common/parser: fix reasoning whitespace bugs + extra parser tests (#21085)

* fix whitespace reasoning issues + add reconstruction tests

* Proper fix

* fix Nemotron autoparser test expectations to include newline in marker

.devops/cpu.Dockerfile

@@ -1,11 +1,13 @@
-ARG UBUNTU_VERSION=22.04
+ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
 ARG TARGETARCH
 
 RUN apt-get update && \
-    apt-get install -y build-essential git cmake libssl-dev
+    apt-get install -y gcc-14 g++-14 build-essential git cmake libssl-dev
+
+ENV CC=gcc-14 CXX=g++-14
 
 WORKDIR /app
 
@@ -55,8 +57,9 @@ RUN apt-get update \
     git \
     python3 \
     python3-pip \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt \
+    python3-wheel \
+    && pip install --break-system-packages --upgrade setuptools \
+    && pip install --break-system-packages -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.github/workflows/docker.yml

@@ -36,18 +36,16 @@ jobs:
       matrix:
         config:
           # Multi-stage build
-          # Note: the arm64 images are failing, which prevents the amd64 images from being built
-          # https://github.com/ggml-org/llama.cpp/issues/11888
-          #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
-          - { tag: "cpu",    dockerfile: ".devops/cpu.Dockerfile",    platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "cuda cuda12", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04", cuda_version: "12.4.0", ubuntu_version: "22.04" }
-          - { tag: "cuda13", dockerfile: ".devops/cuda-new.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04", cuda_version: "13.1.0", ubuntu_version: "24.04" }
-          - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
-          - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
-          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "s390x",  dockerfile: ".devops/s390x.Dockerfile",  platforms: "linux/s390x", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04-s390x" }
-          - { tag: "rocm",   dockerfile: ".devops/rocm.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
-          - { tag: "openvino", dockerfile: ".devops/openvino.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
+          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/arm64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04" }
+          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04" }
+          - { tag: "cuda cuda12", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04", cuda_version: "12.4.0", ubuntu_version: "22.04" }
+          - { tag: "cuda13", dockerfile: ".devops/cuda-new.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04", cuda_version: "13.1.0", ubuntu_version: "24.04" }
+          - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04" }
+          - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04" }
+          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04" }
+          - { tag: "s390x",  dockerfile: ".devops/s390x.Dockerfile",  platforms: "linux/s390x", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04-s390x" }
+          - { tag: "rocm",   dockerfile: ".devops/rocm.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04" }
+          - { tag: "openvino", dockerfile: ".devops/openvino.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v6
@@ -58,7 +56,7 @@ jobs:
         if: ${{ matrix.config.tag != 's390x' }}
         uses: docker/setup-qemu-action@c7c53464625b32c7a7e944ae62b3e17d2b600130 # v3
         with:
-          image: tonistiigi/binfmt:qemu-v7.0.0-28
+          image: tonistiigi/binfmt:qemu-v10.2.1
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3

