Decisions
Decisions
The records behind the AI platform: serving, the agent factory and the split from its host. They moved here from cloud-native-ref with their numbers (split design §8); the next record here is 0059. Host decisions (Flux, Cilium, OpenBao, EKS Pod Identity, …) stay on the host site and are linked, not copied.
A technology choice with a rejected alternative needs a record before merge. Version bumps, chart-value changes and single-file fixes never do. Start from the host’s template.
The records
| ADR | Title | Status | Date |
|---|---|---|---|
| 0003 | Use vLLM Production Stack over KServe + llm-d for v1 LLM Platform | Accepted | 2026-04-30 |
| 0004 | Use Amazon S3 Files for LLM Model Weights Storage | Accepted | 2026-05-01 |
| 0021 | Cloud Storage FUSE for LLM model weights on GCP | Accepted | 2026-08-26 |
| 0041 | Coding agents run in agent-sandbox Sandboxes under gVisor, on a dedicated spot pool per cloud, with OpenHands as the harness profile | Accepted | 2026-09-26 |
| 0042 | Agent Router is the agents’ identity gateway, with role and data class encoded in the token audience and an in-pod proxy holding the tokens | Partly superseded by 0053: Option 1, agent router only | 2026-09-26 |
| 0043 | Agents get GitHub tokens from a self-hosted octo-sts, scoped per repository and role, and two rulesets confine their App to agent branches and no tags | Accepted | 2026-09-26 |
| 0044 | Rooms are an AHP-shaped log we own, stored in CNPG, fanned out with Postgres LISTEN/NOTIFY | Accepted | 2026-09-27 |
| 0045 | palantir/policy-bot, required through a repository ruleset, decides which agent PRs merge | Accepted | 2026-09-27 |
| 0046 | Frontier models through Z.ai and keyless Anthropic per cloud (Bedrock on aws-0, Vertex on gcp-0) | Partly superseded by 0054: internal-data provider only | 2026-09-25 |
| 0047 | Reserved, not yet written: how a request’s complexity is classified for routing (SP4) | Reserved | — |
| 0048 | A custom Task controller with Kueue admission orchestrates the agent factory | Accepted | 2026-09-27 |
| 0049 | Humans reach rooms through a web UI the broker serves, behind oauth2-proxy | Accepted | 2026-09-27 |
| 0050 | Token budgets on Envoy Gateway’s global rate limit, costed from response metadata | Partly superseded by 0053: Option 1, agent router only | 2026-09-25 |
| 0051 | An OpenTelemetry Collector is the agent trace gate | Accepted | 2026-09-27 |
| 0053 | The agent router runs on agentgateway, pinned to a minor that passes our live gate; ai-gateway stays on Envoy Gateway | Accepted | 2026-10-01 |
| 0054 | Internal agent work calls the Anthropic API directly; public work stays on Z.ai, with OpenRouter optional | Accepted | 2026-10-01 |
| 0056 | The factory is driven from the developer’s local agent through a skill and roomctl, and a room is as visible as its repo on GitHub | Accepted | 2026-10-08 |
| 0057 | The AI platform moves to its own repository, consumed by cloud-native-ref as a pinned artifact | Accepted | 2026-10-12 |
| 0058 | The factory’s reviewer on public tasks keeps the implementer’s GLM model | Accepted | 2026-10-16 |