Skip to content

V073-PM22-SCOUT-SUGGESTIONS.md

Latest commit

 

History

History
38 lines (26 loc) · 4.45 KB

V073-PM22-SCOUT-SUGGESTIONS.md

File metadata and controls

38 lines (26 loc) · 4.45 KB
Raw

PM2.2a: source-verified Scout suggestion inbox

Candidate: v0.7.3-pm22-scout-review1, from operator-verified v0.7.3-pm2-editfix1.

Delivered scope

A new authenticated Bob's Mind → Suggestions dashboard at /mind/memory/suggestions presents saved Bob Scout draft comments which can be staged for private memory review by the operator.

  1. Source: only an existing scout-result-<run-id>.json that still passes bob_publish_common.load_record() verification against its original saved feed snapshot, post ID, generated draft SHA-256 and approval SHA. No internet calls, unsupported inferred Moltbook links or guesses about live source authenticity.
  2. Stage: the operator explicitly chooses a verified Scout run. The app stores Bob's prior draft text as a pending, untrusted suggestion with source run ID, post ID, SHA-256 fingerprints and a separate private inbox ID. No memory is created at staging.
  3. Review: the operator reviews and may rewrite the proposed title/text, category and confidence. A typed reason is required to Approve as private memory or Reject suggestion. Provenance is set by the backend and cannot be changed by form fields.
  4. Approve: reverify the real local Scout source and stored fingerprints; reject absent/changed feeds or drafts, stale form tokens and already-reviewed suggestions. Create one existing PM2.1 private memory, linked to the review record, in the same SQLite transaction as the approved decision. Existing memory revision history still applies.
  5. Reject: keep an audit record and reason, without adding any memory.
  6. Deduplicate: one suggestion per Scout run ID, with 50 pending and 250 total bounded inbox records. Each candidate appears once; reviewed decisions remain visible.
  7. Isolation: no worker/service identity, Ollama, OpenClaw, Moltbook API calls, external publishing, automatic memory recall, automatic generation or direct trust promotion. All external/source text is HTML-escaped and treated as lower-trust context.

Accuracy boundary: this release does not generate novel LLM summaries from discussions. The initial inbox uses Bob's already generated, locally verified Scout draft comments to establish and test review/provenance. A future separately reviewed phase can ask the model to propose distilled notes. A stored Scout snapshot proves provenance and consistency with the source captured locally, not that the public discussion or Bob's draft is factually true.

Data and backups

Additive table bob_memory_suggestions in the same private SQLite database. It does not rewrite bob_memories schema, current memories, existing publication decisions or queue state. The existing updater takes a protected SQLite backup before installing code and restarts only Activity Manager.

Archive/retention/export for suggestion rows and automated suggestion generation are deliberately deferred until explicit review. The inbox is capped rather than silently deleting provenance.

Operator acceptance

  1. Run bobctl update --check then bobctl update; verify the release chain v0.7.3-pm2-editfix1 → v0.7.3-pm22-scout-review1; type UPDATE.
  2. Open Bob's Mind → Suggestions on the authenticated LAN dashboard. Select a locally verified Scout draft in Stage from existing Scout work, then click Stage for memory review.
  3. Refresh: the draft appears in Suggested memories · review inbox, with its original run ID, source post ID and untrusted text. The main Memory list must remain unchanged until approval.
  4. Edit title/content to a short accurate observation, select category/confidence, provide reason, and click Approve as private memory. Verify it appears in Bob's Mind → Memory and the recent decision links to it.
  5. Stage a different verified Scout draft, supply a reason and Reject suggestion. Verify no second accepted memory appears. A duplicate stage or replayed submission should fail closed.
  6. Confirm Scout & review, original publishing and Personality still behave as before. No new Moltbook posts or model calls should be triggered by the inbox.

Next gated phases

  • PM2.2b: model-authored candidate extraction with immutable source IDs, structured outputs, prompt-injection-resistant validation and human review only; privacy-safe export/retention planning.
  • PM2.3: explicit opt-in, read-only, bounded recall of approved active memories, never tool/persona instructions.
  • FM1: only after PM2 complete and operator-tested.