Agent Persona Exploration - 2026-11-07

Aperta Adatta ai principianti
  • Qualcosa non va?
#75,702 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
2/5
Tempo stimato
1-3 ore
Idoneità per principianti
74/100
Tipo di issue
Documentazione
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
github-actions, yaml
Ambito
devtools

Direzione di ricerca

Inizia da .github/aw/create-agentic-workflow.md e confronta le indicazioni sulla gestione degli input con .github/aw/workflow-constraints.md; poi leggi .github/aw/report.md per la raccomandazione separata sui risultati pari a zero. L’issue chiede tre aggiunte alla documentazione: la gestione degli input non attendibili, un valore predefinito per i report pianificati vuoti e controlli per le baseline della prima esecuzione e i limiti di scrittura nei PR dei fork. Il lavoro è completo quando tutte e tre le raccomandazioni sono affrontate nelle indicazioni pertinenti.

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

Descrizione

agent-research
Persona Overview
  • Agent: agentic-workflows, evaluated through the persona-evaluator sub-agent
  • Personas This Run: Program Manager, Designer, Legal / Compliance
  • Selection: explored-personas.json was empty, so all nine personas were eligible. These three non-technical personas had only the fixed Nov 4 baseline and no rotation coverage.
  • Scenarios Tested: 4 of 6 generated (PGM-1, DSG-2, LEG-3, LEG-4; PGM-2 and DSG-3 not tested)
  • Average Quality Score: 4.7/5.0 (previous run 4.4)
Key Findings
  • All four responses proposed read-only scopes, empty network allowlists, and explicit safe-outputs. No scenario scored below 4.6.
  • LEG-3 was the strongest response. It reads the approved license header from the base branch, so a PR cannot approve its own header change.
  • Two of four responses (PGM-1, DSG-2) parse free-text issue or PR bodies without flagging them as untrusted. Earlier high-scoring runs did flag this (DO-1, FE-1).
  • The entry-point guidance the evaluator reads (create-agentic-workflow.md) has no untrusted-input clause, though other .github/aw/ files mention untrusted content.
  • Scores are the parent evaluator judgement. The rise from 4.4 may partly reflect rubric drift rather than agent change.
Top Patterns
  1. Triggers: path-scoped pull_request for PR checks (tokens/*.json, src/**), and weekly schedule plus workflow_dispatch for reports.
  2. Tools: github in gh-proxy mode, with bash for deterministic work (WCAG contrast math, JSON diff run as python3 -I or jq).
  3. Security: read-only scopes, network.allowed: [], rolling comments with hide-older-comments, and a dedup key on every issue.
View High Quality Responses (Top 2-3)
  • LEG-3 (4.8): Filters to added files because paths: also matches modified and deleted ones. Splits results into pass, fail, and uncertain, and leaves uncertain cases to humans.
  • LEG-4 (4.8): Omits close-older-issues so unreviewed license flags stay open. Runs the diff in an isolated, no-network Python process.
  • DSG-2 (4.6): Recomputes contrast ratios from hex values instead of trusting claimed results. Flags tokens that no text pair references as unverifiable.
View Areas for Improvement (Top 2-3)
  • Untrusted input (PGM-1, DSG-2): Neither response labels issue or PR body text as data. Other .github/aw/ files mention untrusted content, such as workflow-constraints.md, but create-agentic-workflow.md does not reference it.
  • Empty-result policy (PGM-1): The response asked whether a zero-blocker week should post a status or noop. report.md defines noop only for zero-event windows, which leaves status digests ambiguous.
  • Diff-workflow edge cases (LEG-4, LEG-3): No first-run behaviour when no prior licenses.json exists. Fork-PR limits on add-comment and add-labels are not mentioned.
Recommendations
  1. Add an untrusted-input clause to .github/aw/create-agentic-workflow.md. Point it at the existing guidance and require authors to state that issue, PR, and comment text is data, not instructions, whenever a workflow parses it.
  2. In .github/aw/report.md, set a default for zero-result scheduled status reports (post a no-items line or noop) beside the zero-event rule.
  3. In .github/aw/create-agentic-workflow.md, add two checks for compliance-style workflows: first-run baseline behaviour for diff-based schedules, and fork-PR write-back limits for PR review workflows.
View Run Notes
  • Rotation state mismatch: explored-personas.json was empty at the start, yet the Nov 5 and Nov 6 results files cover six other personas (Information Worker, Product Manager, QA Tester, Backend Engineer, Frontend Developer, DevOps Engineer). This run appended its three personas to the empty list, as the spec says. The next run may therefore re-include earlier personas.
  • Invocation: one persona-evaluator call covered all four scenarios, and all four succeeded. The sub-agent proposed no workflow files and created none.
  • Stored: scenarios in scenarios-42386527774.json, structured records in results-42386527774.json.

References: §42386527774 · §42243338716

Generated by 🎭 Agent Persona Explorer · claude · routed: haiku55 xhigh · ~6 AIC · ⌖ ~2 AIC · ⊞ 6.7K · ◷

Lingua principale
Go
Stelle
5.6k
Fork
615
Merge medio
5h 44m
PR unite (30g)
808

Preparare l'ambiente

Apri in Codespaces

Avvia il container di sviluppo del progetto nel browser, con il tuo account GitHub.

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.

Altre issue di github/gh-aw

Tutte le issue di github/gh-aw

Issue simili

Altre issue su Go

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.