| name | control-plane-maintenance-base |
| description | Shared maintenance topology and policy for aicoder-opencode — one server per repo, shared plugins, overlay-first changes |
| user-invocable | false |
| models | ["ollama-cloud/kimi-k2-thinking","ollama-cloud/glm-5.1","ollama-cloud/minimax-m2.7"] |
| routing_role | architect |
| routing_complexity | large |
control-plane-maintenance-base
Shared maintenance/control-plane base for repos that attach to the aicoder-opencode
server.
Use this skill when work is about:
- shared maintenance workflows
- shared OpenCode control-plane behavior
- cross-repo maintenance plugin baselines
- per-repo server topology that should stay consistent across product repos
Responsibilities
- Keep one OpenCode server per repo.
- Keep shared maintenance logic in
aicoder-opencode.
- Push repo-specific overlays back into the product repo only when they are truly product-specific.
- Prefer thin install shims in product repos over copied plugin implementations.
- Change maintenance topology in small reversible steps and verify one target
repo path before rolling the same pattern wider.
Contract
aicoder-opencode owns shared maintenance plugins and skills.- Product repos may install shared base shims, then add local overlays.
- Control-plane changes should not silently fork per repo.
Topology
aicoder-opencode on 8080dr-repo on 8082letta-workspace on 8084- OpenChamber attaches to one repo server at a time
Shared Base
- Shared plugin source belongs in
src/plugins/
- Shared skill source belongs in
.agents/skills/
- Product repos consume shared base via repo-local shims
Runtime: Model Provider Priority
Use config/models.jsonc as the authoritative allowlist. Only models listed there are approved. Provider priority within each registry entry follows provider_order[].priority.
Key providers in the registry:
| Provider | Status | Notes |
|---|
ollama-cloud/* | Free (rate-capped daily) | Best when cap not exhausted |
iflowcn/* | Free | Major free provider; priority 1 for many registry entries |
mistral/* | Subscription | codestral, devstral variants |
minimax/* | Subscription (minimax.io) | Reliable fallback |
kimi-for-coding/* | Subscription | User-Agent gated, works via opencode |
xiaomi-token-plan-ams/* | Subscription (direct) | mimo-v2-pro priority 1 via token-plan-ams.xiaomimimo.com |
openrouter/:free | Free tier | Allowed; use curated models only |
openrouter/* (paid) | Pay-per-token | Last resort; only models in registry |
longcat-openai/* | Free | Special use only — not general fallback |
Never use: xai/grok-*, openai/* (ChatGPT), togetherai/*, uncurated openrouter auto-selection.
Autopilot model selection (per task matrix)
| Task type | Primary | Fallback |
|---|
| Autonomous coding iteration (dr-repo) | ollama-cloud/qwen3-coder:480b | xiaomi-token-plan-ams/mimo-v2-pro |
| Maintenance / type fixes (letta-workspace) | ollama-cloud/devstral-2:123b | xiaomi-token-plan-ams/mimo-v2-pro |
| Fast general iteration | ollama-cloud/kimi-k2.5 | xiaomi-token-plan-ams/mimo-v2-pro |
| Architecture / heavy reasoning | ollama-cloud/kimi-k2-thinking | xiaomi-token-plan-ams/mimo-v2-pro |
Model selection is handled by OpenCode plus the model-registry plugin. The plugin can reroute task launches when the requested provider is in quota backoff, when an agent exposes ordered models: metadata, or when the task falls back to curated registry routing. Do not override model choice with env vars.
Runtime: Model Selection
The model-registry plugin exposes two tools:
list_curated_models — all registry entries; use freeOnly: true to scope to free providersselect_models_for_role — models recommended for a named role (e.g. "autonomous-agent", "code-review")get_quota_backoff_status — providers currently in quota backoff and when they expire
Rules:
- Before starting a long autonomous task, call
get_quota_backoff_status. If a needed provider is in backoff, let the plugin reroute task launches using the current agent models: list or the curated registry fallback.
- Never fall back to
longcat, claude, gpt, or grok models automatically. These are subscription-gated or architecturally undesirable as automatic fallbacks.
- Free providers (ollama-cloud, openrouter free-tier) are preferred for autonomous/maintenance loops. Paid providers (openrouter paid models) only when no free alternative covers the capability tier.
- Temperature is set automatically by the
model-registry plugin based on the model's capability_tier. Do not override temperature in prompts.
Runtime: Autonomous Execution
These rules apply to all repo maintenance sessions (dr-repo autopilot, letta-workspace maintenance, aicoder unblock-and-patch):
- Default to sustained execution. Do substantial work before reporting back.
- Take your own best evidence-backed recommendation for reversible decisions.
- Ask only when the next action is destructive, irreversible, or preference-shaped in a way that materially changes the path.
- If confidence is low on a critical sub-decision, spawn a specialist subagent for review before proceeding — do not interrupt the human.
- When a maintenance loop fires and finds no pending work, log and exit cleanly. Do not spin on empty queues.
- If a session was previously interrupted, re-read the last checkpoint or session-id file before starting new work to resume cleanly.
Autopilot continuation
When the autopilot timer fires:
- Check if the runner unit is already active — if yes, exit immediately (do not start a second runner).
- Check the session-id file. If it contains a valid session ID, resume that session with
--session <id>.
- If no session-id file exists, start a new session with the configured title.
- Use
get_quota_backoff_status before starting a long chain — if the configured model's provider is in backoff, switch to the next provider in the priority table above.
- On quota error (429) mid-session, the model-registry plugin records the backoff. The next timer cycle will start a fresh attempt — do not retry in a tight loop within the same session.
Runtime: Quota and Provider Events
The plugin tracks quota events automatically and backs off providers for 1 hour on HTTP 429 or quota-keyword errors. No manual action is needed during normal operation.
If you observe a provider persistently failing (beyond the 1-hour window), call get_quota_backoff_status and route around it using the curated fallback in provider_order.
Verification contract
Downstream skills and agents that consume this base inherit the following verification rules:
- Every verification report ends with a final line:
VERDICT: PASS, VERDICT: FAIL, or VERDICT: PARTIAL. Plain text, no markdown, no hedging.
- PARTIAL is reserved for environmental blockers only (tool missing, service will not start, no test framework). Ambiguous findings are FAIL, not PARTIAL.
- A PASS requires at least one executed adversarial probe across concurrency, boundary values, idempotency, or orphan/dangling state — reading code is not verification.
- Each check must record the exact command run and the observed output. Uncited claims do not count as verified.
See targets/dr-repo/overlay/.opencode/agents/verifier.md for the full probe checklist and output block format.
Avoid
- Copy-pasting shared plugin logic into product repos
- Moving product doctrine into
aicoder-opencode
- Adding extra OpenCode servers when one per repo is enough
- Large topology rewrites without verifying the next real maintenance path
- Overriding temperature manually (the plugin sets it from capability tier)
