[API] Add ML LXM Service API(internal) for large model interactons

[songgot]

This commit introduces the ML LXM Service API, a new C API designed to facilitate interactions with large-scale models such as Large Language Models (LLMs)

[hj210]

This API has a solid foundation and demonstrates clear design goals
and consistent implementation. It is particularly well-aligned with
its specific purpose of interacting with LLMs/LVMs.
With some small refactoring, it could become cleaner and more maintainable code.

[hj210]

This function has a long parameter list. It would be better to reduce the number of parameters by using a structure type if possible.
For example,

typedefstruct {
constchar*config_path;
constchar*instructions;
ml_service_event_cbcallback;
void*user_data;
} ml_lxm_session_config_t;
intml_lxm_session_create(constml_lxm_session_config_t*config, ml_lxm_session_h*session) {
if (!config|| !session|| !config->config_path|| !config->callback)
returnML_ERROR_INVALID_PARAMETER;
returnml_lxm_session_create(config->config_path, config->instructions,
config->callback, config->user_data, session);
}

(This sample code is suggested by Cline)

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR adds an internal "LXM" convenience layer over ml-service: a session (ml_service_new + ml_service_set_event_cb), a prompt (GString), and ml_lxm_session_respond which concatenates instructions + prompt into one tensor and calls ml_service_request. It merges cleanly against current main and CI is green. I read both new files against ml-api-service*.c / ml-api-common.c (destroying input_data right after ml_service_request is safe since the extension clones it), and checked packaging / Android build lists and the companion tests in #647.

1. [Medium] NULL deref + stub — c/src/ml-lxm-service.c:42-47: ml_lxm_check_availability writes *status without a NULL check (ml_lxm_check_availability (NULL) crashes) and always returns ML_LXM_AVAILABILITY_AVAILABLE, making the rest of the enum meaningless. Add the NULL check, and either implement a real check (e.g. check_feature_state (ML_FEATURE_SERVICE) -> SERVICE_DISABLED) or mark it @todo NYI in the header.
2. [Medium] Unsynchronized free/read of instructions — ml-lxm-service.c:209-222 (set_instructions: g_free + g_strdup) vs :250-254 (respond reads and appends). Calling set_instructions from another thread (e.g. from the event callback) while respond runs is a use-after-free. Add a GMutex to ml_lxm_session_internal (init in create, clear in destroy) and hold it in both places, as ml_service_s does.
3. [Medium] No tests in this PR; #647's tests would not run in CI — #647 gates the test binary on dependency('llama', required: false), which no CI workflow, spec, or debian rule provides, and the positive test skips without model files. Please land tests together with this code, and split them so NULL-parameter / missing-config / prompt create-append-destroy / set_instructions / respond negative cases build and run unconditionally under enable-ml-service, with only the token-generating path gated on llama / model availability.
4. [Low] Header docs — c/include/ml-lxm-service-internal.h:257 references ml_lxm_session_set_event_cb(), which does not exist (the callback is passed to ml_lxm_session_create); @example sample_lxm_service.c (line 12) names a file not in the repo; the "complete example" (~60-151) is C++ (<iostream>, static_cast) inside a C header — please make it plain C; enum values (~169-173) lack per-value /**< */ comments unlike every enum in ml-api-service.h.
5. [Low] Prompt buffer has no terminator — ml-lxm-service.c:268-270 sends full_input->len bytes. For flexible inputs the buffer is exactly that size; for static inputs a prompt that exactly fills the tensor has no NUL, and an oversized prompt fails with a bare ML_ERROR_INVALID_PARAMETER. Send len + 1 (or document the contract with the filter) and wrap the failure with _ml_error_report_return_continue.
6. [Low] Style diverges from ml-api-service*.c — no check_feature_state (ML_FEATURE_SERVICE) at entry; bare return ML_ERROR_INVALID_PARAMETER; instead of _ml_error_report_return (...) with descriptive messages; g_malloc0 never returns NULL so the if (!s) / if (!p) branches (~76-79, ~140-141) are dead — use g_new0 (type, 1) as @jaeyun-jung suggested.
7. [Low] ml_lxm_prompt_append_instruction is a byte-identical alias of append_text (:196-200), so "a" + instruction "b" yields "ab" with no separator or role marker. Give it real semantics or drop it for now. Also config_path is stored (:22, :81) but never read, and if (full_input) at :283 is always true.
8. [Low] Build/packaging — java/android/nnstreamer/src/main/jni/Android-nnstreamer.mk:77-81 is not updated (LXM silently absent on Android); the header is neither in the c/include/meson.build install list nor in %files -n capi-machine-learning-tizen-internal-devel in the spec. If "in-tree only for now" is intended, please say so in the PR description.
9. [Low]ml-lxm-service-internal.h:157#include <stdlib.h> is unused; commit title typo "interactons".

