LLM Prompt Modes

Use llm_prompt_mode in shared configs. Non-default modes require llm_additional_metrics or per-agent additional metrics.

Mode	Meaning	Example
`default`	Original primary-metric prompt and strategy text.	`tpcc_llm_8param_default_gemini.json`
`full_metrics`	Primary metric remains the objective; extra metrics are supporting evidence.	`tpcc_llm_8param_full_metrics.json`
`full_metrics_signature`	Full-metrics mode plus explicit metric-signature comparison.	`tpcc_llm_8param_full_metrics_signature.json`
`indirect_recent`	Primary metric is hidden/noisy; only recent entries show extra metrics.	`tpcc_llm_8param_indirect_recent.json`
`indirect_all_plain`	Primary metric is hidden/noisy; all history shows extra metrics without signature-comparison wording.	`tpcc_llm_8param_indirect_all_plain.json`
`indirect_all_signature`	Primary metric is hidden/noisy; all history shows extra metrics with signature-comparison wording.	`tpcc_llm_8param_indirect_all_signature.json`

Default Mode

default is the best first mode. It shows the primary optimization metric, active parameter set, parameter descriptions, current configuration, and recent history.

The default prompt includes OS parameter descriptions. Set include_param_descriptions: false only when you intentionally want a shorter prompt.

Full-Metrics Modes

Use full_metrics when the direct objective is available but you want the LLM to see supporting system signals:

{
  "llm_prompt_mode": "full_metrics",
  "optimization_metric": "latency_p99",
  "optimization_goal": "minimize",
  "llm_additional_metrics": ["instructions_per_cycle", "cache_misses"]
}

Use full_metrics_signature when you want stronger wording around comparing metric signatures across configurations.

Indirect Modes

Use indirect modes when the application metric is unavailable, hidden, or too noisy for direct prompting. The LLM reasons from additional system metrics.

{
  "llm_prompt_mode": "indirect_all_signature",
  "llm_additional_metrics": [
    "instructions_per_cycle",
    "cache_misses",
    "cpu_load_cores_pct",
    "power_socket0_watts"
  ]
}

Indirect modes are LLM-only. Classical tuners still need one scalar optimization_metric.

Where To Inspect Prompt Shape

After a run, inspect:

<results_dir>/llm_api_logs/llm_api_iter<iteration>_<agent>_<timestamp>.txt

That file shows the exact prompt and request config for the call.