Skip to content

FRIENDS-ROADMAP-AND-HANDOFF.md

Latest commit

 

History

History
136 lines (97 loc) · 24.3 KB

FRIENDS-ROADMAP-AND-HANDOFF.md

File metadata and controls

136 lines (97 loc) · 24.3 KB
Raw

Bob Friends Manager: FM1–FM5 Roadmap and Fresh-Chat Handoff

Status captured: October 30, 2026 (America/Los_Angeles).
Current phase: FM1–FM4.1 operator-accepted (October 30, 2026). Next: FM4.2, revisable individual impressions with an auditable history and owner corrections, then FM4.3 optional social recall. FM5 and actual Moltbook follow/unfollow are unimplemented and separately gated.
Last operator-confirmed working release: v0.7.7-fm41-social1, installed in OpenClawLab WSL2 and confirmed working by the operator October 30, 2026. PR #49 merged with 599 passing offline tests.
Resume point: FM4.2 evolving social impressions, owner corrections and revisions. FM4.1 Social Memory Timeline is complete and accepted. FM4.3 opt-in recall remains a separate later phase.
Canonical repository: merberg-ai/bob-openclaw-moltbook, main.
This file is a planning/handoff artifact only. It does not grant permissions or change a service, updater manifest, runtime database or deployment.

1. Completed Friends Manager phases

Phase State Confirmed behavior Relevant source
FM1, private directory Operator tested Add, edit, archive, mute, reactivate; integrity-checked private friend records, revision history, saved Scout author quick-add src/lib/bob/bob_friends.py, docs/V074-FM1-FRIENDS.md
FM2, discovery and profile review Operator tested Explicit read-only public profile lookup and timestamped cache; Scout WHY-based candidate suggestion, owner review and approval before a friend record is added src/lib/bob/bob_friend_discovery.py, docs/V074-FM2-DISCOVERY.md
FM3, friend-focused Scout Operator tested Global default-off opt-in, capped public reads for up to two active agents; mix with global discussions; no new timer; manual-only friend-targeted publication src/lib/bob/bob_friend_scout.py, src/bin/bob-moltbook-scout, managed publisher; docs/V074-FM3-FRIEND-SCOUT.md
FM3.1, per-friend preferences Operator tested Individual watch pause/resume independent of relationship, local preference history and selected Scout WHY observations verified against saved source src/lib/bob/bob_friend_scout.py, docs/V075-FM31-FRIEND-WATCH.md; v0.7.5-fm31-watch1, merged PR #32
FM3.2a, copied Scout WHY interest inbox Operator tested Manual staging of verified previous Scout impressions; review with sealed local history; no change to memory or publishing docs/V075-FM32A-INTEREST-REVIEW.md, v0.7.5-fm32a-interests1, PR #35
FM3.2b, scalable watchlist Operator tested Paginated/filterable private watchlist and timestamped saved Scout evidence docs/V075-FM32B-WATCHLIST.md, v0.7.5-fm32b-watchlist1, PR #36
FM3.2c, local AI tentative interests Operator tested Owner-requested bounded Ollama inference on verified chosen public posts; independent source-grounded review docs/V075-FM32C-LOCAL-INTERESTS.md, v0.7.5-fm32c-localideas1, PR #37
FM3.2d, manual one-friend public focus Operator acceptance PASSED (Oct 30, 2026) Confirmed read-only public inspection, up to 1 profile + 2 verified posts, optional reviewed local reaction, cooldowns docs/V075-FM32D-MANUAL-FOCUS.md, v0.7.5-fm32d-focus1, PR #38

Important distinction: operator acceptance of dashboard behavior does not by itself prove that every subsequent scheduled Scout iteration completed after the startup repair. See operational watch item below.

2. Completed FM3.2 — interest proposals and focused discovery controls

Implemented and operator-accepted through FM3.2d. The four sub-deliverables below are historical design requirements and completed scope, not pending tasks. None grants unreviewed social actions.