common/chat-diff-analyzer.cpp

@@ -287,7 +287,7 @@ void analyze_reasoning::compare_reasoning_presence() {
             return p.literal(reasoning_content) + p.space() + p.optional(p.tag("post", (p.marker() + p.space())) + p.rest());
         });
         auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
-            return p.tag("pre", p.marker()) + p.space() + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
+            return p.tag("pre", p.marker() + p.space()) + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
         });
         // try the more aggressive parse first, if it fails, fall back to the delimiter one
         auto result = parser_wrapped.parse_anywhere_and_extract(comparison->output_B);
@@ -297,7 +297,7 @@ void analyze_reasoning::compare_reasoning_presence() {
         if (result.result.success()) {
             if (!result.tags["pre"].empty() && !result.tags["post"].empty()) {
                 mode = reasoning_mode::TAG_BASED;
-                start = trim_whitespace(result.tags["pre"]);
+                start = trim_leading_whitespace(result.tags["pre"]);
                 end   = trim_trailing_whitespace(result.tags["post"]);
             } else if (!result.tags["post"].empty()) {
                 mode = reasoning_mode::TAG_BASED;
@@ -333,7 +333,7 @@ void analyze_reasoning::compare_thinking_enabled() {
     if (left_trimmed.empty() && !diff.right.empty()) {
         if (!right_trimmed.empty() && string_ends_with(comparison->output_B, right_trimmed)) {
             if (start.empty()) {
-                start = right_trimmed;
+                start = trim_leading_whitespace(diff.right);
                 mode  = reasoning_mode::TAG_BASED;
             }
         }
@@ -344,7 +344,7 @@ void analyze_reasoning::compare_thinking_enabled() {
                 if (seg.size() >= 2 && seg[seg.size() - 1].value == left_trimmed && seg[seg.size() - 2].type == segment_type::MARKER) {
                     start = seg[seg.size() - 2].value;
                 }
-                end = left_trimmed;
+                end = trim_trailing_whitespace(diff.left);
                 mode = reasoning_mode::TAG_BASED;
             }
         }
@@ -363,15 +363,23 @@ void analyze_reasoning::compare_thinking_enabled() {
             size_t len = std::min(base.size(), anchor_len);
             std::string anchor = base.substr(base.size() - len);
             auto pos = extended.rfind(anchor);
-            if (pos == std::string::npos || pos + len >= extended.size()) continue;
+            if (pos == std::string::npos || pos + len >= extended.size()) {
+                continue;
+            }
 
             std::string extra = trim_whitespace(extended.substr(pos + len));
-            if (extra.empty()) continue;
+            if (extra.empty()) {
+                continue;
+            }
 
             auto seg = prune_whitespace_segments(segmentize_markers(extra));
             if (seg.size() == 2 && seg[0].type == segment_type::MARKER && seg[1].type == segment_type::MARKER) {
-                if (start.empty()) start = seg[0].value;
-                if (end.empty())   end   = seg[1].value;
+                if (start.empty()) {
+                    start = seg[0].value;
+                }
+                if (end.empty()) {
+                    end   = seg[1].value;
+                }
                 mode = reasoning_mode::TAG_BASED;
                 break;
             }
@@ -423,7 +431,7 @@ void analyze_reasoning::compare_reasoning_scope() {
         LOG_DBG(ANSI_ORANGE "%s: Detected TOOLS_ONLY reasoning mode\n" ANSI_RESET, __func__);
 
         auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
-            return p.tag("pre", p.marker()) + p.space() + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space()));
+            return p.tag("pre", p.marker() + p.space()) + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space()));
         });
         auto result = parser_wrapped.parse_anywhere_and_extract(comparison->output_B);
         if (result.result.success()) {
@@ -516,7 +524,7 @@ analyze_content::analyze_content(const common_chat_template & tmpl, const analyz
         // Take the more promising diff
         std::string pure_content = rdiff.length() > diff_tools.left.length() ? rdiff : diff_tools.left;
         auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
-            return p.tag("pre", p.marker()) + p.space() + p.literal(response) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
+            return p.tag("pre", p.marker() + p.space()) + p.literal(response) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
         });
         auto result = parser_wrapped.parse_anywhere_and_extract(pure_content);
         start = result.tags["pre"];

common/common.cpp

@@ -656,6 +656,38 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
     return true;
 }
 
+// simple glob: * matches non-/ chars, ** matches anything including /
+static inline bool glob_match(const char * pattern, const char * str) {
+    if (*pattern == '\0') {
+        return *str == '\0';
+    }
+    if (pattern[0] == '*' && pattern[1] == '*') {
+        const char * p = pattern + 2;
+        if (*p == '/') p++;
+        if (glob_match(p, str)) return true;
+        if (*str != '\0') return glob_match(pattern, str + 1);
+        return false;
+    }
+    if (*pattern == '*') {
+        const char * p = pattern + 1;
+        for (; *str != '\0' && *str != '/'; str++) {
+            if (glob_match(p, str)) return true;
+        }
+        return glob_match(p, str);
+    }
+    if (*pattern == '?' && *str != '\0' && *str != '/') {
+        return glob_match(pattern + 1, str + 1);
+    }
+    if (*pattern == *str) {
+        return glob_match(pattern + 1, str + 1);
+    }
+    return false;
+}
+
+bool glob_match(const std::string & pattern, const std::string & str) {
+    return glob_match(pattern.c_str(), str.c_str());
+}
+
 //
 // Filesystem utils
 //

common/common.h

@@ -794,6 +794,8 @@ std::string string_from(const std::vector<int> & values);
 std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
 std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
 
