File size: 6,296 Bytes
4880083 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 | diff --git a/common/chat.cpp b/common/chat.cpp
index 9639af9..971e6d6 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -1979,6 +1979,109 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
return data;
}
+// Cohere Command (Cohere2 / command-a) tool-call + reasoning format:
+// <|START_THINKING|>...plan...<|END_THINKING|>
+// <|START_ACTION|>[{"tool_call_id":"0","tool_name":"X","parameters":{...}}]<|END_ACTION|>
+// <|START_RESPONSE|>...text...<|END_RESPONSE|>
+// The action block is a JSON array of objects keyed by tool_name/parameters, with a
+// generated integer tool_call_id, which maps directly onto standard_json_tools().
+static common_chat_params common_chat_params_init_cohere2(const common_chat_template & tmpl,
+ const autoparser::generation_params & inputs) {
+ common_chat_params data;
+
+ data.prompt = common_chat_template_direct_apply_impl(tmpl, inputs);
+ data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+ data.format = COMMON_CHAT_FORMAT_PEG_NATIVE;
+ data.supports_thinking = true;
+ data.thinking_start_tag = "<|START_THINKING|>";
+ data.thinking_end_tag = "<|END_THINKING|>";
+ data.preserved_tokens = {
+ "<|START_THINKING|>", "<|END_THINKING|>",
+ "<|START_ACTION|>", "<|END_ACTION|>",
+ "<|START_RESPONSE|>", "<|END_RESPONSE|>",
+ };
+
+ const std::string THINK_START = "<|START_THINKING|>";
+ const std::string THINK_END = "<|END_THINKING|>";
+ const std::string ACTION_START = "<|START_ACTION|>";
+ const std::string ACTION_END = "<|END_ACTION|>";
+ const std::string RESP_START = "<|START_RESPONSE|>";
+ const std::string RESP_END = "<|END_RESPONSE|>";
+
+ const bool has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+ const bool extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+ const bool include_grammar = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+ // With tools the template leaves generation open after <|CHATBOT_TOKEN|>; without tools
+ // it primes <|START_RESPONSE|> directly.
+ const bool response_primed =
+ data.generation_prompt.size() >= RESP_START.size() &&
+ data.generation_prompt.compare(data.generation_prompt.size() - RESP_START.size(),
+ RESP_START.size(), RESP_START) == 0;
+
+ auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+ auto gen = p.literal(data.generation_prompt);
+ auto end = p.end();
+
+ if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+ if (response_primed) {
+ return gen + p.content(p.until(RESP_END)) + p.literal(RESP_END) + end;
+ }
+ return gen + p.choice({
+ p.literal(RESP_START) + p.content(p.until(RESP_END)) + p.literal(RESP_END),
+ p.content(p.rest()),
+ }) + end;
+ }
+
+ auto reasoning = p.eps();
+ if (extract_reasoning && inputs.enable_thinking) {
+ reasoning = p.optional(p.literal(THINK_START) + p.reasoning(p.until(THINK_END)) + p.literal(THINK_END));
+ } else if (extract_reasoning) {
+ reasoning = p.optional(p.literal(THINK_START) + p.until(THINK_END) + p.literal(THINK_END));
+ }
+
+ auto tools_parser = p.standard_json_tools(
+ ACTION_START, ACTION_END, inputs.tools,
+ /* parallel_tool_calls = */ true,
+ /* force_tool_calls = */ inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED,
+ /* name_key = */ "tool_name",
+ /* args_key = */ "parameters",
+ /* array_wrapped = */ true,
+ /* function_is_key = */ false,
+ /* call_id_key = */ "",
+ /* gen_call_id_key = */ "tool_call_id",
+ /* parameters_order = */ {});
+
+ if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
+ return gen + reasoning + p.space() + tools_parser + end;
+ }
+ // Content-first, then optional tool action (avoids non-backtracking choice on the
+ // tool-section trigger). The response may be <|START_RESPONSE|>-wrapped or bare;
+ // strip a leading wrapper token and let content run up to any tool action.
+ auto content_before = p.optional(p.literal(RESP_START)) + p.content(p.until(ACTION_START));
+ return gen + reasoning + p.space() + content_before + p.optional(tools_parser) + end;
+ });
+
+ data.parser = parser.save();
+
+ if (include_grammar) {
+ data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+ data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+ foreach_function(inputs.tools, [&](const json & tool) {
+ const auto & function = tool.at("function");
+ auto schema = function.contains("parameters") ? function.at("parameters") : json::object();
+ builder.resolve_refs(schema);
+ });
+ parser.build_grammar(builder, data.grammar_lazy);
+ });
+ data.grammar_triggers = {
+ { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, ACTION_START },
+ };
+ }
+
+ return data;
+}
+
namespace workaround {
static void map_developer_role_to_system(json & messages) {
@@ -2256,6 +2359,13 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
return common_chat_params_init_deepseek_v3_2(tmpl, params);
}
+ // Cohere Command (Cohere2 / command-a) format detection: thinking + action-array tool calls.
+ if (src.find("<|START_ACTION|>") != std::string::npos &&
+ src.find("<|START_THINKING|>") != std::string::npos &&
+ src.find("tool_name") != std::string::npos) {
+ return common_chat_params_init_cohere2(tmpl, params);
+ }
+
// Gemma4 format detection
if (src.find("'<|tool_call>call:'") != std::string::npos) {
if (src.find("{#- OpenAI Chat Completions:") == std::string::npos) {
|