* 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
* feat(llm): keep images away from text-only models
Leave view_image out of the tools, replace MCP image blocks with a text note,
and swap agent_browser's screenshot section for text-only guidance when the
run's model does not accept images.
* refactor(llm): disable view_image via is_enabled and gate skill image guidance by modality
* refactor(skills): render opt-in Jinja skills instead of comment markers
* Tweaked skill
* feat(llm): detect whether a model accepts image input
Add model_supports_images(): OpenRouter models are checked against architecture.input_modalities from OpenRouter's public model list (fetched once per process), others against LiteLLM's supports_vision. Unknown models and lookup failures count as image-capable. Not wired in yet.
* refactor(llm): read image support from litellm's catalog, assuming it for unknown models
* style: drop a trailing space in the image-support docstring
_salvage_stream_to_session previously did pre_run_items + replay, which
duplicated items when stream.to_input_list() already included the
pre-cycle input (the seeded initial_input is in both). Detect the overlap
and avoid duplication by checking if replay suffix matches pre_run_items.
A stop that lands between the silent-wait check and mark_running was
overwritten with running. resume_silent_user_wait does the check and the
state change under the coordinator lock and leaves anything else alone.
Plain text is now the only channel to the user. wait_for_user takes no
arguments and only parks the agent, so a reply can no longer be written
twice (once as text, once as a tool argument). The driver bounces a
wait_for_user call that follows no assistant text back into recovery.
TUI and viewer render wait_for_user as the waiting marker alone and keep
rendering recorded respond_to_user events with their message.
docker.from_env() ignores the current docker context, so Docker Desktop on
macOS (socket under ~/.docker/run unless the default-socket option is on),
OrbStack and Colima reported DOCKER NOT AVAILABLE while `docker ps` worked.
Every failure also printed the same "ensure Docker Desktop is running" text
followed by a RuntimeError traceback.
- resolve the endpoint as the CLI does: DOCKER_HOST, then DOCKER_CONTEXT or
the current context, then the default socket; the sandbox backend uses the
same resolution so startup and scan talk to the same daemon
- classify the SDK error (socket missing, permission denied, connection
refused, Windows named pipe) and print the fix for the current platform,
the endpoint that was tried and the underlying error
- exit 1 cleanly instead of raising after the panel
- telemetry reports docker_unavailable_<reason>
parse_arguments() rejected --fail-on without -n before main() could switch
to headless, so a CI job running `strix -t x --fail-on high` without -n
stopped at an argument error instead of scanning. The terminal check now
lives in cli_args as terminal_attached(); the parser only enforces the
headless-only flag when a terminal is attached and the TUI would open.
Without a tty on stdin and stdout (CI, nohup, pipes, cron, TERM=dumb) the Go
TUI handshakes fine and then exits 1 the moment Bubble Tea tries to take over
the screen. Because the backend was already activated, that exit escaped the
pre-activation mapping and surfaced as a raw traceback:
RuntimeError: Bubble Tea TUI exited with status 1.
Now main() checks for a terminal right after argument parsing. With a target
it switches to the headless path (same as -n) and prints one dim notice.
Without a target the start screen is the only way to enter one, so it prints
a panel telling the user to pass -t <target> -n and exits 1. A bare --resume
is left to the picker, which already lists runs when there is no terminal.
If the sidecar still dies after startup, check_return_code raises
TuiProcessExitedError, run_tui maps it to InteractiveInterfaceExitedError,
and main() prints an INTERACTIVE INTERFACE STOPPED panel with the -n hint
under its own telemetry name instead of re-raising.
preflight_request takes api_base_setting so a dedupe model that timed out
points the user at DEDUPE_LLM_API_BASE, not LLM_API_BASE. warm_up_llm is
now exercised with a dedicated dedupe model: own headers, same preflight
timeout, resolved through resolve_dedupe_model.
The warm-up request used settings.llm.timeout (LLM_TIMEOUT, 300s) both as
the request timeout and the wait_for bound, so a wrong LLM_API_BASE or a
dead proxy hung five minutes and then printed an empty Error line.
Add LlmSettings.preflight_timeout (LLM_PREFLIGHT_TIMEOUT, default 30) and
a shared preflight_request() used for the main and dedupe models. When it
expires the panel names the model, the limit and the setting.