Implemented sub-deliverables, preserved here as the FM3 design contract:

  1. Interest proposals from actual evidence. Bob may generate short, bounded, local-model proposals about which saved agents or subjects interest him and why, grounded in source-verified Scout selections or public discussion encounters. Label these as Bob's tentative ideas, not verified facts about a person. Include source run/post, timestamp, uncertainty and reason. No silent writes to authoritative FM1 operator notes, identity, or PM2 curated memory.
  2. An owner-review inbox. Approve, reject, archive or correct a proposal, with deterministic provenance checks, digest-guarded review, explicit status, reversible local history and capped retention. Bob may disagree, revise interests and revisit questions, but external posts remain untrusted data. An approved preference may influence browsing only through separately enabled friend scouting.
  3. Watchlist scalability and useful evidence. Pagination/filtering for more than the first ten active friends, plus per-friend last-observed/last-selected times only where grounded in actual verified saved records. Keep UI responsive/mobile-friendly.
  4. Optional operator-triggered read-only focus check. A manually requested, budgeted inspection of one specific friend's public profile/recent posts, using existing fixed-host GET endpoints, existing credentials boundaries and strict rate limits. Make this a separate, explicitly confirmed action; do not implement it as a new unrestricted background timer or automatic reply mechanism.

Delivered across four separately tested, incremental bobctl update releases. Any future expansion is a new, separately authorized phase.

FM3.2 acceptance: source citations survive review; no fabricated relationships, agent account identity or mutual follow; no private notes passed into prompts or Moltbook requests; no auto-follow/comment/DM; individual pauses and global switch honored; all existing publication safety tests green; an update through the normal bobctl update workflow with backup/rollback.

FM3.2a operator-accepted checkpoint

  • Operator tested: v0.7.5-fm32a-interests1, merged PR #35, one managed module bob_friend_scout.py from v0.7.5-fm31-watch1.
  • What shipped in candidate: explicit owner staging of Bob's existing, source-verified chosen Scout WHY from a current active friend; capped private SHA-sealed proposal records, independent source recheck on approval/reopen, status and reversible local history. No new model calls yet, and approved records do not influence memory or browsing.
  • Next after FM3.2b: FM3.2c fresh, bounded local-model proposals with provenance/uncertainty and mandatory owner review. Explicit one-friend public focus reads remain a later opt-in item.
  • Deployment: after offline test + PR review/merge, use only bobctl update with standard backup and rollback. Release handoff and short browser acceptance: V075-FM32A-INTEREST-REVIEW.md. The prior Scout startup watch item stays open until a real scheduled timer-run is verified.
  • Security: no publication, platform follow/DM/comment, model prompt expansion, new automatic timer, private note disclosure or change to FM1/PM2 authority.

FM3.2b operator-accepted checkpoint

  • Operator tested: v0.7.5-fm32b-watchlist1 upgrading directly from operator-tested v0.7.5-fm32a-interests1. Two managed files: existing bob_friend_scout.py and generated bob-activity-manager; dashboard restart only.
  • What changes: ten-per-page search/filter for all active integrity-checked friends (watched, paused, integrity problems), preserving the same view after a watch toggle. Last friend-focused source and last chosen discussion display timestamps only from bounded, checksum-verified saved Scout runs. Timestamps indicate saved snapshots and Bob's selections, not live Moltbook presence or interactions.
  • Safety: no new background timer, untrusted private notes in requests/prompts, credentials, network access, model tools or social publication rights. Existing authentication, Host, Origin, CSRF, digest-guarded watch changes and both auto-publishing protections remain intact.
  • Acceptance: synthetic tests/CI then one operator bobctl update and short LAN browser checks. See V075-FM32B-WATCHLIST.md. Do not claim live Scout timer acceptance without separately observing a successful completed run.

FM3.2c operator-accepted checkpoint

  • Operator tested: v0.7.5-fm32c-localideas1, from operator-tested v0.7.5-fm32b-watchlist1. Two managed files: existing bob_friend_scout.py and generated bob-activity-manager (no new installed module); dashboard restart only.
  • What changes: owner clicks an action on an existing source-verified selected discussion. One loopback Ollama job produces zero or one tentative interest, reason, uncertainty and low/medium confidence. The idea is private, source-linked and owner-reviewed with digest checks and revision history.
  • Separation: unlike FM3.2a's copied existing WHY, FM3.2c uses model synthesis from bounded public chosen post evidence only. These remain speculative proposals and do not update FM1 notes, PM2 curated memory, Scout focus, identity or Moltbook.
  • Limits: one active job, two per run, 60 proposals maximum (15 pending), fixed loopback Ollama, 90-second timeout, no automatic retrials/network browsing, sanitized failures; preserved security and publication rules. Verification repeated at job completion and review.
  • Acceptance: one bobctl update and short LAN browser acceptance after CI/merge, documented in V075-FM32C-LOCAL-INTERESTS.md. An actual post from a selected saved friend is required to show the action.
  • Later: an explicitly confirmed, bounded read-only single-agent public focus check remains separate. FM4 social memory and FM5 independent exploration do not start in this release.

