image_path was opened as a raw filesystem path, so any readable
png/jpg/jpeg/webp/gif on the host could be copied into the operator
transcript. The path is now resolved from the task working directory
and refused when it contains .. or EvalSymlinks outside that tree.
The goal gate accepts a goal that names its target as a bare domain (goalDomainRE), but the in-goal set was built only with commandTargets, which does not read bare hosts. A question citing the operator's own target as a URL was then tagged target-derived. goalTargets adds the goalDomainRE matches, widened to the whole dotted name, lowercased, port and path dropped. The question-side parser is unchanged.
An ask_operator question that names a host or URL not among the task's
goal targets is tagged "contains target-derived text": the hosts go into
the transcript activity's Detail payload as target_derived_hosts, and the
transcript shows the tag on the request card when the field is set.
Hosts are read with intercept.ExtractTargets (the target gate's
command-string rules) from both the question and the task goal, and
compared without the port. The text the model reads, the pause, and the
operator's decision are unchanged: the tag is a visible hint only.
Audit F12, 0.7.0.
When a worker hits an environmental failure (sandboxed or filtered
network path, missing binary or key, TLS or proxy interception, missing
runtime permission), nothing today tells it how to stop: transcripts
show turn after turn of retries with different flags, equivalent tools,
or fallback transports, all burning budget against a wall the agent
cannot move.
Changes:
- the ask_operator tool description now covers environmental blocks
explicitly: do not retry or work around; include the exact single
command the operator should run in their own terminal as one
copy-pastable line, plus what to paste back, then wait. The question
parameter description gains the same guidance.
- worker and pentest seeds gain an "Operator handoff on environmental
blocks" block: stop after the first environmental failure, no
flag-hopping, no bypass hunting, one command per ask_operator request,
never batch.
- tests: the tool description must keep the handoff markers
(TestAskOperatorEnvHandoffGuidance) and the executor seeds must keep
the block (TestPromptSeedsCarryOperatorHandoff).
Pattern adapted from bug-bounty-hunting-prompts (MIT), whose operator
handoff rule produced exactly this single-command format.
A worker hitting a wall only a human can clear (captcha, email/SMS OTP, 2FA,
a decision) calls ask_operator with a precise question and an optional
screenshot. The question lands in the intent's transcript as a user-role turn
with an attachment card (the same {text, attachments} shape chat uploads use,
so the existing transcript UI renders it natively), and the intent pauses. The
operator replies through the existing message-to-worker flow; the reply resumes
the run with the answer as the next input. Images are whitelisted (png/jpg/
jpeg/webp/gif), capped at 10MB, and copied into the task uploads dir.
Closes#41