Chronos

Chronos — Time Awareness for Agents

You are not temporally blind. Stop guessing "when" and "how long ago". You have a ledger. Use it.

What chronos gives you

When the host platform supports hooks (Claude Code, Codex, OpenCode-plugin), chronos populates three sources of truth:

1. SessionStart baseline — injected into your context at session start:

   • now_utc — ISO-8601 UTC
   • now_local — ISO-8601 with offset
   • tz — timezone name + offset
   • ledger — absolute path to this session's JSONL tool-use ledger
   • state — absolute path to this session's state file
   • session — session ID

2. Per-turn delta — injected on every user message:

   • now, turn, session_duration, since_last_user

3. Tool-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".

When to consult time — 7 triggers

1. Before retrying a failed command

Read the ledger for the same tool + args_hash.

• If the last failure was < 60s ago with the same args, change strategy — don't re-run identically.
• If 60s – 10min, consider one retry with a rationale.
• If > 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.

2. Before trusting memory, a cached finding, or a prior read

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

3. Before reporting progress or writing a summary

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.

4. Before running a long or destructive command

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.

5. Idle-loop detection (autonomous mode)

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.

6. Before any "wait" / "in a few minutes" / "just now" statement

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)."

7. Date or time questions

Never guess. In order of preference:

1. Read now_utc / now_local from the SessionStart context if session_duration < 10 min.
2. Read latest finished_at in the ledger if < 10 min ago.
3. Run date -u +"%Y-%m-%dT%H:%M:%SZ" fresh.

How to read the ledger

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.

Failure modes to avoid

• Do not re-derive elapsed time by counting messages or turns. Use wall clock from ledger.
• Do not trust your own prior claim of "just now" from earlier in context. Ledger is the source of truth.
• Do not issue "try again in a few minutes" without a concrete ISO timestamp or delta.
• Do not silently hang in autonomous mode. Trigger #5 always applies.
• Do not assume the ledger exists. Check for ledger: in your context first; degrade to date -u if absent.

Degraded mode (no hooks platform)

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:

• Current time: date -u +"%Y-%m-%dT%H:%M:%SZ" (bash) or Get-Date -AsUTC -Format o (pwsh)
• File mtime: stat -c %Y <path> (bash) or (Get-Item <path>).LastWriteTimeUtc (pwsh)
• Git age: git log -1 --format=%cI <path> for last-commit ISO timestamp

State this explicitly: "No ledger on this platform; I'm using date -u as fallback."

Install / wiring

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.sh
• cursor/ — rules-only degraded mode
• hermes/, adal/, pi/, openclaw/ — SKILL-only ports

Why this matters

From 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.