/insights facet extraction silently fails — AI analysis sections empty despite gateway working for interactive sessions

Summary

The /insights command generates a report but the AI-powered analysis sections (What You Wanted, Session Types, Outcomes, What Helped Most, At-a-glance) are all empty. The ~/.claude/usage-data/facets/ directory is never created.

Environment

• Claude Code version: 2.1.186
• Platform: macOS (darwin 25.2.0, arm64)
• Auth method: OAuth (claude.ai login)
• Gateway: corporate gateway at bifrost.pronto.prod (configured via gateway-models.json)

Steps to reproduce

1. Run /insights in any session
2. Open the generated HTML report
3. Observe that all AI-generated sections show "No data" or are blank

Expected behavior

~/.claude/usage-data/facets/ is created and populated with per-session facet JSON files. AI analysis sections in the report are filled in.

Actual behavior

~/.claude/usage-data/facets/ is never created. The insights command receives {} as insights data. All AI narrative sections are empty.

Investigation

From inspecting the binary, facet extraction:

• Makes LLM calls with isNonInteractiveSession: true and querySource: "insights"
• Logs "Facet extraction failed: <error>" on failure but this is silent in normal usage
• Uses a separate model selector (Y$l()) from the interactive session model

Key finding: Teammates on the same corporate gateway have working /insights with facets populated. This rules out a gateway-wide block. The issue appears specific to this machine/account.

Additional investigation:

• Cleared ~/.claude/usage-data/session-meta/ cache entirely — did not fix it
• 32 qualifying sessions (≥2 messages, ≥1 min duration) exist with transcripts available
• Interactive sessions work perfectly; only the background non-interactive facet extraction fails
• No ANTHROPIC_API_KEY env var set; using OAuth

Workaround

The session-report plugin (claude plugin install session-report) works correctly as an alternative — it uses a standalone Node.js analyzer script that doesn't depend on background LLM calls.

Request

Please surface the actual error from facet extraction somewhere accessible (e.g. a --debug flag output or a log file) so users can diagnose why it fails without needing to inspect the binary.