foldrun 17.0.0 → 17.0.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/docs/security.md +45 -0
- package/package.json +1 -1
package/docs/security.md
CHANGED
|
@@ -153,6 +153,51 @@ use and role change is in the audit log; every step's tool calls, script
|
|
|
153
153
|
runs and refusals are in the run's events. A run can be reconstructed from
|
|
154
154
|
its record.
|
|
155
155
|
|
|
156
|
+
## What our telemetry carries
|
|
157
|
+
|
|
158
|
+
Your run's record — its prompts, replies, files, tool arguments and errors —
|
|
159
|
+
is yours, in your workspace, and goes where you send it (`telemetry:` to your
|
|
160
|
+
own OpenTelemetry collector, `notify:` to your webhook). The platform's own
|
|
161
|
+
logs and metrics are a different thing: they leave the cluster for our
|
|
162
|
+
operations stack, and they never carry your content.
|
|
163
|
+
|
|
164
|
+
| Carries | Never carries |
|
|
165
|
+
|---|---|
|
|
166
|
+
| ids: account, run, job, task, lease, pod | prompts, replies, a step's output |
|
|
167
|
+
| the names of your workspaces, flows, agents and evals | file names or contents, knowledge, memory |
|
|
168
|
+
| counts, durations, bytes, dollars, HTTP status | tool or script arguments and results |
|
|
169
|
+
| an error's class (`timeout`, `oom`, `disk_full`, `model_error`, `network_error`, `policy_denied`, `user_code_error`, `not_found`, `cancelled`, `platform_error`), its type, its errno or SQLSTATE, and where in **our** code it happened | an error's message or stack text, a model process's stderr, a query or its values |
|
|
170
|
+
| a reason word from a fixed set (`paused`, `no_egress_url`) | URLs, email addresses, anything with a space in it |
|
|
171
|
+
|
|
172
|
+
It holds by construction, not by care:
|
|
173
|
+
|
|
174
|
+
- **One writer.** Every platform log line is built from a closed list of
|
|
175
|
+
fields, each with a shape: a number, a short word, or an id with no space,
|
|
176
|
+
quote, slash, colon or `@`. A value that does not fit is dropped, not
|
|
177
|
+
shortened. An error gives its class and code, never its words.
|
|
178
|
+
- **One boundary.** Every hosted process (web, worker, egress, assistant)
|
|
179
|
+
replaces its console and its stdout/stderr at start. Anything else that
|
|
180
|
+
prints — our older code, a dependency, Next.js — becomes "a line was
|
|
181
|
+
printed here", never what it said.
|
|
182
|
+
- **A second line at the collector.** The log shipper forwards a line only
|
|
183
|
+
if it is exactly such a line, and never reads the namespace where steps
|
|
184
|
+
run.
|
|
185
|
+
- **Tests.** A process inside the boundary tries every way to print a
|
|
186
|
+
marker (an error's message, stack, name and code; a forged envelope; a
|
|
187
|
+
raw write; an uncaught rejection) and the test reads the raw bytes; the
|
|
188
|
+
source of core, platform and web is scanned for any log call that prints
|
|
189
|
+
more than a literal; and metrics may only have fixed names and labels.
|
|
190
|
+
|
|
191
|
+
The incident this answers: a judge's model process crashed, and the full
|
|
192
|
+
text of its error — which carried prompt text — went to the server log.
|
|
193
|
+
Today that line is `model.process_failed` with `errorClass: model_error`.
|
|
194
|
+
|
|
195
|
+
A self-hosted install writes the same lines to its own logs;
|
|
196
|
+
`FOLDRUN_LOG_MODE=raw` on a process puts it outside the boundary, for
|
|
197
|
+
debugging on your own machine. On a laptop (`foldrun run`, `next dev`) there
|
|
198
|
+
is no boundary: the reader is the person whose content it is, and errors
|
|
199
|
+
print in full.
|
|
200
|
+
|
|
156
201
|
## People
|
|
157
202
|
|
|
158
203
|
Four roles, least privilege by route, one owner. See
|