Engineering Blueprint
Claude Code + Manual code delivery workflows require... Playbook
Automates code delivery workflows using four specialized agents that handle issue intake, implementation, code review, and infrastructure changes with built-in handoff tracking and escalation.
1 File Included
pasted-prompt.md
5 KB
What problem does this solve?
Manual code delivery workflows require coordination between planning, implementation, review, and infrastructure changes, creating bottlenecks and requiring constant handoffs.
How does it work?
Four specialized agents handle a tracked issue through distinct phases: **Grillist (intake and plan locking):** Reads a new issue, isolates ambiguities, and asks one clarifying question per turn with a recommended answer. Loops with the user until silence signals approval. When the plan is concrete, restates it as acceptance criteria and assigns to the builder. **Architect or DevOps (code/infrastructure implementation):** Checks out the repo via isolated Git worktree, makes the change, runs tests, opens a PR, and assigns to Reviewer. If the plan is wrong, bounces back to Grillist—don't redesign mid-stream. **Reviewer (independent check):** Comes in cold, compares the diff to the locked plan, approves or requests changes. If back-and-forth exceeds three rounds, label the issue blocked and escalate to a human. **Getting started:** Once you have a running multica instance, connect your daemon. Create the four agents (Grillist, Architect, DevOps, Reviewer) through the multica web UI with initial prompts. Reference the grillist skill pattern from https://github.com/mattpocock/skills to structure your Grillist intake logic. For the Reviewer agent, dial in a code-review skill focused on comparing diffs against locked acceptance criteria and flagging deviations or quality issues. Establish cross-agent assignment behavior and handoff flows. **Typical workflow:** Grillist engages user, loops until plan locks, assigns to builder (Architect or DevOps), builder executes and assigns to Reviewer, Reviewer approves or requests changes, then returns to user for final sign-off on the PR.
What's the biggest win?
Enables parallel processing of multiple issues asynchronously while maintaining transparency and audit trails. Reduces ceremony and manual coordination overhead by using the issue itself as the integration point rather than requiring custom orchestration.
What should I know technically?
Separate the runtime harness from the agent prompt. The harness enforces hard guardrails and safety constraints (e.g., preventing destructive operations like deleting production databases). The agent prompt should focus on behavioral reasoning and context ('what you are doing and why'). This keeps critical safety logic out of the prompt layer where it can be misunderstood or overridden, and puts it in the runtime where it can be enforced.
What are the constraints?
Slower per individual task—a 15-minute manual change takes ~1 hour through the squad due to planning, build, and review phases. Parallelism advantage only pays off across batches of concurrent issues, not single quick fixes. Runtime constraints and guardrails must be enforced at the inference tool level itself—whether that's Claude Code's implicit safety behaviors or OpenCode running OhMyOpenAgent's prompts. The harness config specifies tool access and locks, but the actual safety enforcement depends on the inference runtime's own prompt layer. Changing platforms requires ensuring the new tool's behavioral prompts include equivalent guardrails.
Tools in this Blueprint
About This Blueprint
- Industry
- Computer Software