菜单
Use when a command is killed by memrun, exits 137, exceeds an RSS limit, or produces a memrun JSONL log; inspect the log, identify the largest process, preserve diagnostics, and choose the next safe step before rerunning.
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
Turn a repo-local spec, rollout doc, design note, or implementation plan into a compact /goal prompt. Use when the user wants to run Codex /goal from an existing spec-first workflow and needs the goal to preserve scope, non-goals, validations, and local constraints.
Evaluate a repository's first-rung readiness for high-quality AI code contributions and produce concrete, copy-pasteable file changes that make the repo AI-ready. Use when asked to audit AI contribution quality, diagnose why AI PRs are low quality, prepare a repo for AI-assisted engineering, or move a repo from no harness toward a basic v0.5 contribution harness.
Bootstrap a small, deterministic repo-local AI harness by adding only the docs, skills, or checks that have concrete repo-specific content. Use when a repo already has basic AI contribution readiness and the user wants continuous harness improvement without placeholder process.
Evaluate a real AI-assisted work session against the repo-local harness, classify reusable misses, and update or propose updates to docs, skills, scripts, MCP/tooling, or sensors. Use after substantial work, review feedback, failed validation, or human correction in a repo with an AI harness.
Investigate established Pulumi-vs-Terraform parity gaps in bridged providers by creating or refining bridge cross-tests, after routing has already determined that bridge parity work is the next step, or when the user explicitly asks for that help. Use when the agent already knows that Pulumi has the problem and Terraform does not, and the next step is to capture, narrow, and explain that parity gap inside pulumi-terraform-bridge without debating alternate harnesses.
Run Claude Code as a chat-only reviewer for a local diff or branch, including adversarial pressure-testing when requested. Use when the user explicitly wants a Claude second opinion without posting GitHub comments.
| name | memory-guard |
| description | Use when a command is killed by memrun, exits 137, exceeds an RSS limit, or produces a memrun JSONL log; inspect the log, identify the largest process, preserve diagnostics, and choose the next safe step before rerunning. |
Use this after memrun stops a command, or when the user mentions a memrun
JSONL log, RSS limit, exit 137, or memory guard. The make proxy installed at
~/.local/bin/make guards commands in any harness. Logs default to
${XDG_STATE_HOME:-~/.local/state}/toolkit/memrun/.
Do not immediately rerun the failed command. First inspect the log and explain:
sample file was captured;Prefer the bundled summarizer at scripts/summarize_memrun_log.py relative to
this loaded SKILL.md:
python3 /absolute/path/to/memory-guard/scripts/summarize_memrun_log.py <memrun-log>.jsonl
When working from this toolkit checkout, the repo-relative source path is also valid:
python3 skills/memory-guard/scripts/summarize_memrun_log.py <memrun-log>.jsonl
137 from memrun as an intentional guard kill, not a normal test
failure.memrun enabled and lower scope first: a
narrower make target, one package, one test, or the direct generator command.MEMRUN_SAMPLE_AT and the same or
lower MEMRUN_RSS_LIMIT.The runtime-neutral variables are:
MEMRUN_RSS_LIMITMEMRUN_SAMPLE_ATMEMRUN_INTERVALMEMRUN_LOG_DIRMEMRUN_LABELMEMGUARD_DISABLELegacy CODEX_MEMRUN_* and CODEX_MEMGUARD_DISABLE aliases remain supported.
The proxy only guards make resolved through PATH; an explicit executable
such as /usr/bin/make bypasses it.
MEMRUN_RSS_LIMIT=8g MEMRUN_SAMPLE_AT=5g make generate
MEMGUARD_DISABLE=1 make -n
Use MEMGUARD_DISABLE=1 only for a harmless dry run or when the user explicitly
wants the guard bypassed.
Report the diagnosis first, then the safe next command. Keep the distinction clear between what the log proves and what remains an inference.