SuperAPI AI Directory

vibe-soc-loop

Coding & development

View repository

Updated 10-16 09:29

About this skill

vibe_soc loop

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

  1. Read repository AGENTS.md.
  2. 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
  3. Resolve the absolute project root, workspace, state module name, RTL top, and testbench top separately. For chip/top, the state module is vibe_soc_top while the current RTL top is chip_earlgrey_asic; never infer one from the other.
  4. Query existing pipeline_state.json before stage work. Initialize only when absent and only through the validated state helper.
  5. For review requests, select review_mode=quick|normal|strict; default to normal, use quick for dry-run planning, and use strict before 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:

  1. Mark the owned stage in_progress before editing stage artifacts.
  2. Keep artifacts under the approved roots: docs/, de/rtl/, de/run/, de/syn/, dv/tb/, and dv/sim/.
  3. Run required checks through registered MCP tools. Treat MCP unavailability as a stage blocker or failure, not permission to bypass the flow.
  4. On failure, read the real log/report, classify the failure, update state when applicable, and loop back to the earliest affected stage.
  5. Close a stage only when artifacts exist, all recorded checks pass, and pipeline_state.json records the real evidence.
  6. If a downstream stage repairs RTL, apply the invalidation rules from 01_swarm_flow.md and 05_pipeline_state.md before declaring closure.
  7. For post-stage review or commit readiness, dispatch soc-reviewer after the stage owner finishes. Treat review findings as follow-up work; do not add a review stage to pipeline_state.json.
  8. 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.
  9. 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 from 13_review_gate.md. Only soc/review/rule_library/ may supply project rules; other knowledge-base sources are reference evidence, and empty placeholders require Need 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>