Chronos: time awareness for agents

You are temporally blind by default. You cannot feel elapsed time, you cannot tell a call you made ten seconds ago from one you made an hour ago, and you will confidently say "just now" about something that happened forty minutes back.

Chronos fixes that by reading the transcript the host already writes. Every tool call in this session is on disk with a start timestamp, an end timestamp, the arguments, and whether it failed. Stop guessing. Go look.

What you are given

In your context, without asking:

• At session start: the real wall-clock time, how long you were away if this is a resumed session, and how much active time has already gone into today.
• On every turn: chronos: t+<active time>, turn <n>, <n> tool calls, plus a LOOP: line when you have repeated an identical call, and a STALE: line when a file changed after you last read it.

Those lines are facts measured from disk. Prefer them over your own impression.

On demand:

Active time excludes gaps longer than five minutes. A session resumed after two days has not been running for two days, and reporting that it has is the kind of wrong number that makes a person stop trusting the tool.

When to look at the clock

1. Before retrying anything that just failed

Run chronos when <the command>, or read the LOOP: line if one was injected.

• Same call, same arguments, failed under 60 seconds ago: change the approach. Re-running an identical call cannot produce a different result unless something external changed, and nothing external changed in 60 seconds.
• 60 seconds to 10 minutes: one retry is defensible if you say what you expect to differ this time.
• Over 10 minutes: circumstances may genuinely have moved. Re-run is reasonable.

The same rule governs deploys, pushes and CI triggers. A pipeline that failed 40 seconds ago will fail again for the same reason. Read the log first.

2. Before trusting anything you already looked at

A STALE: line in your context already did this check for one case: a file that was written after you last read it. The other rows in this table are yours to apply, and chronos will not remind you about them.

3. Before summarising or reporting progress

Let the elapsed time set the length:

• under 5 minutes: answer, no recap;
• 5 to 30 minutes: one paragraph on what changed;
• over 30 minutes: chronos report, and use its real numbers.

4. Before starting something long or destructive

Check whether the same tool has been getting slower. chronos report lists the slowest calls and the average per tool. If the last few runs of a command are each well above its average, something is degrading. Say so rather than waiting quietly.

5. Idle detection in autonomous loops

When the injected context says the user has been away and you are still looping unattended: stop, summarise, list the open decisions, and ask. Chronos tells you about an absence once it passes 30 minutes; the statusline shows one from 1 minute. Always escalate past two hours of session time with no user input, whatever either threshold says.

6. Never say "just now" or "in a moment"

Replace every vague interval with a measured one.

• Wrong: "I just tried that."
• Right: "That exact command failed 47 seconds ago. Repeating it will not help."
• Wrong: "Let's wait a bit and retry."
• Right: "Retry at 14:32 local, three minutes from now."

7. Date and time questions

Never answer from memory or from your training data. In order:

1. The now line injected at session start, if the session is young.
2. chronos report, whose numbers come from the transcript's own timestamps.
3. date -u +"%Y-%m-%dT%H:%M:%SZ", or Get-Date -AsUTC -Format o on PowerShell 7, or Get-Date -Format o on Windows PowerShell 5.1.

Failure modes to avoid

• Do not count turns or messages to estimate elapsed time. Message count and wall clock are unrelated; one long tool call can span more time than fifty messages.
• Do not trust your own earlier "just now". The transcript is the record; your context is a summary of it.
• Do not treat a resumed session's span as time worked. Use active time.
• Do not go silent in an autonomous loop. Rule 5 always applies.
• Do not assume chronos is installed. If you never saw a chronos: line in your context, it is not running: fall back to the shell commands below and say so.

Degraded mode: no hooks on this platform

Chronos ships hooks for Claude Code. On a host without them you still follow every rule above, you just gather the facts yourself:

State it plainly when you do: "No chronos context in this session; this timestamp comes from date -u."

Why this is a skill and not a note in a prompt

From arxiv 2510.23853, "Your LLM Agents are Temporally Blind" (October 2025): even with timestamps present in context, the strongest frontier models reach only 65% alignment with human temporal judgment, and timestamps appear in under 4% of reasoning traces. Putting a clock in the context is not enough. The model has to be told when to look at it. The seven rules above are that instruction.