Coordinate work; do not implement RTL or testbench content in the coordinator.
Preflight
- Read
../../rules/01_swarm_flow.md,../../rules/02_toolchain.md,../../rules/05_pipeline_state.md, and../../rules/13_review_gate.mdwhen review or commit readiness is in scope. Read coding style or exceptions only when relevant. For any architecture, documentation, RTL, verification, synthesis, physical-design, or material refactoring task, also read../../rules/06_design_knowledge.mdand querysoc-ai-kbbefore design decisions. - Resolve the absolute module workspace and module name.
- Query
pipeline_state.json; initialize it only when absent. Never overwrite existing state implicitly. - For chip-level, subsystem-level, or multi-module requirements without an approved architecture handoff, dispatch
soc-architectfirst for IP selection, technology/process selection, and the overall SoC integration architecture plan. Treat it as pre-doc planning, not apipeline_state.jsonstage. - Select the RTL role:
- normal logic:
soc-rtl-designer - top integration:
soc-integrator - CRG:
soc-crg-engineer, only whencrg-genis registered - OpenROAD physical-design handoff:
soc-pd-engineer
- normal logic:
- Select
soc-reviewerfor post-stage, pre-commit, or validation-evidence audit. The reviewer reports findings only and never closes a pipeline stage. Use.agents/scripts/check_loop_state.py <workspace> --mode normalafter stage work and--mode strictbefore commit readiness when a workspace is available. - Reviewer dispatch includes relevant review domains, reports and delivery inputs, knowledge-base scope, and the structured output contract from
13_review_gate.md. Missing knowledge-base rules requireNeed Human Confirmation; they are never synthesized from memory.
Delegation
Use named role agents when supported. Otherwise spawn a generic subagent with the matching file under ../../agents/ as its role contract. Respect host delegation policy and do not silently replace a required role agent with coordinator-authored RTL.
For large stage tasks, the coordinator may split work across multiple agents automatically when the host allows it. Keep one role agent as the stage owner. Sidecar agents must have explicit, disjoint write ownership and may not independently close the stage. The stage owner integrates sidecar work, runs the required registered MCP checks, validates artifacts, and updates pipeline_state.json.
Every dispatch prompt includes:
- absolute
workspace task_name- objective and approved assumptions
- single- or multi-module state mode
- for pipeline-stage agents only, requirement to update state and quote the
update_state.pystdout line
After each pipeline-stage role returns, query state immediately and verify status, artifacts, and all checks. After soc-architect returns, verify that architecture artifacts exist and are ready for the doc stage. When review is requested or commit readiness matters, dispatch soc-reviewer after the stage owner reports and before claiming the loop is closed. Do not dispatch downstream work after failure.
Execution contract
- Canonical artifact paths:
docs/,de/rtl/,de/run/,de/syn/,dv/tb/,dv/sim/. - Architecture handoff artifacts stay under
docs/and must cover IP selection, technology/process selection, and the integration architecture before the doc stage consumes them. - EDA execution uses registered MCP tools. Verification calls
soc-build.soc_sim; synthesis callssoc-build.soc_syn; physical-design handoff callssoc-openroad.soc_openroad_*. No direct EDA shell fallback. doc -> rtl; verification and synthesis may run independently after RTL passes.- Treat estimated timing and synthetic PASS markers as failures of validation integrity.
- For an approved doc-stage exception, record
doc skippedwith a concrete note before starting RTL.
Stop and report a precise blocker if required MCP capability, an approved interface decision, or an EDA dependency is unavailable.