Claude models are served on /messages, which the OpenAI SDK can't speak, so
those runs go through LiteLLM's Anthropic route instead. Prompt caching moves
with them, since LiteLLM consumes the injection points the raw SDK rejects.
Zen and Go now show up by name instead of both reading 'OpenCode
subscription', and Zen keeps its cost tracked: it bills prepaid credits per
request, so those runs were never actually free.
* fix(llm): cap the tool calls one assistant response may queue
* fix(llm): cap the subscription backend's responses too
---------
Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>
* perf(cli): fast startup — lazy heavy imports + onedir standalone build
* perf(cli): drop legacy single-file compat from install/self-update
* perf(cli): simplify — drop constants module and extra lazy-import refactors
* refactor(update): strix --update just re-runs the install script
* perf(cli): drop packaging/install/update changes; deepen lazy imports instead
Reverts the onedir build, install.sh, and self-update changes so release
mechanics stay untouched. Startup cost is addressed purely by deferring
heavy imports (agents/openai, config.models, report state/writer, docker)
until a scan actually runs; DEFAULT_MAX_TURNS moves to strix.config.settings
so argparse no longer pulls the agents SDK.
---------
Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>
Some OpenAI-compatible gateways don't support Server-Sent Events (or
deliver them unreliably), but the SDK run loop Strix uses only issues
streamed requests, so such a gateway fails every turn. Add an opt-in
LLM_DISABLE_STREAMING setting that wraps the resolved model in
_NonStreamingModel: each turn makes one non-streaming get_response and
replays the completed result as a single terminal stream event, so tool
calls, usage, and the rest of the agent loop are unchanged. Subscription
(ChatGPT) models are always streamed and are not wrapped.
A configured tool_output_max_bytes smaller than the truncation notice
itself can't fit a bounded preview, so a persisted result could exceed the
ceiling. Enforce a config floor (ge=1024) so nonsensical values are
rejected at load time instead of being worked around at runtime.
Cap the size of every tool result so a single verbose command (recursive
find, noisy scanner, full page dump) can't pin the conversation near the
model's context window for the rest of a scan.
- New ContextSettings config group with env-tunable caps.
- Default the SDK shell tools' max_output_tokens so exec_command /
write_stdin truncate head+tail instead of returning unbounded output.
- Bound Strix's own FunctionTool/CustomTool results (line + UTF-8 byte
head+tail preview with a truncation notice) and cap error strings.
An httpx.Timeout in ModelSettings.extra_args crashes
ModelSettings.to_json_dict() (PydanticSerializationError) on the Chat
Completions and LiteLLM model paths, which serialize settings for their
tracing generation span — failing every model turn on those paths. Pass
the timeout as a plain float, which httpx-based clients apply as the
read (inactivity) timeout.
The SDK's http_status retry policy only retries errors carrying a known
HTTP status code, but quota/billing (and other provider-side) failures
often surface inside a streamed response as a bare error with no status
code, so they were failing on the first attempt. Add a statusless retry
policy to DEFAULT_MODEL_RETRY (retry count and backoff unchanged) so they
are retried before a genuine exhaustion fails the run; user aborts are
never retried.
The SDK's http_status retry policy only retries errors carrying a known
HTTP status code, but quota/billing (and other provider-side) failures
often surface inside a streamed response as a bare error with no status
code, so they were failing on the first attempt. Add a statusless retry
policy to DEFAULT_MODEL_RETRY so they are retried (before any content is
streamed; user aborts are never retried), restoring the pre-SDK engine's
resilience. If the provider is genuinely exhausted, the error still
propagates and fails the scan after retries.
feat(inputs): implement logic for required tool choice based on model
test(inputs): add tests for force_required_tool_choice behavior
test(runner): update tests to include force_required_tool_choice in settings
_read_json_overrides is documented to let env vars outrank the
persisted cli-config.json, but it decided per-alias and broke on the
first alias found in either env or the file. When a multi-alias field
(e.g. api_key via LLM_API_KEY/OPENAI_API_KEY) was set in the env under
one alias but stored in the file under another, the stale file value
was surfaced as an init kwarg and overrode the live env var. A
lowercase env var was also missed (settings use case_sensitive=False).
