メニュー
Record an EM Supervisor observation (trouble sign, incident, or neutral note) for the current session.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/nirecom/agents --skill supervisor-reportコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Edit source code for the current task. Delegates editing and lint/typecheck/self-repair to a subagent.
Plan and write test cases with high reasoning effort. Test iteration runs in a subagent to minimize confirmations.
Conduct a decision-tree interview with the user to lock in requirements, motivation, scope, and non-goals before planning.
SOC 職業分類に基づく
SKILL.md を表示中
| name | supervisor-report |
| description | Record an EM Supervisor observation (trouble sign, incident, or neutral note) for the current session. |
| user-invocable | false |
Invoked by Claude Code itself when a trigger in rules/supervisor-reporting.md fires — never by a human, which is why user-invocable is false.
SR-1. Resolve $SID: read the Session-ID: field of the worktree's WORKTREE_NOTES.md. Only if that file or field is absent, fall back to $CLAUDE_SESSION_ID — Anthropic bug #27987 makes its propagation into Bash unreliable, so it is the fallback, not the source. If neither resolves, still report: state in your turn output what you observed and that no session id could be resolved.
SR-2. Choose categories (comma-separated, multi-select), severity, detail (what was observed, free text), and reporter (the skill or agent name) from the tables below.
SR-3. Run, as one Bash call: node "$AGENTS_CONFIG_DIR/bin/supervisor-report" --categories <cats> --severity <sev> --detail "<text>" --reporter "<name>" --session-id "$SID". All four flags are mandatory — the CLI aborts when one is missing.
<text> is observation text that may come from tool output or a file, so treat it as untrusted: before substituting it, strip every `, $, \, ", newline and control character from it, and collapse the remainder to a single line — a $(...) or backtick left in the detail executes inside the double-quoted argument.
Text you cannot safely reduce that way must not be interpolated at all: shorten the detail to your own one-line summary and leave the raw text out.
SR-4. Never swallow a non-zero exit. Put the fact that the report failed, plus the CLI's stderr, in your turn output: a lost observation is the worst failure mode of this skill.
| Category | When to use |
|---|---|
intent | Scope or non-goal misalignment with intent.md |
outline | Approach selection or delivery plan issue |
detail | File-level implementation plan inconsistency |
workflow | Workflow rule violation, step skip, sentinel issue |
code | Code writing issue (logic error, naming, structure) |
test | Test failure, flaky test, coverage gap |
security | Credential leak risk, dangerous input handling |
performance | Build/runtime slowdown, resource spike |
env | Missing env var, dependency version mismatch |
other | Does not fit any category above |
| Severity | When to use |
|---|---|
error | Confirmed failure or clear violation |
warning | Suspicious behaviour, likely issue |
notice | Worth recording, not immediately concerning |
Example: --categories intent,code --severity warning --detail "changes touch declared non-goal: async LLM calls" --reporter "write-code"
