string	Log type: `"check-run-logs"`, `"code-review"`, or `"activity-logs"`.
`workspace_type`	string	Your workspace type.
`workspace_id`	string	Your workspace identifier.
`window_start`	string	RFC 3339 start of the export window (UTC).
`window_end`	string	RFC 3339 end of the export window (UTC).
`exported_at`	string	RFC 3339 timestamp of when the export ran.
`part_count`	integer	Number of part files in this window.
`total_records`	integer	Total NDJSON records across all parts.
`parts`	array	The part files that make up this window (empty array if none).

Each entry in parts contains:

Field	Type	Description
`part`	integer	1-based ordinal of the part file.
`file`	string	Object name of the part file, relative to the window’s directory.
`records`	integer	Number of NDJSON records in this part.
`bytes`	integer	Compressed size of the part file, in bytes.

Example manifest:

{
  "schema_version": "1",
  "type": "check-run-logs",
  "workspace_type": "github_organization",
  "workspace_id": "123456789",
  "window_start": "2026-07-16T18:00:00Z",
  "window_end": "2026-07-16T19:00:00Z",
  "exported_at": "2026-07-16T19:00:07Z",
  "part_count": 1,
  "total_records": 42,
  "parts": [
    { "part": 1, "file": "check-run-logs_20260716T180000Z_20260716T190000Z.part-0001.ndjson.gz", "records": 42, "bytes": 15734 }
  ]
}

The consumer contract

Read the manifest first, then read only the parts it lists. Any .part-* object under a window’s directory that is not listed in that window’s manifest is stale and must be ignored. For example, it may have been left behind by a previous, larger export of the same window. Reading part files by prefix glob without consulting the manifest can mix generations and yield inconsistent data.

Because the manifest is written last, a consumer that reads the manifest and then reads exactly the parts it names always sees a consistent set. Re-exporting a stable (closed) window overwrites the same part objects with byte-identical content, so a reader can never observe a torn or mixed-generation window.

Check Run Logs

Type: check-run-logs One NDJSON record per concluded check run in the window. Each record contains:

Field	Type	Description
`schema_version`	string	Record schema version, currently `"2"`.
`workspace_type`	string	Your workspace type (e.g. `"github_organization"`).
`workspace_id`	string	Your workspace identifier.
`check_run_node_id`	string	GitHub node ID of the check run.
`check_run_type`	string	Type: `"custom"`, `"correctness"`, or `"approvability"`.
`custom_check_name`	string?	Name of the custom check, if applicable. Null for non-custom types.
`pr_number`	integer	Pull request number.
`repo_id`	string	Repository identifier.
`status`	string	Always `"completed"` (only concluded runs are exported).
`conclusion`	string	Check run conclusion (e.g. `"success"`, `"failure"`).
`created_at`	string	RFC 3339 timestamp of when the check run was created.
`concluded_at`	string	RFC 3339 timestamp of when the check run concluded.
`tool_calls`	array	Tool calls made during the check run. Empty array if none.

Each entry in tool_calls contains:

Field	Type	Description
`call_index`	integer	Sequential index of the call within the check run.
`tool_name`	string	Human-readable name of the tool used.
`internal_name`	string	Internal tool identifier.
`description`	string?	Description of what the tool call did.
`context`	string?	Context or reasoning for the call.

Code Review Analytics

Type: code-review One NDJSON record per code review that concluded in the window. This covers both standard reviews (correctness, approvability) and per-PR custom check (CRA) reviews. Each record contains:

Field	Type	Description
`schema_version`	string	Record schema version, currently `"2"`.
`workspace_type`	string	Your workspace type.
`workspace_id`	string	Your workspace identifier.
`repo_id`	string	Repository identifier.
`repo_name`	string?	Repository name. Null if unavailable.
`pr_number`	integer	Pull request number.
`pr_node_id`	string?	GitHub node ID of the PR.
`run_id`	string?	Review run identifier. Present for standard reviews; null for CRA per-PR records.
`pr_author_login`	string?	GitHub login of the PR author.
`pr_github_status`	string?	GitHub status of the PR (e.g. `"open"`, `"merged"`).
`head_sha`	string?	Head commit SHA reviewed.
`base_sha`	string?	Base commit SHA.
`head_branch`	string?	Head branch name.
`review_type`	string	Review type: `"correctness"`, `"approvability"`, or `"custom"`.
`review_started_at`	string?	RFC 3339 timestamp when the review started.
`review_completed_at`	string?	RFC 3339 timestamp when the review completed.
`cost_centimills`	integer?	Review cost, in centimills (1 centimill = 1/100,000 USD).
`context_bytes`	integer?	Size of the review context (diff), in bytes.
`incremental_context_bytes`	integer?	Size of the incremental review context, in bytes.
`num_objects_to_review`	integer?	Number of code objects selected for review.
`num_objects_reviewed`	integer?	Number of code objects actually reviewed.
`num_objects_billed`	integer?	Number of code objects billed.
`num_comments`	integer?	Number of comments produced by the review.
`num_fix_prs_generated`	integer?	Number of Fix It For Me PRs generated.
`num_fix_prs_accepted`	integer?	Number of Fix It For Me PRs accepted.
`num_fix_prs_not_accepted`	integer?	Number of Fix It For Me PRs generated but not accepted.
`any_custom_instructions`	boolean	Whether any comment in the review used custom instructions.
`inline_comments`	array	Inline (file/line) comments. Empty array if none.
`pr_level_comments`	array	PR-level (issue) comments. Empty array if none.

