meted mode
Show or change how Meted acts on its own decisions.
meted mode
› observe Compute and report the allocation. No ceiling is set on the request.
conservative Apply the engine's ceiling × 1.5, with a floor of 256 tokens.
balanced Apply the engine's ceiling unchanged, with a floor of 96 tokens.
aggressive Apply the engine's ceiling × 0.8, with a floor of 48 tokens, and hint the target length.
meted mode balanced
✓ Mode: observe → balanced
Apply the engine's ceiling unchanged, with a floor of 96 tokens.
Meted will now change your upstream requests. Restart `meted dev` to pick this up.
Writes the new mode to meted.config.json. A running meted dev keeps its
current mode until restarted.
Options
| Flag | |
|---|---|
--json | Machine-readable output, including every mode's full policy |
Trying a mode without committing
meted dev --mode conservative
Runs the gateway in that mode without touching the configuration.
meted inspect "What is a semaphore?" --mode aggressive
Shows what a mode would apply to a single prompt, without any provider request.
meted eval --mode aggressive --limit 20
Measures a mode against real prompts before adopting it.
Per-environment
METED_MODE=observe meted dev # staging
METED_MODE=balanced meted dev # production
The environment variable wins over the configuration file. The same committed config can behave differently per deployment.
Choosing a mode
Each active mode applies a different ceiling, so each trades output tokens
against the chance of truncating an answer. To see that trade on your own
requests, assemble a set representative of your traffic and run
meted eval once per mode:
meted eval ./evals.json --mode conservative --out conservative.json
meted eval ./evals.json --mode balanced --out balanced.json
meted eval ./evals.json --mode aggressive --out aggressive.json
Compare output tokens, sufficiency, under-allocation and median latency across the three reports.
Changing mode takes effect on the next restart of meted dev.