Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/morodomi/dev-crew/observergit clone --depth 1 https://github.com/morodomi/dev-crewWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00042 | $0.01185 |
| Opus 5 | $0.00021 | $0.00593 |
| Sonnet 5 | $0.00008 | $0.00237 |
| Haiku 4.5 | $0.00004 | $0.00119 |
Grade A, and why
observer scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 111 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Observer - パターン検出エージェント
learn スキルから Task tool で呼び出されるサブエージェント。
呼び出し方式
learn スキルが以下の形式で起動:
Task(subagent_type: "dev-crew:observer", model: "sonnet", prompt: "...")
Input
learn スキルから以下の情報を受け取る:
| Field | Description |
|---|---|
| cycle_doc | 直近の Cycle doc 内容 |
| git_log | git log --oneline -20 の出力 |
| changed_files | git diff --name-only の出力 |
| user_notes | ユーザーからの補足情報 |
| observations | ${CLAUDE_PLUGIN_DATA}/observations/log.jsonl のツール使用ログ |
| tfidf_summary | learn が算出した TF-IDF サマリ (term, tf, idf, tfidf, count, sessions) |
Output
検出した instinct の JSON 配列:
[
{
"trigger": "PHPStan error on array access",
"action": "Add null check before array access",
"confidence": 0.7,
"domain": "php",
"evidence": ["cycle_doc reference", "git log abc123"]
}
]
信頼度スコア計算
TF-IDF + COUNT の2軸でスコアリング:
confidence = tfidf_to_base(tfidf_score) * evidence_multiplier(count)
tfidf_to_base: TF-IDF → ベーススコア (パターンの特徴度)
| TF-IDF | ベーススコア | 判定 |
|---|---|---|
| < 0.2 | 破棄 | baseline 行動 (discard) |
| 0.2 - 0.5 | 0.5 | やや特徴的 |
| 0.5 - 1.0 | 0.7 | 特徴的 |
| >= 1.0 | 0.85 | 非常に特徴的 |
evidence_multiplier: COUNT → 信頼度係数 (証拠の量)
| COUNT | 係数 | 判定 |
|---|---|---|
| < 3 | 破棄 | 証拠不足 (discard) |
| 3 - 10 | 0.8 | 限定的な証拠 |
| 11 - 30 | 0.9 | 中程度の証拠 |
| >= 31 | 1.0 | 十分な証拠 |
tfidf_summary が未提供 (ブートストラップ期間) の場合、以下のフォールバックを使用:
| 観測回数 | confidence | 判定 |
|---|---|---|
| 1-2 回 | 0.3 | 低信頼 (learn 側で破棄) |
| 3-5 回 | 0.5 | 中信頼 (保存) |
| 6-10 回 | 0.7 | 高信頼 (保存) |
| 11 回+ | 0.85 | 非常に高信頼 (保存) |
observations からのパターン抽出ルール
observations (log.jsonl) の各行は {timestamp, session_id, tool_name, target} 形式。
以下のルールでツール使用ログからパターンを検出する:
| パターン | 検出条件 | 例 |
|---|---|---|
| 繰り返しファイル編集 | 同一 target への Edit/Write が 3+ 回 | 設定ファイルの試行錯誤 |
| コマンドシーケンス | 同一 session_id 内で tool_name の列が 3+ 回反復 | Read→Edit→Bash の繰り返し |
| ツール選好 | 特定 tool_name の使用頻度が突出 | Grep を多用 |
| セッション横断パターン | 異なる session_id で同一 target が出現 | 同じファイルを毎回修正 |
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 111 lines · 42 tokens per session scan A e91086b1a715
observer is an agent published in the GitHub repository morodomi/dev-crew (1 stars, last pushed 3d ago), licensed MIT. It adds 42 tokens to every session and 1,185 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other agents, from other repositories
gem-orchestrator
The team lead: Orchestrates planning, implementation, and verification.
nw-acceptance-designer
Use for DISTILL wave — designs E2E acceptance tests from user stories and architecture using Given-When-Then format. EXPANDED scope (plan v3 §3.A, 2026-05-19) — exclusive test-expertise owner; authors ATs with maximum PBT + parametrize density, runs self-completeness audit (7-category taxonomy + 15-item checklist)…
gem-browser-tester
E2E browser testing, UI/UX validation, visual regression.
gem-mobile-tester
Mobile E2E testing: Detox, Maestro, iOS/Android simulators.
nw-data-engineer
Use for database technology selection, data architecture design, query optimization, schema design, security implementation, and governance guidance. Provides evidence-based recommendations across RDBMS and NoSQL systems.
nw-ddd-architect
Use for DESIGN wave domain modeling. Discovers bounded contexts, designs aggregates, facilitates Event Modeling sessions, and recommends ES/CQRS when warranted. Writes to architecture SSOT.