farnsworth-bench

fl-bench — model throughput benchmark

Thin wrapper over bin/fl-bench.mjs. It measures tokens/second for each selected model on a cold call and an immediate hot call, prints a table, and appends every result to <plugin>/.bench/results.jsonl.

What to run

Resolve the plugin root (the dir containing plugin.json for farnsworth-loop), then run the benchmark script with node. Pass the user's selection through verbatim; default to --models all.

node "<plugin-root>/bin/fl-bench.mjs" --models <selection> [--profile light|heavy]

<selection> (comma-separated, de-duped):

• all — every callable model (local MLX list discovered live). Default.
• a provider: anthropic | glm | local | codex | minimax
• <provider>:<id> — e.g. glm:glm-5.1, codex:codex-high, anthropic:opus, local:<omlx-id>
• a bare id — opus, glm-5.2, minimax-m3, codex-high, a local id

Profiles (--profile, default light; shorthand --heavy / --light):

• light — a ~200-word paragraph; fast/cheap throughput smoke (output cap 2048).
• heavy — a representative coding/agentic workload: a fixed >5k-token input context plus an instruction that elicits a long structured deliverable (>5k-token decode), output cap 8192, longer timeouts. Use this when the light profile's few-hundred-token decode is too small to characterise real coding throughput. The profile name is stored on every result row.

Useful flags: --list (dry-run; prints the resolved plan + profile, makes NO model calls — cheap way to confirm the selection before spending), --timeout <secs>, --help.

Guidance

• For a quick, cheap check first, run --list with the same --models selection and show the user the plan before the real (paid) sweep.
• The script handles auth from the environment exactly as the runners do (ZAI_API_KEY, MINIMAX_API_KEY, OMLX_AUTH_TOKEN; Anthropic uses the session's own auth; codex uses ~/.codex/auth.json). A provider whose key is unset is recorded as a failed row and the sweep continues — surface those rows.
• Results accumulate across runs in the append-only JSONL; point the user at <plugin>/.bench/results.jsonl for history.
• Report the printed table back to the user, including any failures and the * estimated-token note (codex fallback). The table shows cIn (cold input tokens), cOut/hOut (cold/hot output tokens) — under --profile heavy confirm cIn and the output columns are both comfortably over 5k.
• Heavy profile is much slower and pricier (each call generates thousands of tokens, and slow local models can approach the 1200s local timeout). For an all-models heavy sweep, warn the user and consider a representative subset.

See bin/README.fl-bench.md for the full usage and results-format reference.