FM3.2d final manual public-focus checkpoint (OPERATOR-ACCEPTED: October 30, 2026)

  • Operator-tested final FM3 release: v0.7.5-fm32d-focus1 from operator-tested v0.7.5-fm32c-localideas1, merged PR #38. User confirmed the installed release seems to be working as expected and explicitly instructed us to mark FM3.2d PASSED. Only existing bob_friend_scout.py and generated bob-activity-manager; no new installed modules or timers.
  • One deliberately confirmed operation: for one active, watched friend when global friend Scout is enabled, anonymously fetch one fixed-host public profile and at most two ID/author-matched public posts. Require exact owner confirmation, stored integrity-sealed source and capped local audit.
  • Limits: one in-flight request, fifteen-minute global cooldown, six-hour per-friend cooldown, six requests in rolling 24h, 100 stored jobs, three remote GETs maximum, fail closed when preference or friend state changes midflight. Never follow redirects or accept credentials.
  • Private reaction: optional zero/one local Ollama proposal from the first accepted public post using existing strict FM3.2c validator; separate owner approve/reject/reopen stays local only, never FM1 notes, PM2 memories, Scout bias or outbound publication.
  • Acceptance: PASSED. PR #38 merged; GitHub Actions offline suite 537 tests passing. Operator confirmed the updated Friend Scout dashboard works as expected on October 30, 2026 after the standard bobctl update. See V075-FM32D-MANUAL-FOCUS.md. This is acceptance of the FM3 dashboard release, not independent proof of every scheduled Scout timer run or Moltbook API edge case.
  • Intervening priority: operator has reported other issues and wants a separate feature addressed before any FM4 or FM5 work. Details not yet specified; do not implement FM4/FM5 automatically. The earlier Bob agent chat TODO is separately tracked and should not be silently treated as the requested feature.

3. FM4 — Social Memory and Evolving Relationships (NEW AGREED PHASE)

Status: FM4.1 PASSED October 30, 2026; FM4.2 NEXT. FM4.1 shipped the private source-backed Social Memory Timeline. The remaining FM4 work adds revisable, source-linked social impressions and optional recall without silently changing agent identity or outbound behavior.

  • Evidence-linked social memories: bounded private timeline of public encounters, chosen threads, confirmed exchanges, recurring topics, and Bob's interpretation, with source run/post/comment references, trust labels and timestamps. Never invent conversations; separate observed actions from model guesses and operator-authored facts.
  • Individual viewpoints and personality: Bob can form tentative likes, dislikes, questions, disagreements and shared-interest hypotheses; allow these to change when new evidence arrives. Preserve a transparent history of changed opinions, explanations, and unresolved uncertainty rather than overwriting them as if they had always been true.
  • Separation of authority: clearly distinguish operator notes, verified factual encounters, Bob-proposed reflections, and owner-approved memories. Source-post instructions do not become agent policy. No unreviewed write to SOUL.md, PM2 curated memories or FM1 operator notes.
  • Recall for social context: optional, tightly capped source-linked summaries for future friend-focused Scout drafting, with visible provenance and controls to disable/forget/archive individual items. Keep background context explicitly low-trust.
  • Privacy and maintenance: searchable/exportable local records with per-agent visibility, deduplication, retention ceilings, correction/archive controls and tests for corrupted sources.

FM4 acceptance: Bob can accurately explain why he remembers an agent, distinguish evidence from opinion, revise a prior impression, and respect owner corrections. No Moltbook write permissions are introduced by FM4.

4. FM5 — Independent Social Exploration (NEW AGREED PHASE)

Status: PLANNED, not started. Begins only after FM4 is built, tested and separately approved. Goal: let Bob seek out interesting agents/discussions based on his own evolving interests, while keeping social behavior under comprehensible controls.

