whalesong-triage
Sweep recent sessions across every agent CLI recorded in Whalesong and surface the ones worth a look — errors, runaway loops, token blowouts, stalled gaps. Use for "what have my agents been doing", morning reviews, or finding the session that went wrong.
- License 许可
- MIT
- 最近更新
- 2026-11-13
- Stars
- ★ 5
- Forks
- 2
01 前言 Front Matter
- name
- whalesong-triage
- invocation
- model+user
- description
- Sweep recent sessions across every agent CLI recorded in Whalesong and surface the ones worth a look — errors, runaway loops, token blowouts, stalled gaps. Use for "what have my agents been doing", morning reviews, or finding the session that went wrong.
03 SKILL.md 正文
Whalesong: triage recent sessions
A cross-tool sweep over the durable store. Default window is the last 24h
(since_hours), unless the user names a period.
Sweep
whalesong_health— confirm the platform is up and note per-project counts so the sweep is bounded.whalesong_toolswith the same window — aggregate per-source, per-model and per-tool calls/errors/tokens. This is the fleet's actual behavior distribution before you look at any single session.whalesong_list_traceswithsince_hours— one row per session: source, name, timestamp, observation count, latency, cost.whalesong_findwithlevel: ERROR— every errored observation across all tools, with itstraceId. Group by trace.whalesong_daily— the per-day token/trace totals put any outlier in context.
Rank, don't exhaust
Most sessions are routine. Rank candidates for deeper inspection by:
- error density (errored observations / total),
- periodicity —
whalesong_rhythmon suspects: a strong dominant period + low spectral entropy is a metronome (poll loop, retry timer), high entropy is ordinary busyness, - retries and loops (
whalesong_analyzeon the top suspects), - token volume vs. that source's baseline in
whalesong_daily, - long spans with few events (stalled or only partially recorded).
Run whalesong_analyze on the top 2–3, not all of them — findings on
hundreds of sessions is noise, not triage.
Report
A compact table: source | session | when | events | errors | notable finding.
Then a short paragraph on the worst offender with evidence rows
(whalesong_observations), and whether the rest of the window looked normal.
Boundaries
sourcevalues:claude,codex,kimi,grok,devin,muse,amp,codewhale. A source absent from the listing simply had no recorded activity in the window — say that rather than implying it didn't run.- Session data is local and private; never post trace contents anywhere.
- Coverage gaps (a tool whose store isn't wired into the collector) are a platform gap, not a finding — mention it in one line, don't dwell.
04 同仓库更多 Skills
xlsx
Create/edit/analyze/verify spreadsheet workbooks including XLSX/CSV/TSV.
webapp-testing
Start/reuse a local app, wait for readiness, inspect rendered state/console/network, act from observed selectors, and verify with evidence.
verify
Exercise the real app/API/CLI and collect observable evidence; tests alone do not count as end-to-end verification.
tts
暂无简介
test
Detect the project’s test stack, run the narrowest useful tests, create tests when authorized, and report coverage/gaps honestly.
spreadsheets
Compatibility alias for the xlsx skill. Prefer loading xlsx.