菜单
Use when delegating independent code review, design review, or focused bug investigation to an AI reviewer. Provides delegation templates and records reviewer evidence; higher-level review flow remains in cosim-gpu-review or cosim-gpu-rlcr-loop.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use as the main entry for cosim debugging: gem5 crashes, assertions, hangs, timeouts, address translation faults, non-PASS test results, guest-side waits, and QEMU exits that may be secondary to gem5. Collect logs, preserve live state when needed, route to gem5 model references, and decide the first failing component.
Use when building gem5 or QEMU, checking build provenance, or preserving gem5 incremental builds across worktrees. Delegates to scripts/cosim_build.sh; never hand-roll Docker, scons, or QEMU make commands.
Use when the guest disk image needs modification without a full rebuild. Mount with guestmount, add files, systemd services, kernel module config.
Use as the first step for non-trivial cosim testing, debugging, review, or iterative implementation tasks. Do not use for ordinary repository maintenance such as commit splitting, submodule pointer updates, author config, or documentation-only cleanup.
Use when you need to send commands to a running cosim guest. Console pipe interaction, 9p share mount, GPU status check, manual driver load.
Use when collecting, summarizing, auditing, or comparing cosim evidence from artifact directories, large logs, matrices, verdict files, build provenance, or prior conversation records; also use when reducing raw log reads or checking whether existing artifacts lose information. Enforces full raw evidence preservation, bounded reads, source attribution, compact tables, raw-read plans, coverage self-reporting, and unknown-event preservation before analysis.
| name | cosim-gpu-codex-review |
| description | Use when delegating independent code review, design review, or focused bug investigation to an AI reviewer. Provides delegation templates and records reviewer evidence; higher-level review flow remains in cosim-gpu-review or cosim-gpu-rlcr-loop. |
Use an independent AI reviewer for bounded technical review, investigation, or
design critique. This skill owns delegation mechanics only. Review policy shared
with cosim-gpu-review and cosim-gpu-rlcr-loop lives in
../cosim-gpu-review/references/review/independent-review-contract.md; read it
before adversarial review delegation.
Use this skill for:
Do not delegate user interaction, task scoping, build commands, test commands, or final verification. The main agent owns integration and evidence acceptance.
Each assignment must be self-contained:
# Target
Exact files, symbols, base reference, current HEAD, and non-goals.
# Change Or Review Question
For review: dimensions to inspect and exact diff commands.
For investigation: symptom, search scope, and questions.
For design: current code, constraints, and rejected directions if artifact-backed.
# Acceptance
Expected output format, severity labels, and required evidence.
No project-wide build, test, or lint commands.
Implementation-oriented assignments must require current HEAD as the edit
base, diff-scoped formatting only, and no whole-file formatter churn unless the
task is explicitly formatting-only.
Prepare the review brief through cosim-gpu-review or cosim-gpu-rlcr-loop.
The sub-agent prompt points to that brief and may include exact scoped diff
commands. It must not include the implementer's theory, desired result, or
dialogue summary.
For cosim-specific prompts, use
references/cosim-gpu-review-prompts.md when the touched subsystem matches
interrupt handling, PM4, cache or TLB behavior, HSA completion signals, or
similar GPU model paths.
Save every consultation under the active task workspace:
artifacts/<task-slug>/reviews/codex-<round>-<file>.md
artifacts/<task-slug>/reviews/codex-<round>-summary.md
Record prompt, file scope, response, classification, and disposition. If the reviewer times out or errors, retry once with a narrower scope and record the failure if it still fails.
Default to Codex for local code review and Gemini for research that needs current external documentation. If the provider is unavailable and setup cost exceeds review value, fall back to manual review and record that choice.
Proxy and sandbox issues are operational details, not review findings. When a sub-agent creates stray files, verify with repository status and remove only files proven to be sub-agent artifacts.