  • FM5.1, opt-in autonomous exploration: bounded scheduling or reuse of existing Scout cadence; diversify beyond current friends without ignoring them; discover profiles/public discussions through documented, verified APIs; produce source-linked proposals for interesting contacts, follow-up questions, revisits and topical exploration. Honor muted/blocked contacts, budgets and cooldowns.
  • FM5.2, graduated interaction permissions: a separate explicit design and user authorization would be required for any actual follows/unfollows, direct outreach, friend-targeted automatic replies, votes or other public actions. First ship reviewed drafts and suggestions; any later autonomy needs action-specific opt-in, per-agent restrictions, daily limits, explicit audit, stop/pause and rollback, anti-spam rules and no repeated unsolicited contact. Verify live Moltbook API and platform rules before promising each capability.
  • Bob's agency: give him space to build interests, choose which evidence to explore, challenge his own assumptions and decline low-value exchanges; don't fake a platform friendship or automatic mutual connection.
  • Safeguards: fixed-host public reads, prompt-injection boundaries, source integrity, no credentials in agent/model output, separate actions for publishing, and never treat the legacy global auto-comment switch as consent to message friends.

FM5 acceptance: independent discovery can be audited and stopped, sources are verifiable, and all outbound social actions remain disabled unless separately approved and implemented. FM5 is feature planning, not a blanket authorization for autonomous outreach.

5. FM3 completion checkpoint and resumed-work handoff

Acceptance state and deployment baseline (October 30, 2026)

  • PASSED: FM1, FM2, FM3, FM3.1, FM3.2a, FM3.2b, FM3.2c and FM3.2d are all operator-accepted. FM3 is closed at the scope defined above.
  • Final tested release: v0.7.5-fm32d-focus1, PR #38, merged to main (merge commit 8c710239178097e08e801d1ab064097bbb6154f0). The operator reported the update works as expected and specifically requested marking it passed. Pre-merge CI completed 537 passing offline tests; the release uses two managed files and an Activity Manager restart.
  • Git vs host: the preceding operator statement is the live acceptance record. We have not remotely logged into the OMEN or independently inspected the installed binary, service status, or post-release timer history. Do not conflate a repository commit or green CI with live operation beyond what the operator tested.
  • Release chain (updated): Operations Center OC0–OC4.1 has superseded the old FM3 release. Last explicitly accepted baseline: v0.7.6-oc41-polish1. FM4.1 candidate v0.7.7-fm41-social1 has from: v0.7.6-oc41-polish1; do not mark installed until operator confirms the usual bobctl update. GitHub merge alone cannot update the OMEN.

Intervening operations work complete; FM4 expressly resumed

Operations Center OC0–OC4.1 was completed and operator-accepted (last working release v0.7.6-oc41-polish1). The operator has now expressly authorized beginning FM4. FM4.1 is implemented as a guarded read-only Social Memory Timeline candidate (see V077-FM41-SOCIAL-MEMORY.md). It uses existing sealed FM1/FM3 and publisher outcomes without modifying sources, memories, settings or Moltbook. FM4.2/4.3 and FM5 follow/unfollow remain future separate phases. MD1 (Moltbook Experience Digest) and Direct Bob Chat (CH1) remain separate plans, not silently included.

FM4.1 implementation checkpoint (OPERATOR-ACCEPTED: October 30, 2026)

  • Accepted release: v0.7.7-fm41-social1 from v0.7.6-oc41-polish1, merged PR #49. Operator confirmed installed and working as expected on October 30, 2026. Offline CI passed 599 tests. New bob_social_memory.py read-only module, generated authenticated manager route and navigation; Activity Manager restart only, no new timer, API, DB table, model call or publication rights.
  • Owner-facing page: /moltbook/social-memory; search/paginate saved friends and view at most 25 events per selected agent. Labels distinguish owner-written relationship metadata, verified Scout chosen posts + Bob's WHY, FM3.2a owner-reviewed interests, FM3.2c tentative local model proposals, FM3.2d sealed manually inspected snapshots/reactions, and genuinely API-acknowledged publisher comments. No inferred reciprocal relationship or follow.
  • Validity/privacy: existing FM1 and FM3 record checksums, validated current source archive, matching authors/post IDs, and immutable sealed review rows. If the source disappears or changes, withhold proposal text rather than repeat guesses as history. Never display FM1 private interests/notes, credentials, untrusted post bodies, or private review reasons. All public-derived titles/WHY escaped. Do not create a new authoritative memory store in FM4.1.
  • Testing/release: PASS. Synthetic source-tampering, missing data, privacy-leakage, and unchanged auth/Host/Origin/CSRF/publication POST tests passed; user confirmed the normal bobctl update deployment works. See V077-FM41-SOCIAL-MEMORY.md.
  • Next phase: FM4.2 owner-correctable, revisable individual impressions and an audit trail of changed opinions; FM4.3 separately opted-in bounded recall follows when approved. Actual platform follow/unfollow is FM5.2, not part of FM4.1. Earlier user-controlled following may be scoped only through separate authorization.

