You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit fa813f6
Browse filesBrowse the repository at this point in the historyBrowse files
fix(opencode): allow long refusals in the text-only image control
The 500-character reply stop is right where only a short token is expected
(unset and explicit text+image), but the explicit text-only negative control
legitimately produces verbose refusals; a live smoke run halted on one. Use an
8000-character cap there, still stopping on credential-shaped content, and
truncate stored replies to 500 characters after classification and leak
detection.
Co-authored-by: OpenCode Agent <[email protected]>
Copy file name to clipboardExpand all lines: integrations/isolation/acq-kits/opencode/docs/synthetic-image-validation.md
+8-2Lines changed: 8 additions & 2 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -191,10 +191,16 @@ interrupt.
191
191
The script enforces these stop conditions and halts rather than continuing: a
192
192
literal credential in the rendered provider block, or the key variable unset;
193
193
a provider `baseURL` that is not `https://api.gsa.usai.gov`; a reply that is
194
-
oversize or carries credential-shaped data; an explicit `["text"]` run that
194
+
longer than 500 characters (unset and explicit `["text","image"]` runs, where
195
+
only a six-character token is expected) or carries credential-shaped data; an
196
+
explicit `["text"]` run that
195
197
returns the token; three consecutive models failing the text control; and an
196
198
OpenCode that cannot be run. A model whose own text control fails is recorded
197
-
and skipped (auth or routing, not an image result).
199
+
and skipped (auth or routing, not an image result). In the explicit
200
+
`["text"]` negative control a long, harmless refusal is the expected outcome, so
201
+
that run's length cap is 8000 characters instead (credential-shaped content
202
+
still stops the run); stored replies are truncated to 500 characters after
203
+
classification.
198
204
199
205
The script never edits `catalog.json`, a schema, or any committed capability
200
206
value, and its `candidate` column applies the decision rule above mechanically (and, stricter than the prose, also requires the text-only control samples to be plain non-token replies, so an outage cannot pass as a negative control).
0 commit comments