Keep folders out of Probe

A .probe.config file sets how much Probe may touch the paths under its folder. Put one in any folder: it covers that folder and everything below it, in the same syntax as a .gitignore. It works whichever folder your coding agent was started in.

~/code/client-x/.probe.config

# nothing in this repo is read or recorded
*

Levels

Every line gives the paths it matches a level:

Level	Written as	What Probe does
`off`	a plain line, or `<pattern> @probe off`	Never reads the path and records nothing about it.
`read`	`<pattern> @probe read`	The Probe daemon may see your agent’s work on it, as context. Nothing about it is recorded.
`hash`	`<pattern> @probe hash`	Runs over it are recorded, but each file only by its identity: path, SHA-256 (a fingerprint over 1 GiB), size and row count. Never its bytes, a preview, an upload or its text. The daemon treats it as `off`.
`on`	`!pattern`, or `<pattern> @probe on`	Recorded as usual.

@probe <level> on a line of its own sets the level of every line after it, so written first it is the file’s default:

.probe.config

@probe read
notebooks/
experiments/
notebooks/private/ @probe off
notebooks/README.md @probe on

read and off both stop recording: no run is opened there, and no artifact, note or daemon record is made about those paths. Only the daemon tells them apart: with read it still sees the work, so it understands the session; with off it never does. Your agent’s own searches of your team’s history still work, and session transcripts are still uploaded.

`hash`: lineage for data Probe must not read

Use hash for data whose contents must stay on your machines but whose history you still want: patient records, customer data, anything licensed. Steps over it still become runs, and each file a step reads or writes is recorded as a hash and a row count, so the record shows that one step read raw.jsonl (30,000 rows) and wrote clean.jsonl (29,156 rows) without Probe ever seeing a record (probe run inputs lists a run’s reads). Probe does not link runs by matching those hashes. A run’s link to the dataset or checkpoint it used comes from a use that your agent or the SDK records, naming a fixed version (probe dataset use dataset:<name>@<N> --run <run>, or run.use_dataset in the SDK).

data/.probe.config

@probe hash
*

A step that is not a Python script (head, jq, cat | shuf) names its files so they are hashed too: probe exec --input raw.jsonl --output clean.jsonl -- sh -c "...".
probe exec keeps no console log when its folder, or any path it names, is hash, read or off: the step may print records.
A file under 64 bytes is recorded with its path and size only, no hash. A hash reveals a file’s contents only to someone who can already guess them exactly; for a short, predictable file that is easy, so Probe does not send one.
Uploads (probe dataset add --upload, log_dataset(upload=True)) of a hash path are refused. A checkpoint saved in a hash folder is recorded as a reference (path, hash, size, step), never uploaded.
One path has one level, so read and hash do not combine.

Common setups

One repo out
Only some repos in
Some paths out
Look, don't record
Machine-wide

In the repo’s root:

.probe.config

In the folder that holds your repos:

~/code/.probe.config

/*
!/repo-a/
!/repo-b/

Use /*, not *. As in git, * matches at every depth, so * then !/repo-a/ still excludes every file inside repo-a.

.probe.config

data/
secrets/
*.ckpt

The daemon may use the code as context, but records nothing about it:

.probe.config

* @probe read

~/.probe.config applies to everything under your home folder.

How a path is decided

Walk up from the path

Every .probe.config from the path up to / counts, including files above a repository’s root. A folder without one inherits its parents’ rules.

The last matching line wins

A later line beats an earlier one, and a deeper file beats a shallower one. /probe on cannot override a level; edit the file instead.

An off folder is closed

As in git, nothing inside a folder at off is looked at, its own .probe.config included, and no line turns anything inside it back on. To keep one file, close the contents instead of the folder:

.probe.config

secret/*
secret/important.txt @probe on

A folder at read or on stays open: what is inside takes the folder’s level unless a line matches it directly. A deeper .probe.config may narrow a read folder to off, but never turns it back on.

No match: the session decides

Otherwise the session’s /probe on, read or off decides, as before. A level only ever narrows what the session allows.

To see a path’s level and the line that set it:

probe ignore check data/rows.jsonl

Files above where the agent started

A .probe.config in the folder the agent started in, or below it, always applies. One above that folder is yours to follow or set aside for each session: the agent asks you once, the first time Probe comes up, and follows the file until you answer.

What is enforced where


`probe` CLI	A write whose working directory or a path it names is `read` or `off` is refused. `probe exec` runs the command without opening a run. At `hash`, `probe exec` opens the run and records the files’ identities, with no console log.
Python SDK	`probe.init()` in a `read` or `off` folder (or from such a script) opens no run and says so; code, output and read capture skip those paths; `log_artifact` on one records nothing. At `hash`, the run opens and reads and writes are recorded by identity only; nothing about the file’s contents is sent.
Coding agent	Told at session start which paths `.probe.config` blocks Probe from recording; may create or edit a `.probe.config` when you ask or when it judges a path should not be recorded, and tells you each change.
Probe daemon	Drops the agent’s work done in an `off` folder, naming an `off` file, or searching a folder that holds one (`rg`, `grep -r`, `find`, `git diff`), and its own commands never read one. Sees a `read` path’s work tagged; any write naming it is refused.

The daemon works from what the agent’s commands say. A program that opens an off file without its command naming it (a training script reading data/) is outside this, as is text you or the agent type into the chat. Session transcripts still upload whole.

A .probe.config that exists but cannot be read excludes everything under its folder. @probe levels need CLI 0.231.0 and plugin 0.127.0 or later: older versions read x @probe read (and x @probe off) as a pattern that matches nothing, so they record those paths. Plain lines work from CLI 0.227.0 and plugin 0.125.0, so write off as a plain line. @probe hash needs CLI 0.253.0 or later: older versions read it as off, so they record nothing there (the safe direction). Update with probe wizard --action update --yes.

Replacing folder defaults

probe session default --folder and the .probe/config.json files it wrote are retired. An old one set to off or read still keeps its folder out of Probe in coding-agent sessions, the only place it ever applied. probe doctor lists them, and probe doctor --fix turns each into a .probe.config.

Getting started

Setup wizard

Tracking your work

Python SDK

CLI

Coding agents

Dashboard

Integrations and imports

Deploy and operate

Reference

Levels

`hash`: lineage for data Probe must not read

Common setups

How a path is decided

Files above where the agent started

What is enforced where

Replacing folder defaults

​Levels

​hash: lineage for data Probe must not read

​Common setups

​How a path is decided

​Files above where the agent started

​What is enforced where

​Replacing folder defaults

Levels

`hash`: lineage for data Probe must not read

Common setups

How a path is decided

Files above where the agent started

What is enforced where

Replacing folder defaults