No description
- Rust 88.5%
- CSS 7.7%
- Starlark 2.6%
- Shell 0.8%
- Go Template 0.3%
New crate df-agent-channel (openspec iter1-run-orchestration): the control-plane side of the agent protocol, mounted by `dark-factory serve`. - Agent port: mTLS via tokio-rustls with client certs bound to the run (URI SAN, peer IP via SandboxProvider::resolve_identity). TLS material is reloaded on rotation (checked at most every 10 s), handshakes are bounded in time and count. - Session: Hello/HelloAck, HarnessCredentials and Assign, batched persistence with cumulative acks, dedup and gap detection, transcript storage with blob extraction, prompt (re-)delivery and wall-clock Cancel counted from the runtime's first Hello. Effects of committed messages run even before a gap; tool calls and credentials are refused once the run ended. - Scheduler: SKIP LOCKED run start on task_message_queued and a 15 s tick, background provisioning, provider events, connect timeout and hard stop; continuation summary keeps the newest messages. - Factory tools with schema validation and at-most-once execution (the forge is called outside any transaction), PR lifecycle (draft on push, ready + completion summary on completion), Git credential grants. - Graceful shutdown: sessions drain in-flight tool calls, the scheduler drains provisioning, main waits SHUTDOWN_GRACE before aborting. - Migration: tasks.completion_summary, runs.connected_at. - Config: agent and harness sections deny unknown fields; ChannelConfig holds the df-core sections; Helm configmap sets agent.tls. - Placeholder provider, forge and secret store until iter1-sandbox-provisioning and iter1-integration-plugins land. - Bazel: production library without test-util, separate test-util variant for tests. Remote Bazel is blocked by a pre-existing toolchain issue (docs/bazel-remote-toolchain.md); verified locally. - Deferred review findings are tracked as TD-8 to TD-26 in docs/tech_debt.md. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
||
|---|---|---|
| .agents | ||
| .cargo | ||
| .claude | ||
| .codegraph | ||
| .pi | ||
| .sqlx | ||
| build | ||
| contract | ||
| crates | ||
| deploy | ||
| design | ||
| docs | ||
| openspec | ||
| plugins | ||
| product | ||
| tools/xtask | ||
| .bazelignore | ||
| .bazelrc | ||
| .bazelrc.agents | ||
| .bazelversion | ||
| .editorconfig | ||
| .envrc | ||
| .gitignore | ||
| .hadolint.yaml | ||
| .kube-linter.yaml | ||
| .mcp.json | ||
| _typos.toml | ||
| AGENTS.md | ||
| apm.lock.yaml | ||
| apm.yml | ||
| BUILD.bazel | ||
| Cargo.lock | ||
| Cargo.toml | ||
| deny.toml | ||
| lefthook.yml | ||
| MODULE.bazel | ||
| MODULE.bazel.lock | ||
| multitool.lock.json | ||
| README.md | ||
| rust-toolchain.toml | ||
| rustfmt.toml | ||
dark-factory
A lightweight, Kubernetes-native control plane for AI agents working as "employees" on software projects — an opinionated rethink of Paperclip for a homelab. A Rust control plane provisions agent-sandbox sandboxes, a Rust runtime inside each sandbox drives Claude Code (or other ACP harnesses) and talks back over mTLS, and integrations with forges and issue trackers run as WASI plugins.
Status: design — nothing runnable yet.
Repository layout
| Path | Content |
|---|---|
product/context.md |
Vision, requirements, decision index, future options |
product/iterations/ |
Iteration plans — scope, milestones, spikes |
product/agent-playbook.md |
How agents are expected to work; source for built-in and project skills |
design/tech-stack.md |
The whole tech stack in one place |
design/adr/ |
Architecture Decision Records |
design/ |
Technical deep dives (agent-run.md, egress.md, deployment.md) |
contract/ |
Wire and plugin contracts: protobuf (agent protocol, transcript), WIT (integration plugins), factory tool schemas |
openspec/ |
Specs and implementation changes; order of work per iteration |
Planned as the code arrives: a Cargo workspace (crates/), WASI plugins (plugins/), the Helm chart (deploy/helm/dark-factory/, ADR-0031) and Bazel build files (MODULE.bazel), see ADR-0024.