菜单
Append-only audit log for autonomous agent actions. Every meaningful action (peer call, PR open, file write, cron fire, secret read) must be logged so Rishi can review what agents have done without being present.
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
Peer-to-peer call to another Hermes agent over Fly's private network. Use to delegate work to a specialized agent (e.g., squad for coding) or to relay to the user via the gateway. Same skill installed on every agent.
Push a message into Rishi's Telegram chat from inside an API-server invocation. Use when a peer agent (squad, research, etc.) needs to surface a status update, PR-ready notification, or escalation to Rishi but the work was triggered via /v1/chat/completions (where responses don't otherwise land in Telegram). Fire-and-forget: this does not wait for Rishi's reply.
Orchestrate a coding task end-to-end on the Coding Squad: validate scope, plan, dispatch implementer + reviewer subagents, open a PR, handle review feedback, and escalate to the user via Telegram on failure. Enforces all guardrails (repo whitelist, Microsoft firewall, scope ceiling, no merge).
Search NYC apartment listings using NYBits (no-bot-detection). Works from headless servers where StreetEasy/Zillow/Apartments.com block with captchas. Supports filtering by neighborhood, bedrooms, price, fee status, and amenities. Use this skill whenever the user asks to find, search, or browse apartments in NYC.
Install and configure Chrome for headless browser automation on Fly.io. Fixes "Chrome not found" errors from agent-browser. Covers npm global install, Chrome download, AGENT_BROWSER_EXECUTABLE_PATH setup, and PATH configuration.
Generate project ideas via creative constraints.
| name | audit-log |
| description | Append-only audit log for autonomous agent actions. Every meaningful action (peer call, PR open, file write, cron fire, secret read) must be logged so Rishi can review what agents have done without being present. |
| version | 0.1.0 |
| author | Hakan |
| metadata | {"hermes":{"tags":["observability","logging","audit","foundation"],"related_skills":["call-agent"]}} |
Every Hermes agent on this system writes meaningful autonomous actions to a shared append-only log at /opt/data/logs/audit.log. This is how Rishi reviews what agents have done — without being present for each action.
The log is JSON Lines (one JSON object per line) so it can be grepped, filtered with jq, and turned into dashboards/cron alerts later.
Always log:
call-agent invocations)/opt/data/sessions/Don't log:
From any shell context on any agent:
bash /opt/data/skills/custom/audit-log/scripts/log_action.sh \
<action-category> "free-text description" [outcome]
Examples:
bash .../log_action.sh peer-call "Asked squad to fix PR Hakan#42 review comments"
bash .../log_action.sh pr-open "Opened rishijatia/Hakan#43 (add audit-log skill)" success
bash .../log_action.sh cron-fire "Daily WHOOP brief at 09:15 ET" success
bash .../log_action.sh escalate "Stuck on apartment-search auth flow, asked Rishi" pending
Action categories should be lowercase, hyphenated, and reused across agents.
# Last 20 entries (default)
bash /opt/data/skills/custom/audit-log/scripts/audit_query.sh
# Filter by agent or action
bash .../audit_query.sh --agent squad --action pr-open
bash .../audit_query.sh --since 2026-05-12
# Raw JSON for piping into jq
bash .../audit_query.sh --json | jq 'select(.outcome=="failure")'
{"ts":"2026-05-13T02:30:00Z","agent":"hermes-gateway","action":"peer-call","description":"Asked squad to identify itself","outcome":"success"}
Fields:
ts — ISO-8601 UTC timestamp, second precisionagent — Fly app name (e.g., hermes-gateway, hermes-coding-squad)action — short category, hyphenateddescription — free-text human-readable explanationoutcome — optional: success | failure | partial | pendingagent field disambiguates who did what. Future: ship to a central audit collector for cross-app querying.