Skip to content
sdlc-foundry

Agents and orchestration

Two layers:

  1. cplat, a tested CLI grouped by noun (system, component, api, resource, instance, issue, platform) plus plumbing (feedback, pr, budget, validate, and more). Every command has --dry-run and a clear report (what it will do, what it did, how to undo).
  2. Slash commands and agents, thin prompts that run the script, show you the preview, and relay the result. Orchestration commands (/component:spec, /component:plan, /component:build, /component:fix, /component:release, and the unattended /component:ship and /system:ship) spawn specialised agents (product manager, architect, coder, tester, security, deployment, release) in phases, routed by shape.

Runs are lean by default: a headless session starts its coordinator on Sonnet (--model or coordinator_model in .platform-settings.yml or ~/.config/sdlc-foundry/settings.yml changes it; the architect and reviewer keep their own model), cplat pr open commits, pushes and opens the pull request in one step, cplat quiet <script> prints one line for a passing check and a compact failure summary (not a raw log tail) for a failing one, and agents get a brief with their criteria and task instead of rereading the whole spec. In the platform’s benchmark a small change dropped from about $0.48 to $0.15 and a full build from $1.94 to $1.26 (estimated list prices). Run records feed a weekly retro (cplat platform retro) that proposes the next specs from evidence.

Runs are also bounded. One budget covers a whole run tree: 180 minutes, 60 invocations, 6 reruns and $15 per ship (--budget changes it). A run that fails twice in an equivalent way stops instead of trying a third time, and leaves a recovery snapshot; cplat budget shows what is left. In fix rounds cplat validate runs only the checks a change can affect, while the final gate and the pull request CI stay full. With CPLAT_OTLP_ENDPOINT or telemetry: set, a run can export usage metrics (never prompts or tool content) to OpenTelemetry, local endpoints only unless allow_remote is set, and the lgtm stack ships a Claude Code usage dashboard. Details: Unattended delivery.

Rules every agent follows: use devbox run <recipe> and never call tools directly; ask before pushes, pull requests, repository creation and anything that touches a cluster; stop and offer /issue:report when a platform template or script misbehaves. The full model, including phases and context hand-offs, is in the agent model.