| name | triage-feedback |
| description | Use when asked to triage, run, or process new feedback — the full ingest → tag → cluster → bet → act pass. |
Triage feedback (full pipeline)
Objective: every untriaged feedback item ends the pass tagged, clustered into
a theme, and reflected in evidence-backed bets — with all side effects on the
ledger.
Procedure
- Scope the work. Call
list_feedback with only_untriaged=true. If it
returns nothing and no source has new items, stop and report "no new
signal" — do not re-cluster what is already triaged.
- Ingest new raw items from every available source connector, then write
them with ONE
ingest_feedback call per source batch. Map each item to the
contract shape: source in github|intercom|slack|call|doc, external_id
(the dedupe key — always set it), author, text verbatim (never
summarize at ingest), url when known. Duplicates are expected and
harmless; the store skips them.
- Tag every untriaged item with
tag_feedback (batch the updates):
urgency rubric — 3: outage, data loss, or explicit churn threat
("canceling", "switching"); 2: blocking bug or broken integration with
no workaround; 1: real friction or a repeated request; 0: nice-to-have
or praise.sentiment: pos|neu|neg from the text itself, not from the topic.
- Cluster into themes and upsert with
save_themes. Quality bar:
- 3–8 themes total; merge near-duplicates, split only on different root
causes.
- Title states the user's problem ("Checkout 500s on business plans"),
never a solution ("Fix pricing lookup").
feedback_ids lists every supporting fb_ id; one weak item is not a
theme — leave it tagged but unclustered.- Theme
urgency = max urgency of its evidence; trend is "rising" only
when new items joined an existing theme this pass.
- Surface contradictions. When two items demand opposite behavior, keep
both fb_ ids in the theme summary and name the conflict explicitly —
never silently pick a side.
- Propose bets and persist with
save_bets: each bet cites at least 2
evidence_ids (1 is acceptable only for an urgency-3 outage), links its
theme_ids, and scores impact/effort 1–5 with honest confidence
(cap at 0.9). No evidence → no bet; drop it.
- Act under directives. Call
get_directives. If escalation is ordered
and warranted, follow the escalate-churn-risk skill; if the brief must
stay current, follow write-weekly-brief.
- Report: counts (ingested, tagged, themes, bets, actions), the single
most urgent theme with its evidence count, and anything skipped plus why.
Failure handling
- A tool returns
"status": "error" → fix the arguments and retry ONCE; if
it still fails, continue the pass and put the failure in the report.
- Never invent fb_/th_/bet_ ids — use only ids returned by tools.
