zlh-grind/OPEN_THREADS.md

122 lines
5.8 KiB
Markdown

# Open Threads — zlh-grind
This file tracks **active cross-repo and platform-level work only**.
Repo-specific work belongs in:
- `Codex/API/OPEN_ITEMS.md`
- `Codex/Portal/OPEN_ITEMS.md`
- `Codex/Agent/OPEN_ITEMS.md`
Keep this file short.
---
## Cross-Repo Active
### Final launch smoke test
- create Minecraft server
- confirm it reaches `Ready` / `connectable=true`
- verify public game hostname is shown only when connectable
- upload datapack on vanilla or install mod on supported modded runtime
- create backup
- restore backup
- stop/start/restart host lifecycle actions
- delete server
- confirm Velocity unregister, Cloudflare cleanup, and Technitium cleanup
### Backup / restore polish
- happy-path local Minecraft backup create/restore has been verified live
- API restore starts asynchronously and Portal polls restore status
- keep restore wording/status transitions clear through completion and restart
- confirm checkpoint metadata presentation remains clean when exposed to Portal
- later hardening: persist last restore failure/checkpoint state in Agent `/status`
- later hardening: automatic rollback from pre-restore checkpoint if restore apply/start fails after destructive replace
### Service discovery / launch validation
- service discovery migration audit for remaining non-launch hot-path references
- provisioning validation across current API/Agent/Portal assumptions
- keep public exposure model explicit:
- Portal public
- Minecraft game hostnames public as needed
- API/control plane/internal bridge/agent/admin services private
### Monitoring / observability
- normalize game/dev Alloy monitoring contract across API discovery, agent-written Alloy labels, Prometheus targets, and Grafana dashboards
- keep dynamic game/dev discovery on API -> sync script -> file_sd and verify automatic add/remove behavior for new containers
- finish game/dev template cleanup so Alloy is standard and `node-exporter` is removed from those templates
- keep OPNsense plugin and PBS monitoring as explicit platform exceptions while Linux-managed targets converge on Alloy
### Notifications / launch polish
- email notifications across backend contract + Portal UX
- billing launch validation:
- plan limit gating verified in Portal
- still verify checkout/portal/webhook/upgrade-downgrade if Stripe is live
---
## Platform / Infrastructure Active
- stress testing: k6 IDE load, Minecraft bot load, code-server memory baseline
- OPNsense / public exposure audit
- billing endpoint/path cleanup verification
### Backup boundary
- Agent-owned backups are local, app-aware rollback backups for Minecraft worlds/config
- PBS / platform backup strategy is the durability / disaster-recovery layer
- do not track PBS/offsite durability work as agent implementation work unless that ownership changes
---
## Recently Verified / No Longer Considered Blocked
- password reset and logged-in change-password work end-to-end
- password reset tokens are 5-minute, hashed at rest, single-use, and old unused tokens are invalidated on deploy
- API-owned Minecraft connection state derives from agent readiness, edge/DNS state, Velocity registration, and backend ping
- Velocity proxy lifecycle callbacks are live with `registered_with_proxy` and `proxy_ping_ok` landing in API state
- Portal consumes API-owned `connectable` / `connection` state and no longer infers Minecraft readiness itself
- Portal server creation redirects to `/servers` and tracks setup progress there
- Portal status labels no longer treat all non-connectable states as `Needs attention`
- local Minecraft backup create/restore works end-to-end on live validation
- restore creates intentional pre-restore checkpoint and API now starts restore asynchronously instead of holding the full request open
- backup timestamps are normalized and pre-restore checkpoints are filtered from the default backup list
- agent-backed file edits create shadow copies for revert and API route/stream forwarding issues were fixed
- vanilla datapack upload works
- vanilla Mods UI is hidden and direct vanilla `mods/` upload is rejected by API
- NeoForge mod search/install/list works
- delete/teardown lifecycle removes Velocity, Cloudflare, and Technitium records
- public exposure model is in place: Portal public, control plane private
- vanilla / fabric runtime split is restored:
- `vanilla` = Fabric-based internal profile with proxy/API/config injection
- `fabric` = plain Fabric jar delivery only
- Forge / Neoforge first-start flow avoids premature readiness gating, applies post-start property enforcement, and restarts through the readiness-aware path
- current validation indicates Minecraft server creation succeeds across supported runtime variants
- current validation indicates dev container creation succeeds and hosted IDE access still works after the latest API/Portal runtime and cleanup passes
---
## Platform Future / Phase 2
- SSH / CF tunnel power-user access
- Portal SSH config snippets
- true interactive shell confinement / workspace-boundary decision
- dev-container backup ownership and user-facing restore contract
- likely direction for dev backups: LXC snapshot-based backup/restore instead of agent-managed dev backups
- artifact version promotion
- runtime rollback support
- Cloudflare R2 for large artifact/mod delivery
- admin panel
- referral / dev pipeline reward system
- uptime history
- revisit DDoS mitigation later if needed
---
## Cleaning Rule
- Root keeps only cross-repo/platform work
- Repo-specific items must be removed from root once they live only in one Codex tracker
- Completed items should be removed, not left in place as historical clutter
- Use `CURRENT_STATE.md` for durable implemented behavior
- Use `DECISIONS.md` for settled choices
- Re-open old items only when there is current evidence they are still unfinished or have regressed