refactor(v1): share the dialects' content walker, mediation and preamble

[hallerite]

Stacked on #2503 (which is stacked on #2496). Behaviour-preserving; see the equivalence check at the bottom.

The Anthropic and Responses dialects each carried their own copy of the same three pieces, and two other exact twins sat in responses.py and between chat.py and clients/train.py.

Now shared in dialects/base.py

• blocked_path(value, path, policy, blocked_item): the list walk, non-dict pass-through and non-direct caller gate that both blocked_content_path functions started with verbatim. Each dialect keeps its blocked_content_path name and its own per-item rules (_blocked_block / _blocked_item), so every call site and the recursion through nested content are unchanged.
• mediate_parts(value, path, policy, blocked, wrappers=()): the list mediation both mediate_content functions implemented. Anthropic's extra branch (check a wrapper block without its content, then mediate its content in place) is the wrappers parameter; Responses passes none, which is exactly its old body.
• user_and_tool_messages(request): the two list comprehensions both rewrite_request methods opened with.

Exact twins removed

• responses.py: the content → input_text/input_image conversion appeared twice inside rewrite_request; it is _content_to_input now.
• chat.message_to_wire and clients/train.serialize_completion had the same tool-call comprehension. It is tool_calls_to_wire in chat.py, which train.py already imported from. Only that comprehension is shared: serialize_completion keeps its own content and reasoning_content handling, because it deliberately differs from message_to_wire (content: null stays null, provider_state is not consulted).

Net lines are roughly flat (+121 / −126): the shared functions carry docstrings. The point is one implementation of each rule instead of two.

Equivalence check. With no dedicated dialect tests, I ran the same 81 inputs through the base branch and this branch and diffed the JSON output: mediate_content and blocked_content_path for both dialects across three network policies and nested/wrapper/caller/file/url cases, rewrite_request for both dialects with string and list inputs (changed and unchanged), message_to_wire for assistant/tool/user messages including content=None with and without tool calls, and serialize_completion with and without usage. Identical. Also ruff check, ruff format --check, ty check verifiers, pytest tests/v1 -m "not e2e" (82 passed).

🤖 Generated with Claude Code

Note

Extract shared content walker, mediation, and message selectors into base dialect

• Adds shared helpers to base.py: blocked_path for recursive policy-path traversal, mediate_parts for content mediation, and user_and_tool_messages for selecting UserMessage/ToolMessage instances
• Updates anthropic.py, responses.py, and chat.py to call these shared helpers, passing provider-specific block checks and wrapper types as callbacks
• Replaces inline tool-call wire-dictionary construction in train.py and chat.py with the shared tool_calls_to_wire serializer
• Extracts Responses typed-content conversion into _content_to_input to remove duplication in request rewriting
• Risk: provider-specific block callbacks (_blocked_block, _blocked_item) must preserve their original checks exactly; any missed condition changes what content gets blocked or removed at runtime

^{Macroscope summarized dfa7ec3.}

---

Note

Medium Risk
Changes only move network-policy blocking and content mediation into shared code; a regression in the dialect-specific block callbacks could alter what provider capabilities are stripped at runtime.

Overview
Behavior-preserving deduplication across Anthropic, Responses, chat, and the train client—no intended runtime changes.

New shared helpers in dialects/base.py: blocked_path (recursive policy traversal plus non-direct caller gate), mediate_parts (strip blocked content parts; Anthropic passes _CONTENT_WRAPPERS), and user_and_tool_messages for rewrite_request alignment.

Anthropic and Responses keep their own blocked_content_path / mediate_content entry points but delegate to those helpers via _blocked_block / _blocked_item. Responses also adds _content_to_input so typed content → input_text/input_image isn’t duplicated in rewrite_request.

tool_calls_to_wire in chat.py replaces identical comprehensions in message_to_wire and train.serialize_completion.

^{Reviewed by Cursor Bugbot for commit dfa7ec3. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved at dfa7ec3

Macroscope's review found this PR approvable — This PR consolidates duplicated content-walking, mediation, message-selection, and serialization logic without changing the resulting provider payloads or network-policy decisions. It introduces no new capability, schema, configuration, or deployment behavior.

^{You can add or adjust custom eligibility rules. Learn more.}