common/source/starBenchserver/docs/recording.md

Optional interaction recording and results

Interaction history is useful, but it is not a requirement for participating or publishing a result. The default retains supported interactions and native videos on the server, without downloading duplicate media to the client. Native capture is adapter-specific; the shared service does not create a missing camera video. Standalone frame images are normally temporary, with the fallback exceptions below. These settings never change the actor's observations, robot actions, native horizon or score.

Reference driver settings

Add these options to a normal python -m AgentPolicies.gpt6_astra.run ... command:

--recording-mode standard
# Explicitly download another copy of retained server media:
--recording-mode standard --copy-server-media
# Or essential results, without optional client interaction/media retention:
--recording-mode minimal
# Independent overrides, for example:
--recording-mode standard --no-record-videos --no-record-images

Setting	Retains when supported	Standard / minimal
--[no-]record-interactions	Robot trajectory events, decisions and ACKs	on / off
--[no-]record-model-io	Visible model input/output, notes and method records	on / off
--[no-]record-tool-traces	Allowlisted visible tool events	on / off
--[no-]record-images	Retained client image artifacts / requested server image copies	off / off
--[no-]record-videos	Select video downloads when media copying is enabled	on / off
--[no-]copy-server-media	Download selected media instead of keeping server references	off / off

Explicit switches override the preset. Resolve them before the run; do not change them mid-episode. The Python SDK uses RecordingOptions and RunRequest(recording=...). No additional model call is made for recording. The essential closed-file inventory still identifies server media when client video copying is disabled; --no-record-videos does not delete server videos.

One canonical visual record

For new service runs, videos stay under server/episodes/<session>/. The client retains the server identity, session, relative media paths and integrity metadata in its artifact receipt; a reference is not a downloaded or independently verified local copy. Keep the server result directory on durable storage. Reports and exports must not assume a video exists in client/.../sim_server_files/.

Native adapters may need PNG files during execution. They remain available until native finalization and close. The shared service then drops eligible temporary frame images before publishing its final inventory, and records omitted paths and hashes in server_media_retention.json. Videos, scores, logs and provenance are not deleted. Frame pruning requires successful native finalization/close, nonempty videos numbering at least the declared cameras, and stable file hashes. Missing or zero-byte videos, insufficient camera count, or failed cleanup keep frames with a recorded reason. A pruning error preserves remaining frames and records the warning and any completed removals. These file/count checks do not prove video decoding, camera correspondence or complete frame coverage; inspect native readiness separately. This is close-time retention, not a promise of zero temporary image I/O.

To retain frame images explicitly, start the service with --retain-frame-images. If local image copies are also wanted, use client --record-images --copy-server-media. --copy-server-media intentionally creates an additional copy; it does not remove the server original. Model-side public-image history is also optional to retain; live images remain available to the actor regardless.

These defaults apply to newly launched sources. Frozen runs, existing server inventories and historical media are not rewritten or cleaned retroactively. Older servers may still retain their frames; a new client's media references do not silently change an older server's retention policy.

Essential records remain: method/configuration, benchmark/protocol and selected cases, lifecycle, outcomes, termination/errors, counters and known usage. The reference runner also keeps its existing internal source/configuration snapshots; minimal is not a promise of only four files or zero storage. Full source freezing is not required by the basic result exporter.

What the switches do not promise

• Client retention is not native capture. Servers may generate temporary images and videos while running. The client always finalizes once and retains the close receipt; its flags do not alter server capture. Service-side frame retention is selected separately at launch. New clients reference server media by default and download it only on explicit opt-in.
• Retention is not the live input channel. RGB and state still reach the actor. Direct's public image/history workspace remains available for the whole episode, even in minimal mode, and its owned scratch is removed after closure. Suppressing files before closure would change the method.
• Support varies by method. The optional recorder SDK is not a mandatory model proxy. Direct records visible turn inputs, action submissions/tool replies and allowlisted completed tool/message events. It excludes reasoning events, arbitrary provider envelopes and file diffs; it does not reconstruct a complete provider transcript. Other profiles do not automatically gain raw IO capture.
• Mixed files are conservative. Native metrics and provenance must not disappear merely because images are disabled. Unknown non-media evidence is retained conservatively; recognized interaction streams still respect their switches. Direct's mixed public JSON export requires interactions and model IO; robot events/ACKs remain separately configurable. Disabling images leaves references to unavailable images, not invented retained pixels. Components without selective-recording support may retain mixed artifacts and report that limitation rather than silently dropping their non-media evidence.
• The optional recorder is best effort. Its I/O/serialization failures produce fixed warnings rather than changing robot control. This does not relax existing core result persistence, private transport diagnostics, artifact integrity or service lifecycle checks; failures there remain infrastructure errors.
• This is not automatic secret removal. Essential private error diagnostics can contain participant text. Visible prompts/tool output can also be sensitive. Keep raw records private and review before sharing. Do not submit credentials, hidden reasoning or evaluator-private data to the interaction recorder.

Basic result without event logs

python -B -m benchkit.reports.basic_result \
  --run-dir /absolute/closed-run/client \
  --output-dir AgentPolicies/my_method/results/my_benchmark/report/my-run

The output is a standalone index.html and result.json. It checks essential closed-record consistency without opening events or source snapshots. Select an optional finite scalar native metric with --metric KEY; native success and a partial task score are not interchangeable. Only allowlisted public fields are projected, not arbitrary native info, paths, receipts or raw exceptions.

Keep the source run unchanged. The destination must be new and outside it. A record-consistent report does not prove native execution, provider identity or complete suite coverage. Its denominator is explicitly the client-selected schedule, not a guessed larger server/native panel.

An external driver can write the documented JSON record contract as the reference runner; it need not import or subclass the runner. The exporter reads manifest.json, status.json, episodes.jsonl, summary.json and the matching per-episode result.json files. Schema, identity, count and outcome consistency are checked, including every attempted failure/error and unattempted suffix. See the exact basic-result implementation and its contract fixtures for required field types; unsupported layouts fail explicitly, not by guessing metadata.

For sufficiently complete event records, the separate python -m benchkit.reports.interaction_history exporter provides a trajectory timeline. It currently does not merge agent_interactions.jsonl or copy media. Redacted/missing events may not satisfy its stronger audit: the basic result remains usable. Existing RoboDojo full-evidence publication and hash/source audits retain their narrower requirements; they are not global publication gates.