+bool glob_match(const std::string & pattern, const std::string & str);
+
 //
 // Filesystem utils
 //

tests/test-chat-auto-parser.cpp

@@ -1330,7 +1330,7 @@ static void test_nemotron_reasoning_detection(testing & t) {
     analysis.analyze_template(tmpl);
 
     // Check reasoning markers
-    t.assert_equal("reasoning_start should be '<think>'", "<think>", analysis.reasoning.start);
+    t.assert_equal("reasoning_start should be '<think>\\n'", "<think>\n", analysis.reasoning.start);
     t.assert_equal("reasoning_end should be '</think>'", "</think>", analysis.reasoning.end);
 
     // Check reasoning mode detection

tests/test-chat.cpp

@@ -805,7 +805,8 @@ struct peg_test_case {
     common_chat_templates_inputs params;
     std::string                  input;
     common_chat_msg              expect;
-    bool                         is_partial = false;
+    bool                         is_partial            = false;
+    bool                         expect_reconstruction = false;
 };
 
 struct make_peg_parser {
@@ -828,6 +829,12 @@ struct make_peg_parser {
     }
 };
 
+// Global template filter for --template flag
+static std::string g_template_filter;
+
+// When true, run reconstruction test on every non-partial test and report results
+static bool g_force_reconstruction_test = false;
+
 static void test_peg_parser(common_chat_templates *                      tmpls,
                             const std::function<void(peg_test_case &)> & init,
                             bool                                         detailed_debug) {
@@ -1119,10 +1126,57 @@ static void test_peg_parser(common_chat_templates *                      tmpls,
             }
         }
     }
-}
 
-// Global template filter for --template flag
-static std::string g_template_filter;
+    // Reconstruction test: verify that appending the parsed message to the original
+    // messages and re-rendering the template (without generation prompt) reproduces
+    // the original prompt + input exactly, or as a proper prefix (the template may
+    // append end-of-turn tokens after the assistant message).
+    if ((tc.expect_reconstruction || g_force_reconstruction_test) && !tc.is_partial) {
+        // Start from tc.expect but copy tool call arguments from the actual parser
+        // output, which preserves original JSON formatting (e.g. {"arg1":1} vs {"arg1": 1}).
+        auto reconstruction_msg = tc.expect;
+        auto parsed_msg         = parser.parse(tc.input, false);
+        for (size_t i = 0; i < reconstruction_msg.tool_calls.size() && i < parsed_msg.tool_calls.size(); i++) {
+            reconstruction_msg.tool_calls[i].arguments = parsed_msg.tool_calls[i].arguments;
+        }
+        common_chat_templates_inputs reconstruction_inputs = tc.params;
+        reconstruction_inputs.messages.push_back(reconstruction_msg);
+        reconstruction_inputs.add_generation_prompt = false;
+
+        auto reconstruction_params = common_chat_templates_apply(tmpls, reconstruction_inputs);
+        std::string expected_text  = parser.params_.prompt + tc.input;
+        bool match = reconstruction_params.prompt == expected_text ||
+            (reconstruction_params.prompt.size() > expected_text.size() &&
+             reconstruction_params.prompt.compare(0, expected_text.size(), expected_text) == 0);
+        if (!match && g_force_reconstruction_test && !tc.expect_reconstruction) {
+            // In forced mode, report mismatch but don't fail
+            // Find the first difference position
+            size_t diff_pos = 0;
+            size_t min_len  = std::min(expected_text.size(), reconstruction_params.prompt.size());
+            while (diff_pos < min_len && expected_text[diff_pos] == reconstruction_params.prompt[diff_pos]) {
+                diff_pos++;
+            }
+            size_t ctx_start = diff_pos > 60 ? diff_pos - 60 : 0;
+            size_t ctx_end_e = std::min(expected_text.size(), diff_pos + 40);
+            size_t ctx_end_r = std::min(reconstruction_params.prompt.size(), diff_pos + 40);
+            LOG_ERR("\x1b[31m[RECONSTRUCTION FAIL]\x1b[0m "
+                    "first diff at byte %zu (expected len=%zu, reconstructed len=%zu)\n"
+                    "  expected:      ...%s...\n"
+                    "  reconstructed: ...%s...\n",
+                    diff_pos, expected_text.size(), reconstruction_params.prompt.size(),
+                    expected_text.substr(ctx_start, ctx_end_e - ctx_start).c_str(),
+                    reconstruction_params.prompt.substr(ctx_start, ctx_end_r - ctx_start).c_str());
+        } else if (!match) {
+            std::string error_msg =
+                "Reconstruction mismatch:\n\n"
+                ">>> Expected (prompt + input):\n" + expected_text +
+                "\n\n>>> Reconstructed:\n" + reconstruction_params.prompt;
+            throw std::runtime_error(error_msg);
+        } else if (g_force_reconstruction_test) {
+            LOG_INF("\x1b[32m[RECONSTRUCTION OK]\x1b[0m\n");
+        }
+    }
+}
 