No back-door or suspicious behavior found.

[songgot]

This commit introduces the ML LXM Service API, a new C API designed to facilitate interactions with large-scale models such as Large Language Models (LLMs)

[hj210]

This API has a solid foundation and demonstrates clear design goals
and consistent implementation. It is particularly well-aligned with
its specific purpose of interacting with LLMs/LVMs.
With some small refactoring, it could become cleaner and more maintainable code.

[hj210]

This function has a long parameter list. It would be better to reduce the number of parameters by using a structure type if possible.
For example,

typedefstruct {
constchar*config_path;
constchar*instructions;
ml_service_event_cbcallback;
void*user_data;
} ml_lxm_session_config_t;
intml_lxm_session_create(constml_lxm_session_config_t*config, ml_lxm_session_h*session) {
if (!config|| !session|| !config->config_path|| !config->callback)
returnML_ERROR_INVALID_PARAMETER;
returnml_lxm_session_create(config->config_path, config->instructions,
config->callback, config->user_data, session);
}

(This sample code is suggested by Cline)

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR adds an internal "LXM" convenience layer over ml-service: a session (ml_service_new + ml_service_set_event_cb), a prompt (GString), and ml_lxm_session_respond which concatenates instructions + prompt into one tensor and calls ml_service_request. It merges cleanly against current main and CI is green. I read both new files against ml-api-service*.c / ml-api-common.c (destroying input_data right after ml_service_request is safe since the extension clones it), and checked packaging / Android build lists and the companion tests in #647.

1. [Medium] NULL deref + stub — c/src/ml-lxm-service.c:42-47: ml_lxm_check_availability writes *status without a NULL check (ml_lxm_check_availability (NULL) crashes) and always returns ML_LXM_AVAILABILITY_AVAILABLE, making the rest of the enum meaningless. Add the NULL check, and either implement a real check (e.g. check_feature_state (ML_FEATURE_SERVICE) -> SERVICE_DISABLED) or mark it @todo NYI in the header.
2. [Medium] Unsynchronized free/read of instructions — ml-lxm-service.c:209-222 (set_instructions: g_free + g_strdup) vs :250-254 (respond reads and appends). Calling set_instructions from another thread (e.g. from the event callback) while respond runs is a use-after-free. Add a GMutex to ml_lxm_session_internal (init in create, clear in destroy) and hold it in both places, as ml_service_s does.
3. [Medium] No tests in this PR; #647's tests would not run in CI — #647 gates the test binary on dependency('llama', required: false), which no CI workflow, spec, or debian rule provides, and the positive test skips without model files. Please land tests together with this code, and split them so NULL-parameter / missing-config / prompt create-append-destroy / set_instructions / respond negative cases build and run unconditionally under enable-ml-service, with only the token-generating path gated on llama / model availability.
4. [Low] Header docs — c/include/ml-lxm-service-internal.h:257 references ml_lxm_session_set_event_cb(), which does not exist (the callback is passed to ml_lxm_session_create); @example sample_lxm_service.c (line 12) names a file not in the repo; the "complete example" (~60-151) is C++ (<iostream>, static_cast) inside a C header — please make it plain C; enum values (~169-173) lack per-value /**< */ comments unlike every enum in ml-api-service.h.
5. [Low] Prompt buffer has no terminator — ml-lxm-service.c:268-270 sends full_input->len bytes. For flexible inputs the buffer is exactly that size; for static inputs a prompt that exactly fills the tensor has no NUL, and an oversized prompt fails with a bare ML_ERROR_INVALID_PARAMETER. Send len + 1 (or document the contract with the filter) and wrap the failure with _ml_error_report_return_continue.
6. [Low] Style diverges from ml-api-service*.c — no check_feature_state (ML_FEATURE_SERVICE) at entry; bare return ML_ERROR_INVALID_PARAMETER; instead of _ml_error_report_return (...) with descriptive messages; g_malloc0 never returns NULL so the if (!s) / if (!p) branches (~76-79, ~140-141) are dead — use g_new0 (type, 1) as @jaeyun-jung suggested.
7. [Low] ml_lxm_prompt_append_instruction is a byte-identical alias of append_text (:196-200), so "a" + instruction "b" yields "ab" with no separator or role marker. Give it real semantics or drop it for now. Also config_path is stored (:22, :81) but never read, and if (full_input) at :283 is always true.
8. [Low] Build/packaging — java/android/nnstreamer/src/main/jni/Android-nnstreamer.mk:77-81 is not updated (LXM silently absent on Android); the header is neither in the c/include/meson.build install list nor in %files -n capi-machine-learning-tizen-internal-devel in the spec. If "in-tree only for now" is intended, please say so in the PR description.
9. [Low]ml-lxm-service-internal.h:157#include <stdlib.h> is unused; commit title typo "interactons".

