Module api 

OpenAI-compatible /chat/completions client with a tool-calling loop.

Builds the request body, applies bearer auth, parses the first choice, and when the assistant returns tool_calls runs them via crate::tools::execute_tool and feeds the results back until it returns plain content. Prompt-cache markers and token-budget trimming are intentionally omitted for this first cut.

Structs

ApiStatusError

HTTP status failure returned by the OpenAI-compatible endpoint.

PendingToolCall 🔒

Spinner 🔒

A stderr spinner shown until the first streamed response chunk arrives. Active only on a TTY; a no-op otherwise so logs and redirected output stay clean.

StreamTextPrinter 🔒

StreamedChat 🔒

ToolLoopConfig

Read/edit sandbox + iteration cap for the tool loop.

Enums

StreamPrintState 🔒

Constants

DUMP_MAX_CHARS 🔒

Max characters of any single field printed in a verbose transcript dump.

SLOW_TOOL_LOG_AFTER 🔒

TOOL_CALL_CLOSE 🔒

TOOL_CALL_OPEN 🔒

Functions

append_json_text 🔒

apply_bearer 🔒

apply_stream_delta 🔒

arguments_to_string 🔒

arguments may arrive as a JSON string (OpenAI) or an inline object; normalize to a string.

chat

Run a chat turn with an optional tool loop. Returns the final assistant text. Every request’s token usage is summed into usage (the tool loop makes one request per iteration).

error_summary

One-line provider/request error for logs and reports.

find_sse_delimiter 🔒

floor_char_boundary 🔒

is_api_error

True if an error came from the provider/request layer rather than local tool execution or response parsing. The controller can skip a round for these.

message_content_to_string 🔒

normalize_tool_call_arguments 🔒

OpenAI-compatible chat history requires function.arguments to be a string.

normalize_tool_calls 🔒

parse_one_text_call 🔒

Parse one <tool_call> block body ({"name":..., "arguments"/"parameters":...}) into a synthetic OpenAI-shaped tool call, or None if it isn’t a valid call.

parse_text_tool_calls 🔒

Fallback for models that emit tool calls as <tool_call>{...}</tool_call> text in the message content instead of the structured tool_calls field (the Hermes/Qwen convention). Returns the parsed calls (in OpenAI shape) plus the content with the blocks stripped out.

path_with_line_range 🔒

preview 🔒

Truncate s to max chars with an elision marker, for verbose dumps.

read_streamed_chat 🔒

sanitize_tool_calls_for_history 🔒

Return a copy of calls safe to store in the chat history that gets re-sent on the next request.

sse_data 🔒

system_message 🔒

Build the system message. With cache=true the content is an Anthropic-style block carrying cache_control: ephemeral so the (large, constant) system prompt can be cached across requests; with cache=false it’s a plain string, byte-identical to a no-caching request.

tool_call_label 🔒

A short, human-friendly label for a tool call (shown in non-verbose output): prefers the most salient argument (path/pattern) over the raw JSON.

tool_loop

Convenience: a tool loop config rooted at sandbox.