[merberg-ai]

Overview

• Owner-triggered local Ollama analysis from one verified saved Scout discussion by an active FM1 friend. The only model input is bounded public post title/content, author handle, chosen post ID and saved run.
• Zero or one structured tentative idea per job, with first-person interest, reason, uncertainty and low/medium confidence. Explicitly not a verified fact or established Bob memory.
• New SHA-sealed private local AI interest inbox and revision history, with manual approve/reject/archive/reopen, optimistic compare-and-swap, source rechecks and safe rendering. Approvals do not affect PM2 memory, FM1 notes, Scout preferences or publication.
• Worker runs after existing CSRF/Origin-checked POST commits, via fixed loopback 127.0.0.1:11434; no agent tools, credentials, new Moltbook GETs, timers, social activity or direct posts.
• Limits: one active generation, two attempts per source, 90-second timeout, 32 KiB reply, 450 output tokens, 60 private proposals / 15 pending. Stale jobs fail without retries; invalid model source IDs or instructions rejected.
• Two managed files: existing bob_friend_scout.py and generated bob-activity-manager from pinned builder. Release manifest v0.7.5-fm32b-watchlist1 → v0.7.5-fm32c-localideas1 using bobctl update and existing rollback.
• Added synthetic tests for local HTTP, secrets isolation, source/checksum tampering, owner review, anti-replay/limits, sanitization and manager exact security delta.
• Documentation: docs/V075-FM32C-LOCAL-INTERESTS.md; roadmap handoff updated.

Safety boundaries

All existing Auth/Host/Origin/CSRF/watch/global opt-in/publisher safeguards are preserved. There is no autonomous relationship change, follow, DM, comment, posting or unreviewed memory edit. FM4 and FM5 are deferred.

Acceptance

After CI passes and merge, one bobctl update, then click Ask local Bob for a tentative interest under a verified saved friend discussion in the LAN dashboard; refresh status and test owner review. A no-idea response is valid.