github / github/copilot-cli

Rubber duck reviews leave no verifiable record: independence and findings are unauditable after the session

Aperta
#4,621 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

Nessuno ha ancora preso questa issue.

area:agents area:sessions
Lingua principale
Shell
Stelle
11.2k
Fork
1.9k
Merge medio
14h 16m
PR unite (30g)
6

Descrizione

Describe the feature or problem you'd like to solve

Rubber duck reviews leave no verifiable record. The critique, the model that produced it, and what the session did with each finding all die with the session. #4380 and #4432 show why this matters: the independence guarantee (complementary family) can fail silently, and today the only way users can even detect that is by hand-extracting evidence from process logs and events.jsonl. The review's most valuable properties, who reviewed whom and what remained unresolved, are currently unauditable after the fact.

Proposed solution

Persist a small structured record per rubber duck invocation, next to the session or exportable to the repo: orchestrator model and family, reviewer model and family, how the reviewer was selected (complementary, inherited, explicitly overridden, and by whom), a hash of the review prompt, and the findings with their disposition (addressed, dismissed, left open).

Benefits: silent independence failures like #4380 and #4432 become visible in the record instead of requiring log forensics; teams can sample reviews after merge; and what the review left OPEN (unverified assumptions, accepted tradeoffs) survives as provenance instead of dying in the terminal.

Example prompts or workflows
  1. After a rubber duck pass, /rubber-duck export writes the review record to a file the user can commit alongside the change.
  2. A team samples merged PRs weekly and checks the records: which reviews ran cross-family, which findings were dismissed without evidence.
  3. CI warns when a PR's rubber duck record shows reviewer and orchestrator from the same family (the #4380 case, caught at review time instead of never).
  4. An auditor asks "what did the review leave open?" and gets an artifact, not a chat scrollback.
Additional context

Open disclosure: I maintain disensor (https://github.com/NicolasRocchia/disensor, MIT), a provider-neutral CLI and Action that formalizes exactly this record for adversarial review rounds (schema residue/v0.4: reviewer family and declared independence, prompt hash, findings with terminal states, residue declaration, CI gate). The method behind it is documented with data from 91 review events (DOI 10.5281/zenodo.21633495). Happy for any part of the schema to serve as reference or prior art; the point of this request is that the review record should exist, whoever implements it.

Guida per i contributori

Apri la guida per i contributori

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Direzione di ricerca

Inizia tracciando le invocazioni di rubber-duck e rivedendo i log del processo e le evidenze di events.jsonl menzionate nell’issue. Definisci come ogni invocazione registra i modelli, il motivo della selezione, l’hash del prompt, i risultati e le disposizioni, quindi esamina come /rubber-duck export debba emettere l’artefatto. Il lavoro è completato quando il record persiste dopo la sessione e supporta la verifica dell’indipendenza e dei risultati irrisolti.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Valutazione

Stack tecnologico
shell
Ambito
cli, observability
Tipo di issue
Funzionalità
Difficoltà
5/5
Tempo stimato
Più di una settimana
Stato di attività
Attiva
Chiarezza
Abbastanza chiara
Idoneità per principianti
35/100

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.