Overview
Use this skill as the high-level dispatcher for vibe_soc. It selects the correct repo rules, skills, MCP tools, and state transitions without duplicating the lower-level skill contracts.
Preflight
- Read repository
AGENTS.md. - Read required rules before planning:
- always for pipeline work:
../../rules/01_swarm_flow.md,../../rules/02_toolchain.md,../../rules/05_pipeline_state.md - material RTL change:
../../rules/10_rtl_change_gate.md - verification failure/recovery:
../../rules/11_verif_recovery_gate.md - synthesis or PD:
../../rules/12_syn_pd_gate.md - review or commit readiness:
../../rules/13_review_gate.md - manual RTL style:
../../rules/04_coding_style.md - design decisions:
../../rules/06_design_knowledge.md
- always for pipeline work:
- Resolve the absolute project root, workspace, state module name, RTL top, and testbench top separately. For
chip/top, the state module isvibe_soc_topwhile the current RTL top ischip_earlgrey_asic; never infer one from the other. - Query existing
pipeline_state.jsonbefore stage work. Initialize only when absent and only through the validated state helper. - For review requests, select
review_mode=quick|normal|strict; default tonormal, usequickfor dry-run planning, and usestrictbefore commit or PR.
Classification
Classify ownership before selecting an executor. Any generated top, wrapper, register RTL, CRG RTL, RTL/filelist, interface, or constraint change is pipeline-governed and must be owned by soc-pipeline plus the applicable stage role. Lower-level skills and MCP tools execute work for that owner; they do not bypass state gates.
| Task class | Owner | Executor |
|---|---|---|
| architecture, material RTL, generated RTL/top/wrapper, multi-stage recovery | soc-pipeline / stage role |
matching generator, soc-integrate, or soc-build |
| standalone lint/compile/simulation/regression/coverage/synthesis request | applicable stage role | soc-build |
| read-only port extraction, snapshot, or interface diff | coordinator | soc-integrate |
| OpenROAD physical-design handoff | soc-pd-engineer |
soc-openroad |
| loop audit, validation evidence, commit readiness | soc-reviewer |
check_loop_state.py and read-only inspection |
| approved source-table conversion with no RTL output | coordinator | excel-yml-gen, crg-req-to-design, or cr-tree-diag-gen |
If a required owner role or executor is missing, stop with a precise blocker. Do not hand-roll generated tops, generated CRG logic, direct simulator runs, direct synthesis runs, or OpenROAD shell fallbacks.
Loop Contract
For pipeline-governed work:
- Mark the owned stage
in_progressbefore editing stage artifacts. - Keep artifacts under the approved roots:
docs/,de/rtl/,de/run/,de/syn/,dv/tb/, anddv/sim/. - Run required checks through registered MCP tools. Treat MCP unavailability as a stage blocker or failure, not permission to bypass the flow.
- On failure, read the real log/report, classify the failure, update state when applicable, and loop back to the earliest affected stage.
- Close a stage only when artifacts exist, all recorded checks pass, and
pipeline_state.jsonrecords the real evidence. - If a downstream stage repairs RTL, apply the invalidation rules from
01_swarm_flow.mdand05_pipeline_state.mdbefore declaring closure. - For post-stage review or commit readiness, dispatch
soc-reviewerafter the stage owner finishes. Treat review findings as follow-up work; do not add areviewstage topipeline_state.json. - When a module workspace is available, run
.agents/scripts/check_loop_state.py <workspace> --mode <review_mode>as read-only evidence for the reviewer or final response. - Reviewer dispatches include the implicated review domains, available reports/deliverables,
knowledge_scope, and requested top-risk count. Require the structured report sections and source-authority gate from13_review_gate.md. Onlysoc/review/rule_library/may supply project rules; other knowledge-base sources are reference evidence, and empty placeholders requireNeed Human Confirmation.
Reporting
Final responses for loop work must include:
- task classification and selected lower-level skill/tool
- files changed
- state transition or reason no state transition was needed
- checks run with pass/fail evidence, or why checks were not run
- reviewer outcome when review was run or why review was not needed
- next blocked or recommended stage
Evidence Summary
For completed loop work, include a compact evidence block:
loop_evidence:
module: <workspace or module>
task_class: <classification>
review_mode: <quick|normal|strict|not_run>
state: <stage/status summary>
tools: [<registered MCP tools or checker scripts>]
artifacts: [<relative evidence paths>]
result: <pass|needs-fix|needs-validation|blocked|not_run>