* 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>
* 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
Siblings differ only in the skills they were spawned with, but those came
first in <specialized_knowledge>, so their prompts diverged at 39%. Shared
skills and the catalog now come first, and the requested skills follow a
cache point, so siblings share 93%.
The extra system message takes a fourth Claude breakpoint, so the Bedrock
tool_config one goes: the first system breakpoint already covers the tools.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* perf(llm): give Claude a cache point before the per-run scope
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(llm): split the system prompt at a generic <cache_point> marker
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* perf(prompt): put per-run scope at the end of the system prompt
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(prompt): assert scope renders once, after the shared prefix
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A tool call for a name Strix does not register raised ModelBehaviorError
from the SDK turn resolver, which nothing retries: the root agent's raise
tore down the whole scan and a sub-agent died before its status was set.
Opt into the SDK's tool_not_found_behavior="return_error_to_model" so the
unknown call comes back as a tool result and the agent self-corrects.
The setting landed in openai-agents 0.19.0, which requires openai>=2.45,
so both pins move.
Co-authored-by: Jonathan Singer <jonathansinger@Jonathans-MacBook-Pro.local>
Co-authored-by: Jonathan Singer <jonathansinger@Mac-3004.lan>
Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>