 // Fluent builder for PEG parser tests
 class peg_test_builder;
@@ -1182,6 +1236,11 @@ class peg_test_builder {
         return *this;
     }
 
+    peg_test_builder & expect_reconstruction(bool val = true) {
+        tc_.expect_reconstruction = val;
+        return *this;
+    }
+
     // Expect setters
     peg_test_builder & expect(const common_chat_msg & msg) {
         tc_.expect = msg;
@@ -1355,16 +1414,18 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         // Ministral-3-14B-Reasoning-2512
         auto tst = peg_tester("models/templates/mistralai-Ministral-3-14B-Reasoning-2512.jinja", detailed_debug);
 
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
 
         tst.test("[THINK]I'm\nthinking[/THINK]Hello, world!\nWhat's up?")
             .expect_content("[THINK]I'm\nthinking[/THINK]Hello, world!\nWhat's up?")
+            .expect_reconstruction()
             .run();
 
         tst.test("[THINK]I'm\nthinking[/THINK]Hello, world!\nWhat's up?")
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .enable_thinking(true)
             .expect(message_assist_thoughts)
+            .expect_reconstruction()
             .run();
 
         tst.test(R"([TOOL_CALLS]special_function[ARGS]{"arg1":1})")
@@ -1394,6 +1455,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
                 { "special_function", R"({"arg1": 1})", {} },
                 { "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
             })
+            .expect_reconstruction()
             .run();
 
         tst.test(
@@ -1418,6 +1480,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "special_function", R"({"arg1": 1})", {} },
             })
+            .expect_reconstruction()
             .run();
     }
 
@@ -1621,9 +1684,9 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         // Google Gemma 2 2B - does not support tool calling
         auto tst = peg_tester("models/templates/google-gemma-2-2b-it.jinja");
 
-        tst.test("Hello, world!").expect(simple_assist_msg("Hello, world!")).run();
+        tst.test("Hello, world!").expect(simple_assist_msg("Hello, world!")).expect_reconstruction().run();
 
-        tst.test("Line 1\nLine 2\nLine 3").expect(simple_assist_msg("Line 1\nLine 2\nLine 3")).run();
+        tst.test("Line 1\nLine 2\nLine 3").expect(simple_assist_msg("Line 1\nLine 2\nLine 3")).expect_reconstruction().run();
     }
 
     {
@@ -1666,7 +1729,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         // Test simple content-only template
         auto tst = peg_tester("models/templates/google-gemma-2-2b-it.jinja", detailed_debug);
 
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
     }
     {
         // IBM Granite (reasoning and tool calling model)
@@ -1778,7 +1841,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         // Qwen3-Coder (tool calling with XML-style format)
         auto tst = peg_tester("models/templates/Qwen3-Coder.jinja", detailed_debug);
 