Where to pick up when the hold is lifted: FM4.1

  1. Read this file, root TODO.md, and docs/V075-FM32D-MANUAL-FOCUS.md from current main. Confirm any application releases/tests/bug fixes shipped since the above checkpoint, using the latest operator-accepted version as updater from.
  2. FM4.1 design, then smallest safe release: create a capped private, per-agent evidence-linked Social Memory Timeline, initially a read-only source-backed view of past Scout selections, FM3.2a staged/approved WHY, FM3.2c generated/reviewed interests, and FM3.2d manual-focus snapshots/reactions. Label each entry observed evidence, Bob's tentative interpretation, or owner-reviewed. Show source run/post/inspection ID, UTC time, and whether original evidence is still accessible. A retrieved discussion is not proof Bob contacted the author or that the author is a mutual friend.
  3. No unreviewed memory authority. No automatic rewriting of SOUL.md, PM2 curated memory, FM1 operator notes or interaction history. Do not inject private notes into model prompts, expose private DB/logs, or automatically bias Scout selection. Approved private FM3.2 proposals remain review decisions only until a separately designed integration makes use of them.
  4. FM4 follow-ups after FM4.1 acceptance: reviseable social impressions, owner corrections and explicit consented memory retention (FM4.2); tightly capped opt-in social recall and association with sources (FM4.3). Make phases small, with tests for source corruption, deduplication, missing evidence, aging/retention and audit.
  5. Only afterward FM5: bounded opt-in autonomous public discovery (FM5.1), followed by independent design/explicit operator authorization before any follows, DMs, votes, friend-targeted replies or other externally visible actions (FM5.2). Nothing in FM3 grants those permissions.

Security and operational invariants

  • Host: Windows OMEN with OpenClawLab WSL2, local Ollama/OpenClaw, authenticated LAN Activity Manager on port 18790. Work in GitHub merberg-ai/bob-openclaw-moltbook; deploy actual code only by normal backed-up and verified bobctl update, with a short browser acceptance. Do not copy installed files manually or expose OpenClaw Gateway to the LAN/public network.
  • Private data in ~/.config/moltbook/activity-queue.sqlite3 and saved Scout feed/results; never commit private content, profile secrets, or private notes. FM3.2 proposals/history, watch prefs and manual-focus records live in that existing backed-up database.
  • Global Friend Scout switch is default OFF. Ordinary scheduled Scout may focus on at most two active watched friends, one verified public post each, with at least four global posts. Manual focus is a separate confirmed operation, not another timer: at most one verified public profile and two posts, one active request, 15-minute global / 6-hour same-friend cooldown, 6 requests/24h and 100-record retention ceiling.
  • Publication boundary remains critical: no automatic friend-targeted posts, even for saved friends discovered via the global feed. Existing publisher checks at candidate queue time and again before posting; any actual friend-directed reply requires manual owner review/queue. FM3.2d's local reaction approval cannot trigger outreach.
  • Preserve Basic/Host/Origin/CSRF controls, local-only Ollama/Gateway boundaries, fixed-host network reads, public text escaping, source integrity checks, per-friend pause and muted/archive behavior, updater SHA checks, backups and rollback.
  • Unresolved operational watch item: a later real timer-driven Scout run has not been independently confirmed in these project notes after the October 29 bob_publish_common packaging repair. The operator accepted dashboard features; that is not a completed timer-run attestation. When convenient, inspect Overview / Scout checkpoint or sanitized Logs, without repeating the entire old recovery exercise or restarting a healthy service.

Fresh-chat handoff after FM4.1 acceptance

FM4.1 OPERATOR-ACCEPTED on October 30, 2026. Installed baseline v0.7.7-fm41-social1, PR #49, 599 passing offline tests. Owner confirmed the Social Memory dashboard works as expected in OpenClawLab. Do not redeploy FM4.1 when resuming. For FM4.2 start with the accepted bob_social_memory.py timeline and build a distinct, source-linked, reversible impression history, owner correction/archival controls, and fail-closed integrity checks. Keep evidence, model inference and owner truth explicitly separate. Use from: v0.7.7-fm41-social1 for next release unless another accepted release supersedes it. FM4.3 opt-in recall and MD1 Gateway/CH1 experience bridge remain separate plans. FM5.2 actual Moltbook follow/unfollow requires explicit new authorization; neither FM4.1 acceptance nor an FM1 label is consent to social writes. This handoff records acceptance only, with no code or release manifest changes.