No back-door or suspicious behavior found.

[songgot]

This commit introduces the ML LXM Service API, a new C API designed to facilitate interactions with large-scale models such as Large Language Models (LLMs)

[hj210]

This API has a solid foundation and demonstrates clear design goals
and consistent implementation. It is particularly well-aligned with
its specific purpose of interacting with LLMs/LVMs.
With some small refactoring, it could become cleaner and more maintainable code.

[hj210]

This function has a long parameter list. It would be better to reduce the number of parameters by using a structure type if possible.
For example,

typedefstruct {
constchar*config_path;
constchar*instructions;
ml_service_event_cbcallback;
void*user_data;
} ml_lxm_session_config_t;
intml_lxm_session_create(constml_lxm_session_config_t*config, ml_lxm_session_h*session) {
if (!config|| !session|| !config->config_path|| !config->callback)
returnML_ERROR_INVALID_PARAMETER;
returnml_lxm_session_create(config->config_path, config->instructions,
config->callback, config->user_data, session);
}

(This sample code is suggested by Cline)

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR adds an internal "LXM" convenience layer over ml-service: a session (ml_service_new + ml_service_set_event_cb), a prompt (GString), and ml_lxm_session_respond which concatenates instructions + prompt into one tensor and calls ml_service_request. It merges cleanly against current main and CI is green. I read both new files against ml-api-service*.c / ml-api-common.c (destroying input_data right after ml_service_request is safe since the extension clones it), and checked packaging / Android build lists and the companion tests in #647.

1. [Medium] NULL deref + stub — c/src/ml-lxm-service.c:42-47: ml_lxm_check_availability writes *status without a NULL check (ml_lxm_check_availability (NULL) crashes) and always returns ML_LXM_AVAILABILITY_AVAILABLE, making the rest of the enum meaningless. Add the NULL check, and either implement a real check (e.g. check_feature_state (ML_FEATURE_SERVICE) -> SERVICE_DISABLED) or mark it @todo NYI in the header.
2. [Medium] Unsynchronized free/read of instructions — ml-lxm-service.c:209-222 (set_instructions: g_free + g_strdup) vs :250-254 (respond reads and appends). Calling set_instructions from another thread (e.g. from the event callback) while respond runs is a use-after-free. Add a GMutex to ml_lxm_session_internal (init in create, clear in destroy) and hold it in both places, as ml_service_s does.
3. [Medium] No tests in this PR; #647's tests would not run in CI — #647 gates the test binary on dependency('llama', required: false), which no CI workflow, spec, or debian rule provides, and the positive test skips without model files. Please land tests together with this code, and split them so NULL-parameter / missing-config / prompt create-append-destroy / set_instructions / respond negative cases build and run unconditionally under enable-ml-service, with only the token-generating path gated on llama / model availability.
4. [Low] Header docs — c/include/ml-lxm-service-internal.h:257 references ml_lxm_session_set_event_cb(), which does not exist (the callback is passed to ml_lxm_session_create); @example sample_lxm_service.c (line 12) names a file not in the repo; the "complete example" (~60-151) is C++ (<iostream>, static_cast) inside a C header — please make it plain C; enum values (~169-173) lack per-value /**< */ comments unlike every enum in ml-api-service.h.
5. [Low] Prompt buffer has no terminator — ml-lxm-service.c:268-270 sends full_input->len bytes. For flexible inputs the buffer is exactly that size; for static inputs a prompt that exactly fills the tensor has no NUL, and an oversized prompt fails with a bare ML_ERROR_INVALID_PARAMETER. Send len + 1 (or document the contract with the filter) and wrap the failure with _ml_error_report_return_continue.
6. [Low] Style diverges from ml-api-service*.c — no check_feature_state (ML_FEATURE_SERVICE) at entry; bare return ML_ERROR_INVALID_PARAMETER; instead of _ml_error_report_return (...) with descriptive messages; g_malloc0 never returns NULL so the if (!s) / if (!p) branches (~76-79, ~140-141) are dead — use g_new0 (type, 1) as @jaeyun-jung suggested.
7. [Low] ml_lxm_prompt_append_instruction is a byte-identical alias of append_text (:196-200), so "a" + instruction "b" yields "ab" with no separator or role marker. Give it real semantics or drop it for now. Also config_path is stored (:22, :81) but never read, and if (full_input) at :283 is always true.
8. [Low] Build/packaging — java/android/nnstreamer/src/main/jni/Android-nnstreamer.mk:77-81 is not updated (LXM silently absent on Android); the header is neither in the c/include/meson.build install list nor in %files -n capi-machine-learning-tizen-internal-devel in the spec. If "in-tree only for now" is intended, please say so in the PR description.
9. [Low]ml-lxm-service-internal.h:157#include <stdlib.h> is unused; commit title typo "interactons".