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
 
         tst.test(
                "<tool_call>\n"
@@ -1790,6 +1853,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
                "</tool_call>")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
 
         tst.test(
@@ -1818,6 +1882,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
                 { "special_function", R"({"arg1": 1})", {} },
                 { "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Test with code content (multiline)
@@ -1838,6 +1903,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "python", "{\"code\": \"def hello():\\n    print(\\\"Hello, world!\\\")\\n\\nhello()\"}", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Test with code content (asian unicode chars)
@@ -1855,6 +1921,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "python", "{\"code\": \"格\"}", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Test with HTML tag content
@@ -1876,6 +1943,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "html", "{\"markup\": \"<html>\\n <head>\\n  <title>Hello!</title>\\n </head>\\n</html>\"}", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Test with TODO list (array of objects)
@@ -1893,6 +1961,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "todo_list", "{\"todos\": [{\"item\": \"Check stuff\", \"selected\": false}, {\"item\": \"Prepare stuff\", \"selected\": true}]}", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Test flexible optional argument ordering (2 required + 4 optional, reversed optional order)
@@ -1909,6 +1978,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "tool_2req_4opt", R"({"req1": "hello", "req2": 42, "opt4": 100, "opt2": 200})", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Test flexible optional argument ordering (2 required + 5 optional, reversed optional order)
@@ -1926,6 +1996,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "tool_2req_5opt", R"({"req1": "world", "req2": 7, "opt5": "last", "opt3": "middle", "opt1": "first"})", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Test flexible optional argument ordering (2 required + 5 optional, all 5 in shuffled order)
@@ -1945,6 +2016,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "tool_2req_5opt", R"({"req1": "test", "req2": 99, "opt3": "c", "opt1": "a", "opt5": "e", "opt4": 4, "opt2": 2})", {} },
             })
+            .expect_reconstruction()
             .run();
     }
     {
@@ -2025,6 +2097,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         tst.test("Hello, world!\nWhat's up?")
             .enable_thinking(false)
             .expect(message_assist)
+            .expect_reconstruction()
             .run();
 
         // Reasoning with content (forced-open mode - input starts after <think>)
@@ -2032,6 +2105,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .enable_thinking(true)
             .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
             .expect(message_assist_thoughts)
+            .expect_reconstruction()
             .run();
 
         // Tool call without reasoning
@@ -2042,6 +2116,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .enable_thinking(false)
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
 
         // Tool call with reasoning (forced-open mode)
@@ -2054,6 +2129,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
             .tools({ special_function_tool })
             .expect(message_assist_call_thoughts)
+            .expect_reconstruction()
             .run();
 
         tst.test(
@@ -2073,6 +2149,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
                 { "special_function", R"({"arg1": 1})", {} },
                 { "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // #20650: tool with no required args, model emits <tool_call>name</tool_call> with no arg tags.
@@ -2090,6 +2167,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
                 .tools({ no_args_tool })
                 .expect_reasoning("Let me read the diff content.")
                 .expect_tool_calls({{ "read_file_diff_md", "{}", {} }})
+                .expect_reconstruction()
                 .run();
         }
     }
@@ -2348,22 +2426,24 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
 
         // Kimi-K2 old template
         auto tst = peg_tester("models/templates/moonshotai-Kimi-K2.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test(
                "<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>"
                "{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>")
             .tools({ special_function_tool })
             .expect(kimi_id_special_func_tool_call)
+            .expect_reconstruction()
             .run();
 
         // Kimi-K2-Instruct
         auto tst2 = peg_tester("models/templates/Kimi-K2-Instruct.jinja", detailed_debug);
-        tst2.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst2.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst2.test(
                "<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>"
                "{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>")
             .tools({ special_function_tool })
             .expect(kimi_id_special_func_tool_call)
+            .expect_reconstruction()
             .run();
     }
 
@@ -2459,6 +2539,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         tst.test("<|tools_prefix|>[{\"special_function\": {\"arg1\": 1}}]<|tools_suffix|>")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
     }
 
