Motivation

• Treat tool JSON embedded in non-code-block prose as executable tool calls (instead of always preserving as plain text) to support mixed-content streaming tool handoff.
• Preserve fenced code blocks as examples and do not execute tool snippets inside them.
• Allow more flexible tool-name matching (qualified names and punctuation variants) so canonical tool names resolve more robustly.
• Map a developer role to system in prompt normalization to keep role semantics consistent.

Description

• Updated streaming sieve logic in internal/adapter/openai/tool_sieve_core.go to emit pending tool_calls events for mixed prose and to preserve surrounding prefix/suffix text properly.
• Relaxed consumeToolCapture constraints so non-whitespace wrapping no longer forces tool JSON to be treated only as plain text when it should be parsed as tool calls.
• Added fuzzy tool-name resolution in internal/util/toolcalls_parse.go with resolveAllowedToolName and a loose punctuation-stripping pattern to accept qualified or punctuation-variant names (e.g. mcp.search_web, read-file) and canonicalize them.
• Introduced normalizeOpenAIRoleForPrompt in internal/adapter/openai/message_normalize.go and applied it in responses_input_items.go and message normalization to convert developer to system.
• Adjusted and added unit tests across the repo: OpenAI handler stream/toolcall tests and responses tests, Gemini adapter tests, toolcalls parser tests, and message normalization tests to reflect the new behavior.
• Documentation updates in API.md, README.en.md, and TESTING.md clarifying that tool JSON in non-code-block contexts can be recognized even when mixed with prose and that fenced code block examples are exempt.

Testing

• Ran go test ./... -count=1 including updated packages; unit test suite passed.
• Executed modified/added Go tests in internal/adapter/openai, internal/adapter/gemini, and internal/util (e.g. toolcalls parsing and message normalization tests) and they succeeded.
• Confirmed streaming behavior checks by exercising updated SSE parsing tests (responses and chat stream tests), which passed under the new expectations.

Codex Task