Content
# since
**Your agent already read the file. `since` tells it when the file changed.**
[](https://github.com/LNSHRIVAS/since/actions/workflows/ci.yml)
[]()
[](https://pypi.org/project/pysince/)
Coding agents act on files they read minutes ago. Those files change: a formatter runs, a teammate pushes, another agent edits, a git pull lands. The agent never sees it, and acts on the stale version with full confidence. `since` tells it, on every tool call, exactly what changed.
```bash
pip install pysince
```
Zero dependencies. Works in Claude Code, Cursor, Copilot, and Antigravity. Any MCP client.
## The problem is real, named, and everywhere
Agents read a file, reason about it for many steps, then act on it but the file changed underneath them, and nothing tells them. It has a name: the [stale world model problem](https://tianpan.co/blog/2026-04-10-stale-world-model-long-running-agents). On long-horizon coding tasks, frontier model success drops from around 70% to roughly 23%, and about 36% of those failures trace to context drift, not reasoning quality. The canonical shape: an agent reads a file at step 3, reasons about it through step 30, and writes it back at step 31, but another process edited it at step 17. The agent silently overwrites the newer version, and the task looks like it succeeded.
This shows up across every major agent tool:
- **OpenAI Codex** [will overwrite any changes it didn't create](https://community.openai.com/t/codex-will-overwrite-any-code-changes-it-did-not-create/1362873): edit files while it works and it restores them, "every time."
- **GitHub Copilot** agents [overwrite their own edits](https://github.com/orgs/community/discussions/163388) because "the editor's state is different from what the AI has in its session-based memory."
- **Claude Code** [subagents read stale file versions](https://github.com/anthropics/claude-code/issues/3032), and [its file cache diverges from disk](https://github.com/anthropics/claude-code/issues/51214) with Read/Grep returning the stale copy - the reporter concludes only an out-of-band disk check catches it.
- **Continue** has [the same class of bug](https://github.com/continuedev/continue/issues/9379).
- With **parallel agents**, this compounds into [silent file overwrites and stale views of the codebase](https://www.augmentcode.com/guides/git-worktrees-parallel-ai-agent-execution), where agents "proceed silently on corrupted data rather than surfacing exceptions."
It's a recognized production blocker - 32% of agent teams cite output consistency as their #1 issue and there are [whole guides](https://stormap.ai/post/how-to-stop-ai-coding-agents-from-overwriting-your-work-2026) written just on stopping agents from overwriting your work.
`since` is the lightweight, single-install out-of-band check: it fingerprints every file the agent reads and, on every tool call, reports which ones changed on disk before the agent acts.
## What it does
Every MCP tool response surfaces all files that changed since the agent last read them. The agent does not have to remember to check. `since` volunteers it:
```text
Files changed since last read:
config.json (content changed, mtime changed) - read 4m ago
alerts.py (content changed, mtime changed) - read 3m ago
```
The agent re-reads those files before acting, instead of writing from a stale copy.
## When you need this
Strongest when files change outside the agent's view: another process, a teammate on the same repo, a formatter, a pre-commit hook, a parallel agent, or context that drifted over a long session.
Less useful for quick single-file edits an agent already re-reads on its own. Skip it for throwaway scripts. Reach for it when agents share files, sessions run long, or more than one actor touches the tree.
## Setup
`since` runs as a local MCP server. Add it to your client's MCP config.
**VS Code:**
```json
{
"servers": {
"pysince": {
"type": "stdio",
"command": "pysince-mcp",
"args": [],
"cwd": "${workspaceFolder}"
}
}
}
```
**Antigravity:**
```json
{
"mcpServers": {
"pysince": {
"command": "python",
"args": ["-m", "since.mcp"]
}
}
}
```
Then add this line to your agent's system instructions so it knows when to call the tools:
> On the first read of any file, call `stamp_file_read`. Before editing a file, call `check_staleness`. When the response lists changed files, re-read them before acting on their contents.
## Tools
| Tool | When to call it | What it does |
|---|---|---|
| `stamp_file_read` | After reading any file | Records mtime and content hash |
| `check_staleness` | Before editing a file | Reports if it changed, and lists every other tracked file that changed too |
| `session_duration` | Anytime | How long the session has been tracked |
| `invalidate_source` | Manually | Marks a source stale on demand |
`check_staleness` is the core. It never answers only about the one file you asked about. It reports the full set of tracked files that have drifted, so the agent cannot stay blind to a change it did not think to check.
## How it works
`since` stamps every file read with its mtime and a SHA-256 hash. On any later call it compares stored fingerprints against the current file: mtime first because it is fast, full hash only if mtime moved. No daemon, no polling, no background process. Just a comparison against disk at the next turn, which is why it catches changes the agent's own cached view cannot.
---
## Also: temporal context for chat apps
The same primitive, aimed at conversations instead of files. Wrap your chat function and the model sees a timeline: when each message happened, how long the gaps were, and what context has gone stale.
```python
from since import Store, since_time
from openai import OpenAI
store = Store("~/.since/chat.db")
client = OpenAI()
@since_time(store=store, timezone="Asia/Kolkata")
def chat(messages):
return client.chat.completions.create(model="gpt-4o-mini", messages=messages)
resp = chat(messages=[{"role": "user", "content": "hello"}])
print(resp.choices[0].message.content)
```
The model receives a compact time block before each turn:
```yaml
Now: Wed Jul 01, 02:36 AM (night)
Session: 9h 2m total, 4m active across 3 sittings, 8 messages
Gap: 6h since the last message
```
So instead of "I don't have information about previous conversations," it can say "welcome back, it has been about 6 hours since we last spoke."
The decorator reads the OpenAI response shape by default. For other providers, pass an `extract_reply` function that returns the reply text from your provider's response object.
## Requirements
- Python 3.10+
- Zero dependencies
## Install
```bash
pip install pysince
```
The PyPI name is `pysince` because `since` was already taken. You import it, and the repo is named, `since`.
