This pull request has been automatically marked as stale due to inactivity.
Please add updates or it will be closed.

Closing as covered by landed equivalent changes.

Implemented and landed:

• 8e37c7021 — add per-agent params overrides merged on top of model defaults in resolveExtraParams.
• Pass sessionAgentId through applyExtraParamsToAgent from run/attempt.ts.
• Add schema support for agents.list[].params in src/config/types.agents.ts.
• Add regression tests covering match / merge-precedence / no-match behavior.

This delivers per-agent cache tuning (e.g. cacheRetention) without affecting agents that rely only on model defaults.

Are the available params documented anywhere? And how would you set it for cron jobs specifically?

@BillChirico Docs are here: https://docs.openclaw.ai/reference/prompt-caching and the config reference covers the schema too.

For cron jobs, they inherit params from whichever agent they run on. So you'd set it on the agent in agents.list[], not on the cron itself:

{
 "agents": {
   "list": [{
     "id": "my-cron-agent",
     "params": { "cacheRetention": "none" }
   }]
 }
}

If your crons don't target a specific agent you can also set it per-model under agents.defaults.models["<MODEL>"].params. Per-agent wins over model defaults when both are set.

Ah, so I'd make a specific cron agent @rrenamed, thank you! Do you have any other recommendations on how to use this to save tokens or anything else? Different types of agents or something.

Depends on your setup but the main win is disabling cache on anything low-traffic. We have a couple of cron agents that fire every 30min to 1h with small prompts, setting cacheRetention: "none" on those saved us roughly 25% on daily API spend since the cache writes were just expiring unused.

For high-traffic agents (main orchestrator, anything getting frequent messages) keep the default caching, that's where it actually pays off.

Other than that, model routing helps more than cache tuning honestly. We run Haiku for simple cron jobs (daily reports, health checks) and only use Opus/Sonnet for the agents that need it. That alone cut our costs way more than the cache stuff did. We went from like $8/day to around $2-3/day just from routing the right models to the right jobs.

@BillChirico Hope this helps:)

Other than that, model routing helps more than cache tuning honestly. We run Haiku for simple cron jobs (daily reports, health checks) and only use Opus/Sonnet for the agents that need it. That alone cut our costs way more than the cache stuff did. We went from like $8/day to around $2-3/day just from routing the right models to the right jobs.

@BillChirico Hope this helps:)

Thank does a ton thank you! You're able to route models that low level?

Thank does a ton thank you! You're able to route models that low level?

Yeah, each agent and cron can have its own model. In your config:

{
  "agents": {
    "list": [
      {
        "id": "main",
        "model": "openrouter/openai/gpt-5.2"
      },
      {
        "id": "daily-report",
        "model": "ollama/llama3:8b"
      }
    ]
  }
}

For crons you set the model directly when you create it:
npx openclaw cron add --name "daily-report" --model "<MODEL>" ...

So you can mix and match however you want. Heavy reasoning on Opus or GPT-5.2, coding on Sonnet or Codex, lightweight crons on Haiku or a local Ollama model. Each one gets its own model + params independently.

One thing that's worked well for us is having a second agent validate the first one's output. Like our main agent runs on one model and a separate tester agent on a cheaper one double-checks its work. Catches a lot of stuff the primary misses.

@rrenamed Can I contact you on Discord? This is a huge help and just have a couple more questions. My Discord is Bapes

@rrenamed I accidentally ignored your Discord invitation. Just sent you another one!