No back-door or suspicious behavior found.

[songgot]

This commit introduces the ML LXM Service API, a new C API designed to facilitate interactions with large-scale models such as Large Language Models (LLMs)

[hj210]

This API has a solid foundation and demonstrates clear design goals
and consistent implementation. It is particularly well-aligned with
its specific purpose of interacting with LLMs/LVMs.
With some small refactoring, it could become cleaner and more maintainable code.

[hj210]

This function has a long parameter list. It would be better to reduce the number of parameters by using a structure type if possible.
For example,

typedefstruct {
constchar*config_path;
constchar*instructions;
ml_service_event_cbcallback;
void*user_data;
} ml_lxm_session_config_t;
intml_lxm_session_create(constml_lxm_session_config_t*config, ml_lxm_session_h*session) {
if (!config|| !session|| !config->config_path|| !config->callback)
returnML_ERROR_INVALID_PARAMETER;
returnml_lxm_session_create(config->config_path, config->instructions,
config->callback, config->user_data, session);
}

(This sample code is suggested by Cline)

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR adds an internal "LXM" convenience layer over ml-service: a session (ml_service_new + ml_service_set_event_cb), a prompt (GString), and ml_lxm_session_respond which concatenates instructions + prompt into one tensor and calls ml_service_request. It merges cleanly against current main and CI is green. I read both new files against ml-api-service*.c / ml-api-common.c (destroying input_data right after ml_service_request is safe since the extension clones it), and checked packaging / Android build lists and the companion tests in #647.

1. [Medium] NULL deref + stub — c/src/ml-lxm-service.c:42-47: ml_lxm_check_availability writes *status without a NULL check (ml_lxm_check_availability (NULL) crashes) and always returns ML_LXM_AVAILABILITY_AVAILABLE, making the rest of the enum meaningless. Add the NULL check, and either implement a real check (e.g. check_feature_state (ML_FEATURE_SERVICE) -> SERVICE_DISABLED) or mark it @todo NYI in the header.
2. [Medium] Unsynchronized free/read of instructions — ml-lxm-service.c:209-222 (set_instructions: g_free + g_strdup) vs :250-254 (respond reads and appends). Calling set_instructions from another thread (e.g. from the event callback) while respond runs is a use-after-free. Add a GMutex to ml_lxm_session_internal (init in create, clear in destroy) and hold it in both places, as ml_service_s does.
3. [Medium] No tests in this PR; #647's tests would not run in CI — #647 gates the test binary on dependency('llama', required: false), which no CI workflow, spec, or debian rule provides, and the positive test skips without model files. Please land tests together with this code, and split them so NULL-parameter / missing-config / prompt create-append-destroy / set_instructions / respond negative cases build and run unconditionally under enable-ml-service, with only the token-generating path gated on llama / model availability.
4. [Low] Header docs — c/include/ml-lxm-service-internal.h:257 references ml_lxm_session_set_event_cb(), which does not exist (the callback is passed to ml_lxm_session_create); @example sample_lxm_service.c (line 12) names a file not in the repo; the "complete example" (~60-151) is C++ (<iostream>, static_cast) inside a C header — please make it plain C; enum values (~169-173) lack per-value /**< */ comments unlike every enum in ml-api-service.h.
5. [Low] Prompt buffer has no terminator — ml-lxm-service.c:268-270 sends full_input->len bytes. For flexible inputs the buffer is exactly that size; for static inputs a prompt that exactly fills the tensor has no NUL, and an oversized prompt fails with a bare ML_ERROR_INVALID_PARAMETER. Send len + 1 (or document the contract with the filter) and wrap the failure with _ml_error_report_return_continue.
6. [Low] Style diverges from ml-api-service*.c — no check_feature_state (ML_FEATURE_SERVICE) at entry; bare return ML_ERROR_INVALID_PARAMETER; instead of _ml_error_report_return (...) with descriptive messages; g_malloc0 never returns NULL so the if (!s) / if (!p) branches (~76-79, ~140-141) are dead — use g_new0 (type, 1) as @jaeyun-jung suggested.
7. [Low] ml_lxm_prompt_append_instruction is a byte-identical alias of append_text (:196-200), so "a" + instruction "b" yields "ab" with no separator or role marker. Give it real semantics or drop it for now. Also config_path is stored (:22, :81) but never read, and if (full_input) at :283 is always true.
8. [Low] Build/packaging — java/android/nnstreamer/src/main/jni/Android-nnstreamer.mk:77-81 is not updated (LXM silently absent on Android); the header is neither in the c/include/meson.build install list nor in %files -n capi-machine-learning-tizen-internal-devel in the spec. If "in-tree only for now" is intended, please say so in the PR description.
9. [Low]ml-lxm-service-internal.h:157#include <stdlib.h> is unused; commit title typo "interactons".