Decide whether a field is already set in the environment by checking
all of its aliases case-insensitively before consulting the file. Add
regression tests for the cross-alias and case-insensitive cases.
Closes#688
* fix: resolve pre-commit check failures
- Change RuntimeError to TypeError for type validation in report/writer.py
- Update pyupgrade to v3.21.2 for Python 3.14 compatibility
* chore: add pytest test infrastructure
Mirror the layout introduced on feature/438-token_budget: pytest +
pytest-asyncio dev deps, asyncio_mode auto, a tests.* mypy override, and
pytest in the mypy pre-commit hook deps so the tests/ package type-checks.
* feat: add --mount and large-target pre-flight for local repos (#492)
Large local targets were copied into the sandbox file-by-file via the SDK
LocalDir entry, which stalls on big repos and could leave /workspace empty.
- --mount <path> bind-mounts a host directory read-only at /workspace/<subdir>
instead of copying it, bypassing the per-file stream.
- A size pre-flight (STRIX_MAX_LOCAL_COPY_MB, default 1024) fails fast with a
clear message suggesting --mount when a non-mounted local target is too big.
* fix: reject empty --mount paths
An empty or whitespace-only --mount value resolves to the current working
directory and would silently bind-mount it into the sandbox. Reject it.
* fix: dedupe local targets so a dir is never both copied and mounted
If the same directory is passed via --target and --mount (or as duplicate
values), it previously produced two targets — copied AND bind-mounted, and
the copied one could trip the size pre-flight. Dedupe by resolved path,
preferring the bind mount.
* fix: treat non-positive STRIX_MAX_LOCAL_COPY_MB as disabled
Previously a value of 0 (or negative) made every local target count as
oversized, aborting all local scans. Now <= 0 disables the pre-flight.
* fix: log unreadable subtrees during size pre-flight
os.walk silently swallowed directory-listing errors, so a permission-denied
subtree could make a large repo under-count and slip past the pre-flight.
Surface such omissions via an onerror warning.
* docs: document --mount and STRIX_MAX_LOCAL_COPY_MB
Add CLI reference + example for --mount, document the size pre-flight env var,
note the read-only-is-not-a-hard-boundary caveat and that remote repos are not
size-checked, and clarify the backends docstring on when bind mounts apply.
* Update strix/interface/main.py
* Update strix/runtime/docker_client.py
---------
Register a litellm.success_callback that captures kwargs['response_cost']
into a new observed-cost bucket on LLMUsageLedger. record() skips the
tokens-times-registry estimate for LiteLLM-routed models so we do not
double-count with the callback; OpenAI direct routes keep estimating
since LiteLLM is not invoked for them. Per-agent attribution for
LiteLLM-routed calls is apportioned by token share at to_record() time.
OpenAI's Responses API rejects reasoning.effort on non-reasoning
models like gpt-4o with `unsupported_parameter`, so any scan with
the default STRIX_REASONING_EFFORT=high against gpt-4o crashed at
the first model call. drop_params=True absorbs the rejected param
on LiteLLM-routed models but the SDK's native OpenAI path has no
equivalent.
Lift model_supports_reasoning to a public helper that strips
litellm/, any-llm/, openai/ prefixes and falls back to last-segment
lookup so prefixed forms like anthropic/claude-opus-4-7 resolve
through the bare model_cost entry. make_model_settings regains
model_name and skips Reasoning() when the registry doesn't confirm
support. uses_chat_completions_tool_schema reuses the same helper
(was duplicating the lookup under a misleading name).
OpenAI's Responses API rejects tools[i].type="custom" on non-reasoning
models like gpt-4o (400 with code=unknown_parameter, param=tools).
Strix's SDK-native Filesystem capability registers CustomTool entries
by default, so a bare STRIX_LLM=gpt-4o run failed at the first tool
invocation even though warm-up (a tool-less call) succeeded.
uses_chat_completions_tool_schema now consults
litellm.model_cost[<name>].supports_reasoning for OpenAI routes and
flips to the chat-completions function-tool schema for models that
don't carry the reasoning flag. Same registry-lookup pattern as
is_known_openai_bare_model. Non-OpenAI prefixes and configs with
LLM_API_BASE are unchanged (still function tools).
yoni