AGENT SKILL

whalesong-listen

Render a recorded agent session to audio with Whalesong's deterministic synthesizer and interpret it — hear a loop, a stall, an error storm, or the difference between two runs. Use when the user wants to feel a session's shape, compare runs by ear, or present a run without scrolling a waterfall.

License 许可
MIT
最近更新
2026-11-13
Stars
★ 5
Forks
2

01 前言 Front Matter

name
whalesong-listen
invocation
model+user
description
Render a recorded agent session to audio with Whalesong's deterministic synthesizer and interpret it — hear a loop, a stall, an error storm, or the difference between two runs. Use when the user wants to feel a session's shape, compare runs by ear, or present a run without scrolling a waterfall.

03 SKILL.md 正文

Whalesong: listen to a run

The sonogram's audio path: the same signal pyramid that drives the picture drives sound. whalesong_listen renders a trace (or session file) to stereo 24 kHz WAV, time-compressed into the requested duration.

The vocabulary — designed, not measured

  • Sustained harmonic voices — model/reasoning, retrieval, subagent spans. A session that is mostly thinking sounds sustained.
  • Short envelopes at fixed registers — tool, file, code, browser, network calls. A busy hands-on phase sounds percussive.
  • Beating / dissonance — error density. A run under failure literally sounds rougher.
  • Stereo pan — fixed per category, so sources separate in space.

Meaning is carried by density, continuity, repetition and transitions. A pleasant chord is not a success signal, and this is an experimental diagnostic aid, not a validated monitor.

Use it well

  1. Render the interesting window, not always the whole run — seconds compresses; very long runs become grainy (the 80 ms score grid gets audible at extreme compression).
  2. For A/B ("did the fix help?") render both at the same seconds — different speeds change the texture.
  3. Play for the user: the tool returns afplay "<file>". Report what the render exposes — e.g. "first third is dense tool chatter, then a long near-silent stretch, then a rough patch at the end" — and tie each audible feature back to findings/rows from whalesong_analyze.
  4. Renders land in ~/.whalesong/renders/ unless out is given. Same trace + settings = same samples, so re-rendering is free.

Pair with rhythm

whalesong_rhythm is the quantitative side of listening: if a passage sounds metronomic, the periodogram's dominant period and the autocorrelation lag say exactly how metronomic, in seconds.

04 同仓库更多 Skills