No back-door or suspicious behavior found.

[songgot]

This commit introduces the ML LXM Service API, a new C API designed to facilitate interactions with large-scale models such as Large Language Models (LLMs)

[hj210]

This API has a solid foundation and demonstrates clear design goals
and consistent implementation. It is particularly well-aligned with
its specific purpose of interacting with LLMs/LVMs.
With some small refactoring, it could become cleaner and more maintainable code.

[hj210]

This function has a long parameter list. It would be better to reduce the number of parameters by using a structure type if possible.
For example,

typedefstruct {
constchar*config_path;
constchar*instructions;
ml_service_event_cbcallback;
void*user_data;
} ml_lxm_session_config_t;
intml_lxm_session_create(constml_lxm_session_config_t*config, ml_lxm_session_h*session) {
if (!config|| !session|| !config->config_path|| !config->callback)
returnML_ERROR_INVALID_PARAMETER;
returnml_lxm_session_create(config->config_path, config->instructions,
config->callback, config->user_data, session);
}

(This sample code is suggested by Cline)

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR adds an internal "LXM" convenience layer over ml-service: a session (ml_service_new + ml_service_set_event_cb), a prompt (GString), and ml_lxm_session_respond which concatenates instructions + prompt into one tensor and calls ml_service_request. It merges cleanly against current main and CI is green. I read both new files against ml-api-service*.c / ml-api-common.c (destroying input_data right after ml_service_request is safe since the extension clones it), and checked packaging / Android build lists and the companion tests in #647.

1. [Medium] NULL deref + stub — c/src/ml-lxm-service.c:42-47: ml_lxm_check_availability writes *status without a NULL check (ml_lxm_check_availability (NULL) crashes) and always returns ML_LXM_AVAILABILITY_AVAILABLE, making the rest of the enum meaningless. Add the NULL check, and either implement a real check (e.g. check_feature_state (ML_FEATURE_SERVICE) -> SERVICE_DISABLED) or mark it @todo NYI in the header.
2. [Medium] Unsynchronized free/read of instructions — ml-lxm-service.c:209-222 (set_instructions: g_free + g_strdup) vs :250-254 (respond reads and appends). Calling set_instructions from another thread (e.g. from the event callback) while respond runs is a use-after-free. Add a GMutex to ml_lxm_session_internal (init in create, clear in destroy) and hold it in both places, as ml_service_s does.
3. [Medium] No tests in this PR; #647's tests would not run in CI — #647 gates the test binary on dependency('llama', required: false), which no CI workflow, spec, or debian rule provides, and the positive test skips without model files. Please land tests together with this code, and split them so NULL-parameter / missing-config / prompt create-append-destroy / set_instructions / respond negative cases build and run unconditionally under enable-ml-service, with only the token-generating path gated on llama / model availability.
4. [Low] Header docs — c/include/ml-lxm-service-internal.h:257 references ml_lxm_session_set_event_cb(), which does not exist (the callback is passed to ml_lxm_session_create); @example sample_lxm_service.c (line 12) names a file not in the repo; the "complete example" (~60-151) is C++ (<iostream>, static_cast) inside a C header — please make it plain C; enum values (~169-173) lack per-value /**< */ comments unlike every enum in ml-api-service.h.
5. [Low] Prompt buffer has no terminator — ml-lxm-service.c:268-270 sends full_input->len bytes. For flexible inputs the buffer is exactly that size; for static inputs a prompt that exactly fills the tensor has no NUL, and an oversized prompt fails with a bare ML_ERROR_INVALID_PARAMETER. Send len + 1 (or document the contract with the filter) and wrap the failure with _ml_error_report_return_continue.
6. [Low] Style diverges from ml-api-service*.c — no check_feature_state (ML_FEATURE_SERVICE) at entry; bare return ML_ERROR_INVALID_PARAMETER; instead of _ml_error_report_return (...) with descriptive messages; g_malloc0 never returns NULL so the if (!s) / if (!p) branches (~76-79, ~140-141) are dead — use g_new0 (type, 1) as @jaeyun-jung suggested.
7. [Low] ml_lxm_prompt_append_instruction is a byte-identical alias of append_text (:196-200), so "a" + instruction "b" yields "ab" with no separator or role marker. Give it real semantics or drop it for now. Also config_path is stored (:22, :81) but never read, and if (full_input) at :283 is always true.
8. [Low] Build/packaging — java/android/nnstreamer/src/main/jni/Android-nnstreamer.mk:77-81 is not updated (LXM silently absent on Android); the header is neither in the c/include/meson.build install list nor in %files -n capi-machine-learning-tizen-internal-devel in the spec. If "in-tree only for now" is intended, please say so in the PR description.
9. [Low]ml-lxm-service-internal.h:157#include <stdlib.h> is unused; commit title typo "interactons".

