FindingsCSV wrote name, class, summary, and task text as-is. A value
starting with =, +, -, @, tab, or CR is evaluated as a formula when
the file is opened in Excel or LibreOffice. Those cells are now
prefixed with a single quote.
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.
insert_assets swallowed ListAssetInterceptRules and
TaskInterceptRulesSplit errors and evaluated an empty rule set, which
allowed every asset. A load failure now refuses the whole call. The
tool result is generic; the driver error stays in the process log.
offset+length > size wraps when length is near MaxInt64, so the cap is
skipped and make([]byte, length) can exhaust memory. Compare length
against size-offset instead.
marshalStringArray quoted double quotes but not backslashes. In a
Postgres array element a backslash escapes the following quote, so a
value such as x\",evil,\"y could close the current element and inject
another. Double backslashes first, matching marshalJSONBArray.
trimErr cut the error text at byte 300, which can split a multi-byte
character. The result is stored as llm_profile_health.last_error, and
PostgreSQL rejects the invalid UTF-8, so the breaker state of a profile
whose provider answers in Chinese is never persisted and the write is
retried every second.
json.Unmarshal of JSON null into a map yields nil. syncBrowserMCPProxy
then assigned NODE_EXTRA_CA_CERTS under traffic capture and panicked,
including during NewManager on every restart once that row was stored.
in_scope_not_asset counts in-scope events that are not asset types, not
gate rejects: a rejected candidate comes back as an insertion forbidden
error in the insert_assets result. EvaluateAssetGate also applies block
rules whether or not an allow-list exists, so the no-allow-list gap is
about the allow rule, not the whole gate.
A turn the provider's safety system declines (Anthropic stop_reason
"refusal", OpenAI-compatible finish_reason "content_filter") has no text
and no tool calls. Norma passes the stop reason through and the harness
ends any tool-less turn with ReasonCompleted, so the run summary said the
model "ended this turn normally", and steerHooks.Stop took the empty turn
for an idle one and resent the conversation up to the empty-response retry
count.
The LLM recorder is the one layer that sees the stop reason. It now hands
each completed call's stop reason to an optional context sink
(llmrec.WithStopSink). captureRunSession installs a per-run tracker
(agent.WithStopTracker) on that sink. With it:
- steerHooks.Stop does not inject the idle-turn rerun after a refusal;
- terminalText reports "provider refusal (stop_reason ...)" and says the
intent was not completed, instead of the completed wording.
Other terminal states and ordinary empty turns are unchanged.
Fixes#327
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
States the trust boundary discussed in #317:
- output.json's scope_distance is the integrated scanner's own verdict
and is ingested as-is; in-scope events are not re-checked against
ARTEX's rule set.
- The enforced boundary is the task allow-list: assets reach the graph
(insert_assets) only through EvaluateAssetGate, and rejects are
recorded as in_scope_not_asset.
- The no-allow-list case is an accepted gap, with operator guidance:
configure an allow-list per task or review output.json before feeding
it to ARTEX. Cross-checking ingested events against ARTEX's rule set
would redefine existing scanner configurations, so it is a deliberate
design question, not an accepted vulnerability.
Option 2 as picked on #317; documentation only.
FindingFilename cut the name at 120 bytes, which can split a multi-byte
character. archive/zip only sets the UTF-8 flag on valid names, so the
entry was written without it and unzip tools decoded the whole name as
CP437, garbling every non-ASCII character in it.
The rebase onto current main keeps the original entries and adds the three
user-facing changes merged after its base: the finding report_state field,
the ordered LLM health persistence, and the framed intent summaries.
The /system/llm/ badge read llm_profile_health, which only the failover
Pool wrote. A task pinned to an explicit chain, a bound agent with
failover off, and the global provider all call the provider directly, so
a dead backend stayed "Healthy".
One policy now lives in llmpool.Registry.RecordCall, used by the Pool and
by the task runtimes: a success clears the state, a failover-eligible
failure trips it, context cancellation and deterministic 400s do not, and
a stream that already delivered output is left alone. The selection
carries the profile whose breaker the call updates, and a provider
already wrapped in the Pool reports nothing to avoid double counting.
The LLM recording page also explains an empty list while the llm_record
toggle is off, after the settings call resolves.
Originally reported in #257.
The wake lines interpolated a planner-written intent summary into the trusted sentence, and the overview kept every summary in the plain JSON, so an inherited summary crossed tasks as if this planner had written it (#124, F3). Frame the summary on the wake lines and move the open/running/recent-done summaries into one framed block; the structured intent fields stay plain.
A newer state for the same profile replaces one that has not been written yet, and a failed write is logged instead of discarded on a detached goroutine.
The status line said the proxy does not apply rules yet. Replace it with
the request and response phases, what exchange_rewrites records and how
the hashes are formed (plain SHA-256, keyed HMAC for sensitive actions),
the SSE and WebSocket behaviour, the 5 MiB response pass-through, the
snapshot limits (live slots are never deleted by the TTL), and the fact
that replays go through the proxy and are rewritten like captures.
It now says that traffic_get shows the originals of a rewritten exchange
and that traffic_search never finds a value a rewrite rule wrote. As for
traffic_search, only the original built-in text is replaced.