Each entry in inline_comments contains:

Field	Type	Description
`comment_id`	string	Macroscope comment identifier.
`github_comment_node_id`	string	GitHub node ID of the comment.
`github_comment_database_id`	integer?	GitHub numeric (database) ID of the comment.
`file_path`	string	File the comment applies to.
`line_start`	integer	Start line of the comment.
`walker`	string	Internal review identifier. It does not indicate the programming language. Older records may hold other values or an empty string.
`review_type`	string	Review type that produced the comment.
`custom_check_name`	string?	Custom check name, if applicable.
`has_custom_instructions`	boolean?	Whether the comment used custom instructions.
`comment`	string	Comment body.
`runtime_impact`	string?	Assessed runtime impact.
`pr_impact`	string?	Assessed PR impact.
`positive_sentiment`	boolean	Whether the comment received positive feedback.
`negative_sentiment`	boolean	Whether the comment received negative feedback.
`is_addressed`	boolean	Whether the comment was addressed.
`addressed_by_action`	string?	How the comment was addressed.
`addressed_commit_hash`	string?	Commit that addressed the comment.
`is_resolved`	boolean	Whether the comment thread was resolved.
`feedback_source`	string?	Source of the feedback signal.
`cost_centimills`	integer?	Comment cost, in centimills.
`cra_config_commit`	string?	Commit of the custom check agent config, if applicable.
`cra_config_file_path`	string?	Path of the custom check agent config, if applicable.
`created_at`	string	RFC 3339 timestamp of when the comment was created.

Each entry in pr_level_comments contains:

Field	Type	Description
`github_comment_node_id`	string	GitHub node ID of the comment.
`github_comment_database_id`	integer	GitHub numeric (database) ID of the comment.
`url_ref`	string?	URL reference for the comment.
`review_type`	string	Review type that produced the comment.
`custom_check_name`	string?	Custom check name, if applicable.
`comment`	string	Comment body.
`created_at`	string	RFC 3339 timestamp of when the comment was created.

Per-window aggregate. In addition to the record parts and manifest, code review analytics writes one gzipped rollup object per window:

{key_prefix}{workspace_type}/{workspace_id}/code-review/{YYYY}/{MM}/{DD}/code-review_aggregate-{window_start}.json.gz

The aggregate is a gzip-compressed JSON object:

Field	Type	Description
`schema_version`	string	Aggregate schema version, currently `"2"`.
`exported_at`	string	RFC 3339 timestamp of when the export ran.
`workspace_type`	string	Your workspace type.
`workspace_id`	string	Your workspace identifier.
`window_start`	string	RFC 3339 start of the export window (UTC).
`window_end`	string	RFC 3339 end of the export window (UTC).
`total_review_runs`	integer	Total review runs in the window.
`total_inline_comments`	integer	Total inline comments across all runs.
`total_pr_level_comments`	integer	Total PR-level comments across all runs.
`total_comments`	integer	Total comments (inline + PR-level).
`inline_positive_sentiment_count`	integer	Inline comments with positive feedback.
`inline_positive_sentiment_pct`	number?	Percentage of inline comments with positive feedback (0-100, one decimal). Null when there are no inline comments.
`inline_negative_sentiment_count`	integer	Inline comments with negative feedback.
`inline_negative_sentiment_pct`	number?	Percentage of inline comments with negative feedback. Null when there are no inline comments.
`inline_addressed_count`	integer	Inline comments that were addressed.
`inline_addressed_pct`	number?	Percentage of inline comments addressed. Null when there are no inline comments.
`inline_dismissed_count`	integer	Inline comments dismissed (resolved without being addressed).
`inline_dismissed_pct`	number?	Percentage of inline comments dismissed. Null when there are no inline comments.
`fix_it_for_me_generated`	integer	Fix It For Me PRs generated in the window.
`fix_it_for_me_accepted`	integer	Fix It For Me PRs accepted.
`fix_it_for_me_acceptance_rate`	number?	Acceptance rate of Fix It For Me PRs (0-100, one decimal). Null when none were generated.
`by_review_type`	array	The same metrics broken down per review type (empty array if none).

Each entry in by_review_type contains review_type (string) plus review_runs, inline_comments, pr_level_comments, the four inline_*_sentiment/addressed/dismissed count-and-percentage pairs, fix_it_for_me_generated, fix_it_for_me_accepted, and fix_it_for_me_acceptance_rate: the same fields and semantics as the top-level totals, scoped to that review type.

User Activity Logs

Type: activity-logs One NDJSON record per workspace activity event in the window. Each record contains:

Field	Type	Description
`schema_version`	string	Record schema version, currently `"2"`.
`event_id`	string	Unique event identifier (UUID).
`workspace_type`	string	Your workspace type.
`workspace_id`	string	Your workspace identifier.
`occurred_at`	string	RFC 3339 timestamp of when the event occurred.
`actor`	object	The actor who performed the action (see below).
`action_category`	string	High-level category of the action.
`action_type`	string	Specific action type.
`resource_type`	string?	Type of the resource acted on, if applicable.
`resource_id`	string?	Identifier of the resource acted on, if applicable.
`metadata`	object	Action-specific metadata. May be an empty object.

The actor object contains:

Field	Type	Description
`github_user_id`	integer?	GitHub numeric user ID of the actor.
`github_login`	string?	GitHub login of the actor.
`ip`	string?	Source IP address of the action.

Pricing

Code Review

Status

API & Integrations

Workspace

Resources

File layout

Object key layout

The window manifest

The consumer contract

Check Run Logs

Code Review Analytics

User Activity Logs

​File layout

​Object key layout

​The window manifest

​The consumer contract

​Check Run Logs

​Code Review Analytics

​User Activity Logs

File layout

Object key layout

The window manifest

The consumer contract

Check Run Logs

Code Review Analytics

User Activity Logs