carbon-panel

Imported from athNdev/carbon-panel (AGENTS.md). Install upstream with npx skills add athNdev/carbon-panel. Copyright stays with the author.

AGENTS.md — carbon-panel agent operating contract

Carbon-panel is a Go backend + SvelteKit web UI for Minecraft server orchestration (Docker-isolated servers, single-port TCP proxy, Connect-RPC API).

Build / test (what CI runs)

• Backend: make test (== go test ./...), plus go vet ./....
• Frontend: cd web/carbon-panel && bun install --frozen-lockfile && bun run check && bun run test && bash ../../scripts/web-build.sh
• Proto: make proto-lint (runs buf lint via Docker); make check type-checks the frontend.
• See Makefile targets (test, fmt, lint, check) and README.md for dev flow.

MANDATORY session protocol (Tier 0 persistent context, file-native, no UI)

Every agent session MUST do session-start and session-end. Total start budget: <=3.5K tokens.

Session-start (read, in order):

1. This file (AGENTS.md).
2. memory/index.md, then memory/decisions.md (active rows only).
3. Semantic recall: qmd top-5 for the task if the qmd daemon is available, else grep -ri "<keywords>" memory/.
4. Full recipe: .agents/skills/memory-ops/SKILL.md.

Session-end (write, <=0.5K tokens):

1. Append one entry to memory/YYYY-MM-DD.md: what changed, why, commit SHA, Plane IDs.
2. Decision-grade changes: also append one active row to memory/decisions.md.
3. Never rewrite history: supersede old rows/logs with new entries, don't edit them.

Product truth

• Roadmap and why live in Plane project MINE (MineServer). Markdown links to Plane work-item IDs; it never duplicates them.
• Durable architecture decisions live in docs/adr/NNNN-title.md.
• Runbook: docs/context-layer.md.

Hard rules

• No secrets, PII, or tokens ever in memory/ or skills. No exceptions.
• Per-project isolation: this context is carbon-panel only; never import other projects' memory.
• Personas/reviewers are read-only: no memory writes, no Plane edits.

Cluster-Wide Standing Rules (Mandatory Across All Agents)

All agents operating in this repository must strictly comply with the cluster standing rules defined in /home/prox/agent_setup/AGENTS.md:

1. Cross-Node Collision Avoidance (agentplane): Run agentplane check <path> before touching files/resources; claim long-running tasks and release them when done.
2. Pre-Task Consultation: Query agent-context (qmd query "<topic>" or qmd search "<term>") before performing cluster, infrastructure, or service operations.
3. Post-Action State Updates (MANDATORY COMPLETION GATE): Any task modifying cluster configuration, infrastructure, services, tooling, skills, provider configs, or architecture MUST update agent-context:
   • Document changes in /home/prox/context-corpus/ (runbooks/, decisions/, sessions/).
   • Sync and reindex: /home/prox/agent_setup/context/context-sync.sh
   • Verify health: /home/prox/agent_setup/context/context-healthcheck.sh
   • No task modifying system state is complete without completing this step.
4. Subagent Delegation via t3-lb: Use t3lb (t3lb spawn, t3lb wait, t3lb nodes) to distribute heavy or parallel workloads across cluster nodes.
5. Forgejo System of Record: Regular git pushes to local Forgejo (port 2222) and run localrepo snapshot before ending sessions.
6. Default Token-Saving Communication: Caveman Mode: All agents operate in Caveman Mode (~65–75% token reduction):
   • Terse, fragment-based, zero conversational filler, no pleasantries.
   • Code blocks, commands, paths, diffs, and errors remain 100% exact and uncompressed.
   • Immediate tool execution without narrative announcements.
   • Switch to full sentences only for safety warnings, destructive actions (rm -rf, DROP TABLE), or when user requests "normal mode".