oz0 replay continue
For: Operators and developers looking up a command.
For a model or prompt change: at up to --points recorded decision points of each of up to --max-jobs finished jobs (model calls whose answer asked for tool calls), the candidate (--model, an alias or <provider>/<model>; or the agent's instructions replaced by --instructions) is asked -k times, and the unchanged control is asked -k times beside it, each with the record up to that point as context.
What they propose goes through the candidate policy (the policy now, or --preset/--policy/--promote: a proposed level change) and is never run: no tool is called.
The report gives, per decision point, the share of the candidate's answers that differ from the record and from every control answer, and the control's own share against the record (the model's own variation, which the candidate's is read against); per agent, the mean divergence. It claims nothing about the turns after a point, and nothing about the answers' quality.
These are real model calls: they run as jobs on the tenant's nodes, through the gateway, metered and under the tenant's budget and --budget-usd in all, and show as the platform's replay spend, apart from the tenant's work. A job about a data subject is never continued: its content is not sent to a model again. Starting one needs comparisons.start and the content permission; oz0 replay continue cancel ID stops one.
Usage
oz0 replay continue (--model M | --instructions FILE) --budget-usd USD [-k 3] [--points 5] [--max-jobs 20] [--since 7d] [flags]
oz0 replay continue [command]
Subcommands
Flags
| Flag | Type | Default | Description |
|---|---|---|---|
--agent | string | only this agent's jobs | |
--budget-usd | float | what the continuation's model calls may spend in all, in dollars (required) | |
--instructions | string | a file of instructions that replace the agent's (at most 64 KiB) | |
--json | bool | print JSON instead of a table | |
--k (-k) | int32 | 3; 3 to 10 | how many times each arm is asked at each point |
--max-jobs | int32 | 20, at most 200 | at most this many jobs, newest first |
--model | string | the candidate model: an alias or <provider>/<model> | |
--plugin | string | only the jobs of this plugin's agents | |
--points | int32 | 5; at most 10 | decision points per job, spread over it |
--policy | string | decide under this tool-policy document (JSON) instead of the policy now | |
--poll | duration | 2s | with --wait, how often to look |
--preset | string | decide with this preset added: a shipped one, or the name of the level change --preset-rules proposes | |
--preset-rules | string | a JSON list of tool rules: a proposed level change | |
--promote | strings | decide with these shadow rules acting | |
--since | string | 7d | how far back to read: 7d, 36h, 90m |
--tenant | string | default | tenant |
--wait | bool | wait for the report and print it |
Inherited flags
| Flag | Type | Default | Description |
|---|---|---|---|
--context | string | $OZ0_CONTEXT or <config dir>/orchestrator-zero/context.json | CLI context file written by server init |
--log-level | string | info | log level: debug, info, warn or error |
See all oz0 commands and the global flags.