Inspect agent conversations

Use recorded interactions to understand an agent’s behavior and preserve evidence deliberately.

A final answer rarely explains why an agent succeeded or failed. Session evidence can show the requests, responses and tool interactions that Synth recorded. Use it to investigate behavior, not as a substitute for an outcome.

Find and inspect a session

synth session list
synth session show <session-id>
synth session streams <session-id>
synth session interactions <session-id>

Use IDs returned by the preceding command. A session may contain multiple streams; incomplete or missing evidence should not be treated as an empty conversation.

Session inspection is different from run inspection: a run organizes execution and scoring; a session organizes recorded interaction history. A session is not a general-purpose workspace or a command to resume an optimizer.

Export deliberately

synth session export <session-id> --output ./session-evidence

Choose a new destination. Read the receipt and exported coverage before assuming the bundle contains every file or interaction. Keep exports private: prompts, tool output and attachments can contain sensitive business information even when authentication tokens are not explicitly exported.

Import or back up local history

synth session import requires a reviewed, normalized synth.session-import.v1 JSONL file supplied with --file; it does not read native provider logs directly. Validate that file without uploading it:

synth session import --file ./reviewed-session.jsonl --dry-run

synth session backup is the path for native local history. It is broader: it can include transcripts, attachments, project files and associated metadata. Read its help and inspect a dry run before consenting to upload anything.

synth session backup --dry-run

A dry-run size estimate does not mean every embedded secret has been removed, and may not include all auxiliary payloads. Review the actual scope before uploading. Never treat a backup as permission to share someone else’s data.

For interactive work, see Claude Code with Synth. For reusable testing, convert a representative situation into a task, separating the actor’s context from private assessment material.