菜单
Gives AI coding agents temporal awareness via a hook-backed ledger and decision rules. Activates whenever an agent reasons about recency, retry windows, memory staleness, deploy cooldowns, idle detection, or any "when / how long ago" question. Required for autonomous or long-running agents.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/OthmanAdi/chronos --skill chronos命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | chronos |
| description | Gives AI coding agents temporal awareness via a hook-backed ledger and decision rules. Activates whenever an agent reasons about recency, retry windows, memory staleness, deploy cooldowns, idle detection, or any "when / how long ago" question. Required for autonomous or long-running agents. |
| version | 1.0.0 |
| license | MIT |
| author | OthmanAdi |
| tags | ["time","temporal","autonomy","ledger","agents","hooks","session-management"] |
| homepage | https://aware-bloom-szmt.here.now |
| repository | https://github.com/OthmanAdi/chronos |
| allowed-tools | Read, Bash(date *), Bash(stat *), Bash(cat *), Bash(jq *), Bash(ls *), Bash(git log *), Bash(git status *) |
You are not temporally blind. Stop guessing "when" and "how long ago". You have a ledger. Use it.
When the host platform supports hooks (Claude Code, Codex, OpenCode-plugin), chronos populates three sources of truth:
SessionStart baseline — injected into your context at session start:
now_utc — ISO-8601 UTCnow_local — ISO-8601 with offsettz — timezone name + offsetledger — absolute path to this session's JSONL tool-use ledgerstate — absolute path to this session's state filesession — session IDPer-turn delta — injected on every user message:
now, turn, session_duration, since_last_userTool-use ledger — JSONL file, one entry per tool invocation:
{tool_use_id, tool, args_hash, started_at, finished_at, duration_ms, success}If the host platform does not support hooks (Cursor, Hermes, ADAL, plain skill-only mode), chronos degrades: run date -u or date -Iseconds at decision points and state explicitly "derived from shell, not ledger".
Read the ledger for the same tool + args_hash.
< 60s ago with the same args, change strategy — don't re-run identically.60s – 10min, consider one retry with a rationale.> 10min, circumstances may have changed; re-run is reasonable.Deploy cooldowns: the same rule applies to deployments and pushes. If the last deploy failed under 60s ago, check logs and fix config before re-deploying. Don't re-trigger a broken pipeline without a changed state.
Staleness thresholds by source:
| Source |
|---|
| Stale after |
|---|
| Action when stale |
|---|
git status, running processes, env vars | 5 min (same session) | Re-query |
| File contents previously Read | 1 hr OR any write to that path since Read | Re-Read |
| Persistent memory (MemPalace, KG facts) | 7 days since last write | Re-verify before use |
| External API responses (prices, PR status, issues) | 5 min | Re-fetch |
| Build/test results | After any source change since that run | Re-run |
| LLM's own prior claim ("earlier I saw...") | Always verify against ledger | Consult ledger |
Use session_duration to pick verbosity:
< 5 min — terse answer, no recap.5 – 30 min — one-paragraph summary of what changed.> 30 min — structured recap with timestamped milestones from the ledger.Scan the ledger for the same tool recently. If the last 3 duration_ms values for that tool are growing (e.g. each > 2× the previous median), something is degrading. Surface it, don't silently wait.
Whenever since_last_user exceeds the idle threshold (default 900s / 15 min, configurable via CHRONOS_IDLE_THRESHOLD_SEC) and you are still in an autonomous loop: pause. Summarize progress, list pending decisions, ask for direction. This prevents silent runaway.
For explicit autonomous loops (/loop, /autonomous), the threshold may be higher — but always escalate when session_duration > 2h without user input.
Replace imprecise language with a concrete delta from the ledger:
❌ "Just tried that."
✅ "Last attempt was 47s ago (tool_use_id abc123, failed with EACCES)."
❌ "Let's wait a moment."
✅ "Retry at 2026-04-23T20:45:00Z (+3 min from now)."
Never guess. In order of preference:
now_utc / now_local from the SessionStart context if session_duration < 10 min.finished_at in the ledger if < 10 min ago.date -u +"%Y-%m-%dT%H:%M:%SZ" fresh.Ledger path is injected as ledger: <path> in SessionStart context.
One line = one event (JSONL, append-only). Latest line for a given tool_use_id wins.
# Last 10 events
tail -n 10 "$LEDGER_PATH"
# Events in last 10 minutes (helper script)
./scripts/ledger_read.sh --since 10m
# Events for a specific tool in last hour
./scripts/ledger_read.sh --tool Bash --since 1h
# Last duration for a specific args_hash
./scripts/ledger_read.sh --tool Bash --args-hash abc123 --last
Entry shape:
{"tool_use_id":"abc","tool":"Bash","args_hash":"f0e1","started_at":"2026-04-23T20:00:00Z"}
{"tool_use_id":"abc","tool":"Bash","args_hash":"f0e1","finished_at":"2026-04-23T20:00:12Z","duration_ms":12034,"success":true}
Two lines per tool call: one on PreToolUse (started_at), one on PostToolUse (finished_at + duration). This is deliberate — async writes avoid adding latency.
args_hash is written by the PreToolUse hook: SHA-256 of the serialized tool arguments, first 12 hex characters. Filter the ledger by it to identify identical repeated calls. Do not compute it yourself — read it from the ledger only.
ledger: in your context first; degrade to date -u if absent.If you don't see chronos baseline in your context, the host platform doesn't support hooks. You still follow the decision rules — just swap ledger reads for:
date -u +"%Y-%m-%dT%H:%M:%SZ" (bash) or Get-Date -AsUTC -Format o (pwsh)stat -c %Y <path> (bash) or (Get-Item <path>).LastWriteTimeUtc (pwsh)git log -1 --format=%cI <path> for last-commit ISO timestampState this explicitly: "No ledger on this platform; I'm using date -u as fallback."
See installers/<platform>/ for platform-specific setup:
claude-code/ — full hook stack (5 events)codex/ — full hook stack (PreToolUse matches Bash only)opencode/ — TS plugin writes ledger; read via ledger_read.shcursor/ — rules-only degraded modehermes/, adal/, pi/, openclaw/ — SKILL-only portsFrom arxiv 2510.23853 "Your LLM Agents are Temporally Blind" (Oct 2025): even when timestamps are present in context, the best frontier models achieve only 65% alignment with human temporal judgment. Timestamps appear in fewer than 4% of reasoning traces. The fix is not more plumbing — it's explicit decision rules that make the agent look at the clock at the right moments.
This SKILL.md is those rules.