@@ -2467,7 +2548,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
     {
         auto tst = peg_tester("models/templates/MiniMax-M2.jinja", detailed_debug);
         tst.test(
-               "</think><minimax:tool_call>\n<invoke name=\"special_function\">\n<parameter "
+               "<minimax:tool_call>\n<invoke name=\"special_function\">\n<parameter "
                "name=\"arg1\">1</parameter>\n</invoke>\n</minimax:tool_call>")
             .tools({ special_function_tool })
             .expect(message_assist_call)
@@ -2517,37 +2598,41 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
     // mistralai-Mistral-Nemo-Instruct-2407.jinja
     {
         auto tst = peg_tester("models/templates/mistralai-Mistral-Nemo-Instruct-2407.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test("[TOOL_CALLS][{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}, \"id\": \"123456789\"}]")
             .tools({ special_function_tool })
             .expect(message_assist_call_id)
+            .expect_reconstruction()
             .run();
     }
     {
         auto tst = peg_tester("models/templates/meetkai-functionary-medium-v3.1.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test("<function=special_function>{\"arg1\": 1}</function>")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
     }
     // Functionary v3.2 - recipient-based format: >>>recipient\n{content}
     {
         auto tst = peg_tester("models/templates/meetkai-functionary-medium-v3.2.jinja", detailed_debug);
-        tst.test("all\nHello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("all\nHello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test("special_function\n{\"arg1\": 1}")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
     }
 
     // FireFunction
     {
         auto tst = peg_tester("models/templates/fireworks-ai-llama-3-firefunction-v2.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test(" functools[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
     }
 
@@ -2608,10 +2693,11 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
          { "models/templates/MiMo-VL.jinja", "models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja",
            "models/templates/Qwen-Qwen2.5-7B-Instruct.jinja" }) {
         auto tst = peg_tester(path, detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test("<tool_call>\n{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n</tool_call>")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
     }
 
@@ -2634,6 +2720,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .enable_thinking(true)
             .expect(simple_assist_msg("Hello, world!\nWhat's up?", "Here are my reasoning steps:\nI'm\nthinking"))
+            .expect_reconstruction()
             .run();
 
         // Reasoning + Tool calls
@@ -2650,42 +2737,45 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
     // Mistral Small 3.2 - FUNC_BRACKET_TAG format: [TOOL_CALLS]func_name[CALL_ID]id[ARGS]{...}
     {
         auto tst = peg_tester("models/templates/Mistral-Small-3.2-24B-Instruct-2506.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test("[TOOL_CALLS]special_function[CALL_ID]123456789[ARGS]{\"arg1\": 1}")
             .tools({ special_function_tool })
             .expect(message_assist_call_id)
+            .expect_reconstruction()
             .run();
     }
     // Devstral
     {
         auto tst = peg_tester("models/templates/unsloth-mistral-Devstral-Small-2507.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test("[TOOL_CALLS]special_function[ARGS]{\"arg1\": 1}")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
         tst.test("Hello, world!\nWhat's up?[TOOL_CALLS]special_function[ARGS]{\"arg1\": 1}")
             .tools({ special_function_tool })
             .expect(message_assist_call_content)
+            .expect_reconstruction()
             .run();
     }
 
     {
         // Llama 3.1
         auto tst = peg_tester("models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").tools({ special_function_tool }).expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").tools({ special_function_tool }).expect(message_assist).expect_reconstruction().run();
     }
 
     {
         // Llama 3.2
         auto tst = peg_tester("models/templates/meta-llama-Llama-3.2-3B-Instruct.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").tools({ special_function_tool }).expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").tools({ special_function_tool }).expect(message_assist).expect_reconstruction().run();
     }
 
     {
         // Llama 3.3
         auto tst = peg_tester("models/templates/meta-llama-Llama-3.3-70B-Instruct.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").tools({ python_tool }).expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").tools({ python_tool }).expect(message_assist).expect_reconstruction().run();
     }
 
