[kelkalot]

Fixes #126, #127 and #128. These are the remaining places where a target passed the system under test something other than the scenario's real input.

Changes

• CallableTarget (CallableTarget passes the scenario's evaluation-only document marks to the wrapped function #126). The wrapped function now gets:

  • the history without the engine's markers (documents, decision, tool_events);
  • the scenario's documents as text through documents, on every turn;
  • its files through file_uri, on the turn that attaches them. That turn keeps its file_uri in the history, so the function can attach the files where they belong.

  No evaluation-only document mark reaches it.

• HTTPAppTarget system prompt (HTTPAppTarget ignores the audit's system prompt #127). The audit's system_prompt is now sent:

  • With the new system_field, it goes to that request field in either body mode. That covers APIs that take the system prompt at the top level, such as Anthropic's native Messages API. A value the template already sets there comes first; content-block lists get the prompt appended as a text block.
  • Without system_field, in messages mode, it joins the template's leading system message or becomes one. Many chat templates reject two system messages in a row. A messages list in request_template is kept ahead of the conversation; before, it was overwritten.
  • Without system_field, in other modes, it can't be sent, and a warning says so.

• HTTPAppTarget template history (A history key in HTTPAppTarget's request_template replaces the conversation history #128). A history list in request_template comes before the conversation's earlier turns instead of replacing them. A history value that isn't a list is now an error.

• Shared helpers.

  • targets/base.py gains ENGINE_MARKERS, plain_history() and history_file_uri().
  • context_marks.document_texts() reduces documents to text.

  The callable, HTTP, decision and OpenEnv targets use them. DecisionTarget now treats an explicit documents=[] as given.

• README. "Writing your own target" describes what a CallableTarget function and an HTTPAppTarget request receive.

Behaviour changes

• CallableTarget functions no longer find documents on history[0], and get file_uri only on the turn that attaches it.
• HTTPAppTarget sends the system prompt when the audit sets one: as part of the leading system message in messages mode, or at system_field.

Tests

Eight new tests in tests/test_targets.py. They cover:

• what a CallableTarget function receives in a two-turn audit with marked documents and a file;
• documents passed directly to send;
• system prompt routing in both body modes, including joining with the template and the warning, also when warnings are errors;
• template messages and history ahead of the conversation, and a non-list history.

pytest -n auto: 1725 passed.

Live checks with claude-haiku-5-5:

• Anthropic's native Messages API (message_field="messages", system_field="system"): the template's and the audit's system prompts went out as one top-level system field, and the model followed them on both turns.
• The OpenAI-compatible endpoint: a template system message and the audit's prompt went out as one system message, with the same result.
• A CallableTarget wrapping a small RAG function: it built its prompt from the documents as text, with no marks in its requests.