run_strix_scan(subagent_turn_limits={skill: turns}) gives each sub-agent the
smallest limit among its skills. The agent is warned 10 turns before the limit
and reminded every 5 turns past it, but never stopped.
* feat(skills): expand XSS coverage for HTML injection and normalization
* Update xss.md
---------
Co-authored-by: alex s <46074070+bearsyankees@users.noreply.github.com>
Verifying http_exchange_ids fetched every request with its full raw
request and response bodies, serially, and timed out on busy proxy
projects; the ids were then dropped and the finding lost its evidence.
Check existence with an id-only GraphQL query, retry once, and when the
proxy still does not answer keep the ids with a warning instead of
discarding them.
A cached sandbox image built before the entrypoint honored
STRIX_CAIDO_LISTEN_ADDRESS still starts Caido on 0.0.0.0:48080 with
--allow-guests, so enabling STRIX_DOCKER_SANDBOX_NETWORK=host on it
would expose guest access on every host interface. The Dockerfile now
labels the image strix.sandbox.caido-loopback=true, and host mode
checks that label on the image before creating the container. An image
without it raises a ValueError instead of starting. The configuration
docs note the requirement.
The reserved host.docker.internal check in STRIX_DOCKER_EXTRA_HOSTS now
compares case-insensitively, since hostname lookup is case-insensitive
and a mixed-case key could otherwise add a second entry for the
reserved name. Tests cover the missing-label rejection and a mixed-case
reserved hostname.
Refs #1402
* feat(agents): let callers remove base tools from scan agents
* feat(agents): let callers choose the report tool instead of removing tools
* refactor(agents): name the report tool via the prompt template, not string replacement
STRIX_DOCKER_SANDBOX_NETWORK=host attaches the host network, drops
published ports, sets STRIX_CAIDO_LISTEN_ADDRESS so the entrypoint binds
the proxy to 127.0.0.1, and resolves that fixed port on the daemon
loopback. Any other value attaches the named network and uses the
container address there; leaving it unset keeps the published localhost
port. STRIX_DOCKER_EXTRA_HOSTS merges a JSON map of hostnames to literal
IPv4 or IPv6 addresses into the sandbox hosts entries, keeps
host.docker.internal reserved for host-gateway, and rejects invalid
JSON, empty names, and non-IP values before the container is created.
The configuration notes that host mode follows a VPN only with a local
Linux Docker engine and that port 48080 allows one host-mode scan per
daemon.
Fixes#1402
BudgetPausedError carries whether the operator pause caused it, so a resume that lands between the hook's refusal and the loop's handler still parks and continues instead of exiting under the stop policy. After the resume the agent drains its mailbox into the session before retrying the call, so a message queued while it was held is not skipped.
* feat(llm): extra request-body fields via LLM_EXTRA_BODY
Some provider parameters, such as OpenAI's service_tier and
safety_identifier, are read only from the request body, so
LLM_EXTRA_HEADERS can't set them. LLM_EXTRA_BODY merges a JSON object
into the body of every request to the configured model, on both routing
paths. A dedicated dedupe model gets its own DEDUPE_LLM_EXTRA_BODY
instead, as with headers.
* fix(config): decode saved JSON-object settings when loading
persist_current() saves env values verbatim, so LLM_EXTRA_HEADERS,
LLM_EXTRA_BODY and their DEDUPE_ variants land in cli-config.json as JSON
text. The loader passed that text unchanged to the dict fields, so a later
launch without the env var failed validation. Decode saved strings for
dict fields, as pydantic-settings does for env vars.
---------
Co-authored-by: Szymon Nastaly <szymon.nastaly@autoform.ch>
* fix(models): replay malformed tool-call arguments as valid JSON
A model that emits a tool call with truncated or otherwise invalid JSON
arguments fails that one call, but the raw string is recorded in the
session. Strict OpenAI-compatible servers (vLLM, SGLang) validate every
assistant tool call in the request, so each later turn is rejected with
"Assistant tool call function.arguments must be valid JSON" and the agent
never recovers. The turn guard now rewrites such arguments on the way out
to a JSON object carrying the original text, alongside the existing
duplicate call-id repair.
* fix(agents): tell the model to re-issue tool calls whose arguments are not valid JSON
A streamed turn that ends mid-tool-call leaves the client with a JSON prefix
as the arguments. Instead of the SDK's generic parse-error result, the tool
wrapper now short-circuits with a clear 'not executed, re-issue with complete
arguments' message so the model recovers in one turn.
* fix(models): treat NaN/Infinity tool-call arguments as malformed
* test(cost): pin model resolution in the provider-prefixed cost test
* fix(models): treat over-nested tool-call arguments as malformed
* feat(llm): log the upstream provider behind OpenRouter
LiteLLM drops OpenRouter's top-level `provider` from stream chunks (including error chunks) and the `metadata.provider_name` of non-2xx replies, so the request log could not say which provider served or failed an attempt. Record both, plus OpenRouter's `metadata.error_type`, per attempt and emit them as `upstream=` / `upstream_error=` on the llm_request line.
* feat(llm): retry 4xx model errors
* test: let the blocked-provider test run out of 4xx retries
* refactor: retry any 4xx or 5xx status
* fix: don't retry 401, 402 or 403
* fix: don't retry 404
LiteLLM drops OpenRouter's top-level `provider` from stream chunks (including error chunks) and the `metadata.provider_name` of non-2xx replies, so the request log could not say which provider served or failed an attempt. Record both, plus OpenRouter's `metadata.error_type`, per attempt and emit them as `upstream=` / `upstream_error=` on the llm_request line.
* fix(execution): only strip images on image-rejection errors
The image-strip recovery fired on any 400/404/422, wiping a session's screenshots and replaying the turn on unrelated errors (e.g. an upstream incomplete-tool-call 400, a context overflow). Require an image-related message too.
* fix(execution): match only replies that mean the model takes no images
* fix(execution): match DeepSeek's image rejection
* fix(execution): guess at Vercel AI Gateway's image rejection
* feat(llm): scrub images from every request to a text-only model
Replace images at the MCP call before any result_transform, give load_skill the
same agent_browser screenshot swap as the system prompt, and add a
call_model_input_filter that turns any image left in the input into a tool error
while chaining any filter already set.
* chore: fix the ruff and bandit failures on main