No back-door or suspicious behavior found.

[songgot]

This commit introduces the ML LXM Service API, a new C API designed to facilitate interactions with large-scale models such as Large Language Models (LLMs)

[hj210]

This API has a solid foundation and demonstrates clear design goals
and consistent implementation. It is particularly well-aligned with
its specific purpose of interacting with LLMs/LVMs.
With some small refactoring, it could become cleaner and more maintainable code.

[hj210]

This function has a long parameter list. It would be better to reduce the number of parameters by using a structure type if possible.
For example,

typedefstruct {
constchar*config_path;
constchar*instructions;
ml_service_event_cbcallback;
void*user_data;
} ml_lxm_session_config_t;
intml_lxm_session_create(constml_lxm_session_config_t*config, ml_lxm_session_h*session) {
if (!config|| !session|| !config->config_path|| !config->callback)
returnML_ERROR_INVALID_PARAMETER;
returnml_lxm_session_create(config->config_path, config->instructions,
config->callback, config->user_data, session);
}

(This sample code is suggested by Cline)

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR adds an internal "LXM" convenience layer over ml-service: a session (ml_service_new + ml_service_set_event_cb), a prompt (GString), and ml_lxm_session_respond which concatenates instructions + prompt into one tensor and calls ml_service_request. It merges cleanly against current main and CI is green. I read both new files against ml-api-service*.c / ml-api-common.c (destroying input_data right after ml_service_request is safe since the extension clones it), and checked packaging / Android build lists and the companion tests in #647.

1. [Medium] NULL deref + stub — c/src/ml-lxm-service.c:42-47: ml_lxm_check_availability writes *status without a NULL check (ml_lxm_check_availability (NULL) crashes) and always returns ML_LXM_AVAILABILITY_AVAILABLE, making the rest of the enum meaningless. Add the NULL check, and either implement a real check (e.g. check_feature_state (ML_FEATURE_SERVICE) -> SERVICE_DISABLED) or mark it @todo NYI in the header.
2. [Medium] Unsynchronized free/read of instructions — ml-lxm-service.c:209-222 (set_instructions: g_free + g_strdup) vs :250-254 (respond reads and appends). Calling set_instructions from another thread (e.g. from the event callback) while respond runs is a use-after-free. Add a GMutex to ml_lxm_session_internal (init in create, clear in destroy) and hold it in both places, as ml_service_s does.
3. [Medium] No tests in this PR; #647's tests would not run in CI — #647 gates the test binary on dependency('llama', required: false), which no CI workflow, spec, or debian rule provides, and the positive test skips without model files. Please land tests together with this code, and split them so NULL-parameter / missing-config / prompt create-append-destroy / set_instructions / respond negative cases build and run unconditionally under enable-ml-service, with only the token-generating path gated on llama / model availability.
4. [Low] Header docs — c/include/ml-lxm-service-internal.h:257 references ml_lxm_session_set_event_cb(), which does not exist (the callback is passed to ml_lxm_session_create); @example sample_lxm_service.c (line 12) names a file not in the repo; the "complete example" (~60-151) is C++ (<iostream>, static_cast) inside a C header — please make it plain C; enum values (~169-173) lack per-value /**< */ comments unlike every enum in ml-api-service.h.
5. [Low] Prompt buffer has no terminator — ml-lxm-service.c:268-270 sends full_input->len bytes. For flexible inputs the buffer is exactly that size; for static inputs a prompt that exactly fills the tensor has no NUL, and an oversized prompt fails with a bare ML_ERROR_INVALID_PARAMETER. Send len + 1 (or document the contract with the filter) and wrap the failure with _ml_error_report_return_continue.
6. [Low] Style diverges from ml-api-service*.c — no check_feature_state (ML_FEATURE_SERVICE) at entry; bare return ML_ERROR_INVALID_PARAMETER; instead of _ml_error_report_return (...) with descriptive messages; g_malloc0 never returns NULL so the if (!s) / if (!p) branches (~76-79, ~140-141) are dead — use g_new0 (type, 1) as @jaeyun-jung suggested.
7. [Low] ml_lxm_prompt_append_instruction is a byte-identical alias of append_text (:196-200), so "a" + instruction "b" yields "ab" with no separator or role marker. Give it real semantics or drop it for now. Also config_path is stored (:22, :81) but never read, and if (full_input) at :283 is always true.
8. [Low] Build/packaging — java/android/nnstreamer/src/main/jni/Android-nnstreamer.mk:77-81 is not updated (LXM silently absent on Android); the header is neither in the c/include/meson.build install list nor in %files -n capi-machine-learning-tizen-internal-devel in the spec. If "in-tree only for now" is intended, please say so in the PR description.
9. [Low]ml-lxm-service-internal.h:157#include <stdlib.h> is unused; commit title typo "interactons".

No back-door or suspicious behavior found.

