Skip to content

todos.md

Latest commit

 

History

History
66 lines (51 loc) · 5.41 KB

todos.md

File metadata and controls

66 lines (51 loc) · 5.41 KB
Raw

Todos

Editor

  • Editor "Agent not found" on expired sessions — the side-panel editor renders "Agent not found" (not a re-auth prompt) when the WorkOS token lapses mid-session: agents.get 401s → loadFailed → not-found copy. Consider distinguishing auth failure from genuinely missing agents.

Harness implementation (archived)

The managed harness agents rollout (Vercel Sandbox) was completed and then the feature was removed; the rollout history lives in git history.

Self-hosting roadmap

Context: Convex is the only app-data provider (Phase 3 done). Enterprise self-hosting runs the official convex-backend (examples/customer-deployment/), and the SQL read mirror ships in docker-compose.mirror.yml (Phase 4 done). Architecture + rationale: internal/develop/convex-only-self-hosting.md. Frozen escape hatch: postgres-self-host branch pinned at 63cf2edb4 — the last commit with the full bespoke Postgres provider (deleted on main); only touch it if a customer explicitly requires a Postgres app-data backend.

  • Confirm FSL scope with Convex before building — email/Discord: orchestrating convex-backend instances that exist only to serve Overlay deployments (including in customer premises) vs. offering managed Convex as a product. Get the answer in writing; enterprise procurement will ask.
  • Big Brain — per-customer lifecycle manager (Phase 5, build when a concrete enterprise need lands). Scope v1 to single-customer deployments: provision (instance name → schema → INSTANCE_SECRET → admin key → container), health-check via /version, image upgrades with pre-upgrade npx convex export, scheduled export backups, standby promote on lease loss. Written against four portable interfaces — containers, Postgres provisioner, object store, DNS — so it works on any cloud. The OSS repo's big_brain_client crate shows the backend is designed for external control.
  • Big Brain — central fleet plane — deferred harder: per-tenant routing, metering (/metrics, off by default), and multi-instance scaling only matter if Overlay hosts many tenants itself. Re-read the FSL before any managed-hosting shape; this version is closest to the license line.

Overlay Cloud agents (Phase 6)

  • Test the E2B adapter against real E2B — it is only tested against a fake SDK. Needs an E2B_API_KEY; set OVERLAY_MANAGED_SANDBOX_PROVIDER=e2b and OVERLAY_CLOUD_AGENT_IMAGE=overlay-agent:v5, run npx tsx infra/agent-image/build-e2b.mts, then create a cloud agent and check: create, wake after idle, pause, delete, writeFiles (Codex auth.json), and the E2B rate card in ManagedAgentSandboxBilling.ts. See internal/develop/codex-subscription.md.
  • Hermes on Overlay Cloud — dropped (2026-10-03). Boat's lazy Hermes install fails, and pre-installing with UV_PYTHON_INSTALL_DIR / UV_CACHE_DIR on the real disk did not help: the base layer's filesystem drops files while uv unpacks (EXTERNALLY-MANAGED, _sysconfigdata_, wheel .dist-info errors). Hermes stays hidden from the pickers (SELECTABLE_AGENT_PROVIDER_IDS); retry only if Boat's base layer changes.
  • Cursor agent live test — needs a CURSOR_API_KEY.
  • Codex live test — needs an approved ChatGPT device code; then two Codex agents on one account.

Workspace billing (deferred)

  • Per-agent spend caps (deferred 2026-10-06). A workspace with a plan draws every run from one shared pool; per-member limits already exist (configureSpendLimit). A cap per agent or automation would stop one runaway resource from draining the pool. Not in the first version of workspace billing: build it when a pool is actually shared by several people. Needs: a limit keyed by resource id beside the member limit, checked in the same reservation path as reserveWorkspace, and a row on the workspace billing tab.

Admin console

  • Separate WorkOS environment for the admin console (added 2026-10-14). Today the owner is the only operator, so apps/admin signs in through the existing WorkOS environment, gated by ADMIN_ALLOWED_USER_IDS and administrativePrincipals. Before anyone else gets operator access, move admin sign-in to its own WorkOS environment: SSO through company Google Workspace, MFA required, no password or magic link. Plan: internal/plans/ADMIN_CONSOLE_SEPARATION_PLAN.md.
  • Staff environment for other team members (added 2026-10-14). When more people join, give them their own staff setup (accounts, roles in administrativePrincipals, access to the overlay-admin Vercel project) instead of sharing the owner's.

Agent permission cards

  • MCP approval gate (Overlay tools called by Claude Code / Codex over MCP) still lets the agent move on. mcp-approval-gate.ts waits 25 s for the answer, then returns "waiting for approval, call again" and the agent keeps responding meanwhile (hosts time MCP calls out in about a minute). The permission cards for the agent's own shell/file/tool requests now stop the transcript at the card (partsUntilPendingRequest); this path needs a real pause (progress notifications to extend the host's timeout, or holding the call until answered). Cards stay; improve, do not remove (owner, 2026-10-07).
  • Hold the agent's own output while a card is pending, on the host. Today only the view stops at the card; sibling tool calls the agent started in parallel still run and appear after the answer.