     // GPT-OSS format tests
@@ -2989,10 +3079,11 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
     // GigaChat V3
     {
         auto tst = peg_tester("models/templates/GigaChat3-10B-A1.8B.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test("<|message_sep|>\n\nfunction call<|role_sep|>\n{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
 
         tst.test(
@@ -3001,16 +3092,18 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         )
             .tools({ special_function_tool })
             .expect(message_assist_call_content)
+            .expect_reconstruction()
             .run();
     }
 
     // GigaChat V3.1
     {
         auto tst = peg_tester("models/templates/GigaChat3.1-10B-A1.8B.jinja", detailed_debug);
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).expect_reconstruction().run();
         tst.test("<|function_call|>{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}")
             .tools({ special_function_tool })
             .expect(message_assist_call)
+            .expect_reconstruction()
             .run();
 
         tst.test(
@@ -3019,6 +3112,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         )
             .tools({ special_function_tool })
             .expect(message_assist_call_content)
+            .expect_reconstruction()
             .run();
     }
 }
@@ -3155,6 +3249,10 @@ int main(int argc, char ** argv) {
             detailed_debug = true;
             common_log_set_verbosity_thold(999);
         }
+        if (arg == "--force-reconstruction-test") {
+            g_force_reconstruction_test = true;
+            only_run_filtered          = true;
+        }
     }
 
     if (only_run_filtered) {

tools/cli/cli.cpp

@@ -224,10 +224,11 @@ struct cli_context {
 };
 
 // TODO?: Make this reusable, enums, docs
-static const std::array<const std::string, 6> cmds = {
+static const std::array<const std::string, 7> cmds = {
     "/audio ",
     "/clear",
     "/exit",
+    "/glob ",
     "/image ",
     "/read ",
     "/regen",
@@ -258,7 +259,7 @@ static std::vector<std::pair<std::string, size_t>> auto_completion_callback(std:
         }
     }
 
-    if (!cmd.empty() && line.length() >= cmd.length() && cursor_byte_pos >= cmd.length()) {
+    if (!cmd.empty() && cmd != "/glob " && line.length() >= cmd.length() && cursor_byte_pos >= cmd.length()) {
         const std::string path_prefix  = std::string(line.substr(cmd.length(), cursor_byte_pos - cmd.length()));
         const std::string path_postfix = std::string(line.substr(cursor_byte_pos));
         auto cur_dir = std::filesystem::current_path();
@@ -339,6 +340,8 @@ static std::vector<std::pair<std::string, size_t>> auto_completion_callback(std:
     return matches;
 }
 
+static constexpr size_t FILE_GLOB_MAX_RESULTS = 100;
+
 int main(int argc, char ** argv) {
     common_params params;
 
@@ -430,7 +433,8 @@ int main(int argc, char ** argv) {
     console::log("  /exit or Ctrl+C     stop or exit\n");
     console::log("  /regen              regenerate the last response\n");
     console::log("  /clear              clear the chat history\n");
-    console::log("  /read               add a text file\n");
+    console::log("  /read <file>        add a text file\n");
+    console::log("  /glob <pattern>     add text files using globbing pattern\n");
     if (inf.has_inp_image) {
         console::log("  /image <file>       add an image file\n");
     }
@@ -441,6 +445,27 @@ int main(int argc, char ** argv) {
 