[songgot]

This commit introduces the ML LXM Service API, a new C API designed to facilitate interactions with large-scale models such as Large Language Models (LLMs)

[hj210]

This API has a solid foundation and demonstrates clear design goals
and consistent implementation. It is particularly well-aligned with
its specific purpose of interacting with LLMs/LVMs.
With some small refactoring, it could become cleaner and more maintainable code.

[hj210]

This function has a long parameter list. It would be better to reduce the number of parameters by using a structure type if possible.
For example,

typedefstruct {
constchar*config_path;
constchar*instructions;
ml_service_event_cbcallback;
void*user_data;
} ml_lxm_session_config_t;
intml_lxm_session_create(constml_lxm_session_config_t*config, ml_lxm_session_h*session) {
if (!config|| !session|| !config->config_path|| !config->callback)
returnML_ERROR_INVALID_PARAMETER;
returnml_lxm_session_create(config->config_path, config->instructions,
config->callback, config->user_data, session);
}

(This sample code is suggested by Cline)

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR adds an internal "LXM" convenience layer over ml-service: a session (ml_service_new + ml_service_set_event_cb), a prompt (GString), and ml_lxm_session_respond which concatenates instructions + prompt into one tensor and calls ml_service_request. It merges cleanly against current main and CI is green. I read both new files against ml-api-service*.c / ml-api-common.c (destroying input_data right after ml_service_request is safe since the extension clones it), and checked packaging / Android build lists and the companion tests in #647.

1. [Medium] NULL deref + stub — c/src/ml-lxm-service.c:42-47: ml_lxm_check_availability writes *status without a NULL check (ml_lxm_check_availability (NULL) crashes) and always returns ML_LXM_AVAILABILITY_AVAILABLE, making the rest of the enum meaningless. Add the NULL check, and either implement a real check (e.g. check_feature_state (ML_FEATURE_SERVICE) -> SERVICE_DISABLED) or mark it @todo NYI in the header.
2. [Medium] Unsynchronized free/read of instructions — ml-lxm-service.c:209-222 (set_instructions: g_free + g_strdup) vs :250-254 (respond reads and appends). Calling set_instructions from another thread (e.g. from the event callback) while respond runs is a use-after-free. Add a GMutex to ml_lxm_session_internal (init in create, clear in destroy) and hold it in both places, as ml_service_s does.
3. [Medium] No tests in this PR; #647's tests would not run in CI — #647 gates the test binary on dependency('llama', required: false), which no CI workflow, spec, or debian rule provides, and the positive test skips without model files. Please land tests together with this code, and split them so NULL-parameter / missing-config / prompt create-append-destroy / set_instructions / respond negative cases build and run unconditionally under enable-ml-service, with only the token-generating path gated on llama / model availability.
4. [Low] Header docs — c/include/ml-lxm-service-internal.h:257 references ml_lxm_session_set_event_cb(), which does not exist (the callback is passed to ml_lxm_session_create); @example sample_lxm_service.c (line 12) names a file not in the repo; the "complete example" (~60-151) is C++ (<iostream>, static_cast) inside a C header — please make it plain C; enum values (~169-173) lack per-value /**< */ comments unlike every enum in ml-api-service.h.
5. [Low] Prompt buffer has no terminator — ml-lxm-service.c:268-270 sends full_input->len bytes. For flexible inputs the buffer is exactly that size; for static inputs a prompt that exactly fills the tensor has no NUL, and an oversized prompt fails with a bare ML_ERROR_INVALID_PARAMETER. Send len + 1 (or document the contract with the filter) and wrap the failure with _ml_error_report_return_continue.
6. [Low] Style diverges from ml-api-service*.c — no check_feature_state (ML_FEATURE_SERVICE) at entry; bare return ML_ERROR_INVALID_PARAMETER; instead of _ml_error_report_return (...) with descriptive messages; g_malloc0 never returns NULL so the if (!s) / if (!p) branches (~76-79, ~140-141) are dead — use g_new0 (type, 1) as @jaeyun-jung suggested.
7. [Low] ml_lxm_prompt_append_instruction is a byte-identical alias of append_text (:196-200), so "a" + instruction "b" yields "ab" with no separator or role marker. Give it real semantics or drop it for now. Also config_path is stored (:22, :81) but never read, and if (full_input) at :283 is always true.
8. [Low] Build/packaging — java/android/nnstreamer/src/main/jni/Android-nnstreamer.mk:77-81 is not updated (LXM silently absent on Android); the header is neither in the c/include/meson.build install list nor in %files -n capi-machine-learning-tizen-internal-devel in the spec. If "in-tree only for now" is intended, please say so in the PR description.
9. [Low]ml-lxm-service-internal.h:157#include <stdlib.h> is unused; commit title typo "interactons".

No back-door or suspicious behavior found.

