Synexar management plane · Paperclip parity · 2026-07-05

Paperclip — the other parts

The collaboration layer is shipped. Paperclip's remaining value is a management plane over the commanders — and two of its pieces map straight onto scars we've already felt. This board is the plan for the rest. Tap any card to steer it.

✅ Already handled — don't rebuild these
Collaboration & observability surface — live Artifact boards + feedback tray + Board Wall (this session). Paperclip's "manage from your phone" is now our board-link workflow.
Governance, verification, isolation, goals — Synexar is AHEAD here: domain-aware bright lines (billing/RLS/PHI), the fresh-context delivery-verifier (Paperclip only tracks activity, we prove staging outcomes), git-worktree isolation, and goals-with-ancestry (CHARTER→GOALS→Pact). Paperclip is the design reference, not a replacement.
🔧 The parts still to address — ranked by our own pain
① Budget & cost hard-stopsDO NOW · P0
Synexar todayNo structural spend guard. Agents/loops can run away on tokens with nothing to stop them.
ProposedA per-agent / per-goal / per-run token+cost ledger with a hard-stop: when a cap is hit, the loop pauses & queued work cancels. Mirror Paperclip's atomic budget-check-at-checkout, wired into /autorun + the commander spawn path.
🩹 Scar: the $1,081 CI-LLM cost incident (2026-06-11) — exactly the runaway this prevents.
Effort: medium · lives in synexar-runtime, above /synexar-task · highest pain-to-effort ratio → the clear first move.
② Reliable heartbeats & schedulingDO NOW · P0/P1
Synexar todayScheduled routines are broken & silent: daily-digest.yml down, X-scan silent since Jun 12 — and nothing tells us they stopped.
ProposedA DB-backed wake queue with coalescing + orphan-run recovery (Paperclip's heartbeat engine), OR the lean version first: fix the crons + a "last-ran" heartbeat monitor that alerts on a missed beat. Every routine run becomes a tracked ticket.
🩹 Scar: routines fail silently — the dangerous kind (like the bird X-auth partial-failure). A heartbeat monitor makes silence loud.
Effort: medium · decision inside: lean (fix + monitor) vs. full (DB queue). I'd start lean, earn the full version.
③ True auto-observability dashboardNEXT · P1
Synexar todayThe Board Wall is curated, not automatic (artifacts can't self-index). No live ops view of agents, cost, or routine health.
ProposedThe self-hosted /runtime dashboard queries a datastore: commander state + the new budget ledger + heartbeat health + fixer audit. PHI-capable, sandbox-free. Becomes worth building once ① & ② produce data to show.
Effort: larger · sequence AFTER budget + heartbeats (they generate its content).
④ Cross-commander coordination (org chart + delegation)LATER · P2
Synexar todayCommanders are siloed; Demis synthesizes across them only on-demand. No standing delegation between domains.
ProposedA light coordinator layer — org-chart-style delegation so a commander can hand a cross-domain task to another (Paperclip's reporting-lines idea), with the bright-line fences intact.
Effort: larger · lowest urgency · revisit once the loop reliably self-runs.
Secrets · plugins · portability · identityALREADY COVERED
Why no work neededKey Vault + managed identity (secrets), MCP/skills (plugins), worktrees + Pact travel (portability), Entra/RLS (identity). Paperclip's versions don't beat ours. Noting so they're consciously ruled out, not forgotten.
🗺️ Recommended sequence
① Budget hard-stops→ ② Heartbeat monitor (lean)→ ③ /runtime auto-dashboard→ ④ Cross-commander delegation
📌 The call I'm recommending: build these into synexar-runtime ourselves, with Paperclip as the design reference — don't adopt its code. Reasons: (1) it's a 4-month-old pseudonymous project (dependency risk), (2) ours must be HIPAA- & domain-aware in ways a generic tool isn't, (3) we already have ~70% — filling three gaps in our own system beats migrating to theirs. The management plane sits ABOVE /synexar-task; the execution fences never move. (Disagree? Flag this card — the alternative is piloting Paperclip itself as the plane.)