     // interactive loop
     std::string cur_msg;
+
+    auto add_text_file = [&](const std::string & fname) -> bool {
+        std::string marker = ctx_cli.load_input_file(fname, false);
+        if (marker.empty()) {
+            console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
+            return false;
+        }
+        if (inf.fim_sep_token != LLAMA_TOKEN_NULL) {
+            cur_msg += common_token_to_piece(ctx_cli.ctx_server.get_llama_context(), inf.fim_sep_token, true);
+            cur_msg += fname;
+            cur_msg.push_back('\n');
+        } else {
+            cur_msg += "--- File: ";
+            cur_msg += fname;
+            cur_msg += " ---\n";
+        }
+        cur_msg += marker;
+        console::log("Loaded text from '%s'\n", fname.c_str());
+        return true;
+    };
+
     while (true) {
         std::string buffer;
         console::set_display(DISPLAY_TYPE_USER_INPUT);
@@ -525,22 +550,60 @@ int main(int argc, char ** argv) {
             continue;
         } else if (string_starts_with(buffer, "/read ")) {
             std::string fname = string_strip(buffer.substr(6));
-            std::string marker = ctx_cli.load_input_file(fname, false);
-            if (marker.empty()) {
-                console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
-                continue;
+            add_text_file(fname);
+            continue;
+        } else if (string_starts_with(buffer, "/glob ")) {
+            std::error_code ec;
+            size_t count = 0;
+            auto curdir = std::filesystem::current_path();
+            std::string pattern = string_strip(buffer.substr(6));
+            std::filesystem::path rel_path;
+
+            auto startglob = pattern.find_first_of("![*?");
+            if (startglob != std::string::npos && startglob != 0) {
+                auto endpath = pattern.substr(0, startglob).find_last_of('/');
+                if (endpath != std::string::npos) {
+                    std::string rel_pattern = pattern.substr(0, endpath);
+#if !defined(_WIN32)
+                    if (string_starts_with(rel_pattern, "~")) {
+                        const char * home = std::getenv("HOME");
+                        if (home && home[0]) {
+                            rel_pattern = std::string(home) + rel_pattern.substr(1);
+                        }
+                    }
+#endif
+                    rel_path = rel_pattern;
+                    pattern.erase(0, endpath + 1);
+                    curdir /= rel_path;
+                }
             }
-            if (inf.fim_sep_token != LLAMA_TOKEN_NULL) {
-                cur_msg += common_token_to_piece(ctx_cli.ctx_server.get_llama_context(), inf.fim_sep_token, true);
-                cur_msg += fname;
-                cur_msg.push_back('\n');
-            } else {
-                cur_msg += "--- File: ";
-                cur_msg += fname;
-                cur_msg += " ---\n";
+
+            for (const auto & entry : std::filesystem::recursive_directory_iterator(curdir,
+                    std::filesystem::directory_options::skip_permission_denied, ec)) {
+                if (!entry.is_regular_file()) {
+                    continue;
+                }
+
+                std::string rel = std::filesystem::relative(entry.path(), curdir, ec).string();
+                if (ec) {
+                    ec.clear();
+                    continue;
+                }
+                std::replace(rel.begin(), rel.end(), '\\', '/');
+
+                if (!glob_match(pattern, rel)) {
+                    continue;
+                }
+
+                if (!add_text_file((rel_path / rel).string())) {
+                    continue;
+                }
+
+                if (++count >= FILE_GLOB_MAX_RESULTS) {
+                    console::error("Maximum number of globbed files allowed (%zu) reached.\n", FILE_GLOB_MAX_RESULTS);
+                    break;
+                }
             }
-            cur_msg += marker;
-            console::log("Loaded text from '%s'\n", fname.c_str());
             continue;
         } else {
             // not a command

tools/server/server-tools.cpp

@@ -101,38 +101,6 @@ static run_proc_result run_process(
     return res;
 }
 
-// simple glob: * matches non-/ chars, ** matches anything including /
-static bool glob_match(const char * pattern, const char * str) {
-    if (*pattern == '\0') {
-        return *str == '\0';
-    }
-    if (pattern[0] == '*' && pattern[1] == '*') {
-        const char * p = pattern + 2;
-        if (*p == '/') p++;
-        if (glob_match(p, str)) return true;
-        if (*str != '\0') return glob_match(pattern, str + 1);
-        return false;
-    }
-    if (*pattern == '*') {
-        const char * p = pattern + 1;
-        for (; *str != '\0' && *str != '/'; str++) {
-            if (glob_match(p, str)) return true;
-        }
-        return glob_match(p, str);
-    }
-    if (*pattern == '?' && *str != '\0' && *str != '/') {
-        return glob_match(pattern + 1, str + 1);
-    }
-    if (*pattern == *str) {
-        return glob_match(pattern + 1, str + 1);
-    }
-    return false;
-}
-
-static bool glob_match(const std::string & pattern, const std::string & str) {
-    return glob_match(pattern.c_str(), str.c_str());
-}
-
 json server_tool::to_json() {
     return {
         {"display_name", display_name},