[songgot]

This commit introduces the ML LXM Service API, a new C API designed to facilitate interactions with large-scale models such as Large Language Models (LLMs)

[hj210]

This API has a solid foundation and demonstrates clear design goals
and consistent implementation. It is particularly well-aligned with
its specific purpose of interacting with LLMs/LVMs.
With some small refactoring, it could become cleaner and more maintainable code.

[hj210]

This function has a long parameter list. It would be better to reduce the number of parameters by using a structure type if possible.
For example,

typedefstruct {
constchar*config_path;
constchar*instructions;
ml_service_event_cbcallback;
void*user_data;
} ml_lxm_session_config_t;
intml_lxm_session_create(constml_lxm_session_config_t*config, ml_lxm_session_h*session) {
if (!config|| !session|| !config->config_path|| !config->callback)
returnML_ERROR_INVALID_PARAMETER;
returnml_lxm_session_create(config->config_path, config->instructions,
config->callback, config->user_data, session);
}

(This sample code is suggested by Cline)

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR adds an internal "LXM" convenience layer over ml-service: a session (ml_service_new + ml_service_set_event_cb), a prompt (GString), and ml_lxm_session_respond which concatenates instructions + prompt into one tensor and calls ml_service_request. It merges cleanly against current main and CI is green. I read both new files against ml-api-service*.c / ml-api-common.c (destroying input_data right after ml_service_request is safe since the extension clones it), and checked packaging / Android build lists and the companion tests in #647.

1. [Medium] NULL deref + stub — c/src/ml-lxm-service.c:42-47: ml_lxm_check_availability writes *status without a NULL check (ml_lxm_check_availability (NULL) crashes) and always returns ML_LXM_AVAILABILITY_AVAILABLE, making the rest of the enum meaningless. Add the NULL check, and either implement a real check (e.g. check_feature_state (ML_FEATURE_SERVICE) -> SERVICE_DISABLED) or mark it @todo NYI in the header.
2. [Medium] Unsynchronized free/read of instructions — ml-lxm-service.c:209-222 (set_instructions: g_free + g_strdup) vs :250-254 (respond reads and appends). Calling set_instructions from another thread (e.g. from the event callback) while respond runs is a use-after-free. Add a GMutex to ml_lxm_session_internal (init in create, clear in destroy) and hold it in both places, as ml_service_s does.
3. [Medium] No tests in this PR; #647's tests would not run in CI — #647 gates the test binary on dependency('llama', required: false), which no CI workflow, spec, or debian rule provides, and the positive test skips without model files. Please land tests together with this code, and split them so NULL-parameter / missing-config / prompt create-append-destroy / set_instructions / respond negative cases build and run unconditionally under enable-ml-service, with only the token-generating path gated on llama / model availability.
4. [Low] Header docs — c/include/ml-lxm-service-internal.h:257 references ml_lxm_session_set_event_cb(), which does not exist (the callback is passed to ml_lxm_session_create); @example sample_lxm_service.c (line 12) names a file not in the repo; the "complete example" (~60-151) is C++ (<iostream>, static_cast) inside a C header — please make it plain C; enum values (~169-173) lack per-value /**< */ comments unlike every enum in ml-api-service.h.
5. [Low] Prompt buffer has no terminator — ml-lxm-service.c:268-270 sends full_input->len bytes. For flexible inputs the buffer is exactly that size; for static inputs a prompt that exactly fills the tensor has no NUL, and an oversized prompt fails with a bare ML_ERROR_INVALID_PARAMETER. Send len + 1 (or document the contract with the filter) and wrap the failure with _ml_error_report_return_continue.
6. [Low] Style diverges from ml-api-service*.c — no check_feature_state (ML_FEATURE_SERVICE) at entry; bare return ML_ERROR_INVALID_PARAMETER; instead of _ml_error_report_return (...) with descriptive messages; g_malloc0 never returns NULL so the if (!s) / if (!p) branches (~76-79, ~140-141) are dead — use g_new0 (type, 1) as @jaeyun-jung suggested.
7. [Low] ml_lxm_prompt_append_instruction is a byte-identical alias of append_text (:196-200), so "a" + instruction "b" yields "ab" with no separator or role marker. Give it real semantics or drop it for now. Also config_path is stored (:22, :81) but never read, and if (full_input) at :283 is always true.
8. [Low] Build/packaging — java/android/nnstreamer/src/main/jni/Android-nnstreamer.mk:77-81 is not updated (LXM silently absent on Android); the header is neither in the c/include/meson.build install list nor in %files -n capi-machine-learning-tizen-internal-devel in the spec. If "in-tree only for now" is intended, please say so in the PR description.
9. [Low]ml-lxm-service-internal.h:157#include <stdlib.h> is unused; commit title typo "interactons".

No back-door or suspicious behavior found.