AI/ML good first issues

Users are experiencing inconsistent GPU performance on GB10 hardware, with decode speeds fluctuating significantly between fast (63 ms/step) and slow (94 ms/step) states. This performance degradation occurs even when `nvidia-smi` reports stable clock speeds and no throttling. The issue is reproducible with a standalone PyTorch script, suggesting a potential hardware or low-level driver problem.

This issue requires adding unit tests to verify the behavior of session continuity for AI streaming requests. The tests should ensure that the `X-Hermes-Session-Id` header is consistently sent for the same session, omitted when no session ID is provided, and that the Sheets application's schema correctly handles both cases.

This GitHub issue is a survey for students to report which LLM applications they use most frequently. Participants are instructed to create a new Markdown file in a specific directory within the course repository, fill in their GitHub username, and mark up to two LLM apps they use regularly. They then need to submit this file via a pull request to the main branch of the course repository.

The user is reporting a lack of clear feedback in the DSH Desktop panel when changing the provider and model for memory consolidation features. Changes to these settings require a DSH restart to take effect, but the panel does not provide a "restart to apply" notification, leading to confusion when errors occur. Additionally, there's no validation to ensure the selected model is compatible with the chosen provider.

This issue addresses three problems discovered after a fix for autoDream's tolerant path handling. The first problem involves `dreamMinExplicitCoverage` incorrectly rejecting weak models that intentionally provide less explicit output. The second issue concerns the `archive` decision type being used inappropriately for items that should be retained long-term. Finally, the third problem highlights a lack of detailed logging for failed autoDream runs, making it difficult to diagnose the cause of rejection.

This issue aims to implement a user interface feature that allows learners to visualize tokenization of non-Latin languages (Hindi, Tamil, CJK) within the existing tokenizer. It involves adding example prompts for these languages and ensuring the byte-fallback fragmentation is correctly displayed, differentiating it from English tokenization.

The user is requesting support for Linux, specifically for use on a platform called 'Xiangong'. This is an enhancement request, indicating a desire for new functionality rather than a bug fix.

The model picker incorrectly displays deprecated OpenAI models that result in a 404 error when used. This issue arises because the system builds the picker directly from an API response that still lists these defunct models. The proposed solutions involve either a deny-list, probing models to cache their availability, or marking unavailable models in the UI.

This issue details a successful port of the t8 project to AMD Radeon GPUs on Windows using ROCm. All four core capabilities, including music generation, audio-to-sheet conversion, score rendering, and voice cloning, have been verified to produce real outputs. The port was achieved with minimal code changes, primarily by leveraging ROCm's HIP backend and ensuring compatibility with existing PyTorch models.

This issue proposes enhancing the graph functionality from a passive display to an active retrieval signal for LLM reasoning. It outlines a four-part plan: anchoring activation through multiple normalized pathways, evolving edge weights based on usage and decay, implementing graded injection budgets with hallucination prevention, and integrating passive confirmation into existing touch mechanisms. This feature is planned for v0.8 after the v0.7 workspace and heat features are stabilized.

The user wants a way to temporarily disable the memory/history feature, similar to incognito mode in browsers. They are looking for a direct option in the UI, like in ChatGPT, to avoid saving conversation history for specific test cases.

This issue proposes adding read-only endpoints to the standalone API to expose user profiles and rules. Currently, external tools cannot access this information, leading to workarounds. The proposed solution involves adding GET routes for '/api/external/.../profile' and '/api/external/.../rules', returning empty values instead of 404 errors for non-existent data, and using existing authentication mechanisms.

This feature request aims to improve user experience by proactively informing users when the 'minimal' preset mode disables runtime context injections (memory, user profiles, etc.). The panel should display a clear message and a link to solutions when this mode is active, preventing silent suppression of these features.

This feature aims to make the dsh-mneme memory capabilities available to any MCP client, decoupling it from the DSH host. The initial version (v1) will implement a stdio MCP server that communicates via a standalone API, providing a set of six memory management tools with zero migration cost for existing DSH users. This is a significant step towards generalizing the plugin.

This issue proposes adding an observability feature to allow users to see what is currently being injected into prompts. This includes a preview of injected content, selection reasons, and volume, as well as explicit notifications when injection is disabled due to host presets like 'minimal'. The goal is to improve transparency and debugging for users.

This issue addresses accessibility improvements for a dashboard, focusing on WCAG compliance. Key areas for enhancement include implementing aria-live for status updates, improving context for screen reader users, adding semantic structure to status cards, managing focus within modals, providing alternative text and summaries for charts, and auditing color contrast. The proposed solutions are broken down into four manageable sub-tasks.

This issue proposes a significant redesign of the conflict queue (status page) UI and interaction. The goal is to make it more user-friendly and understandable for ordinary users, moving away from a developer-centric view. Key improvements include visual enhancements like progress bars for similarity, color-coding for A/B content, word-level diff highlighting, and clearer explanations for 'archived' states. The structure will also be improved with a clickable status card and educational content for empty states.

This issue requests the creation of a new shell script runner, `evals/agents/cursor.sh`, to integrate the Cursor CLI into the evaluation framework. This runner will translate Cursor's output into a format compatible with the existing scorer, allowing for the measurement of Cursor's agent activation performance.

This issue proposes creating a new shell script runner for the Codex CLI to enable its integration into the evaluation framework. This runner will translate Codex CLI commands into a format that the evaluation scorer can understand, specifically by counting the number of times the agent activates its skills. The goal is to measure the activation of installed packs on Codex, similar to how it's done for Claude Code.

This issue requests the creation of a new executable script, `gemini-cli.sh`, to integrate the Gemini CLI with an existing evaluation framework. This script will act as a runner, translating Gemini CLI's headless execution into a format the scorer can read, specifically focusing on measuring the activation of installed skills.

This feature request addresses issues with the automatic scope tagging of memories in Mneme v0.8.0. The current system over-assigns scopes based on the writing agent's "snapshot" and lacks a manual way to correct or promote memories that should be global. The user proposes adding manual scope editing and improving the system's ability to identify and correct mis-scoped memories.

This issue proposes adding a new 'upsert-node' operation to the document editing system. This operation will allow agents to either add a new node if it doesn't exist or update an existing node if it does, simplifying logic and enabling idempotent retries. The change involves adding the operation definition, a corresponding test, and updating documentation.

This issue proposes adding a 'pose' mode to an existing system, allowing users to generate skeleton outputs alongside depth information. This would provide an alternative control signal for video generation, particularly useful for dance and fight clips, by leveraging a commercially usable pose estimation model. The implementation would involve CLI and web UI changes, a new backend, and comprehensive testing.

This issue requests the translation of the project's README file into Japanese and Korean to reach new communities. The task involves creating new README files for each language, maintaining existing content like images and code blocks, and updating the language links in the English and Chinese READMEs. The translations should be done carefully, especially for specific technical sections.

This issue requests the addition of a new "target preset" for an unspecified video model within the reshot project. This involves finding the model's official specifications for control/reference videos (FPS, size rounding, max length), adding a corresponding entry in `reshot/targets.py`, updating the web interface and documentation, and writing a test to verify the functionality.

This issue is a "good first issue" to add reproducible tests for a lightweight keyword search script (`scripts/search_corpus.py`). The goal is to create tests using only temporary, fictional data to ensure changes can be safely verified without relying on external resources or real data. The tests should cover various search functionalities like keyword matching, filtering, and limiting results.

This issue proposes implementing end-to-end integration tests for the FastAPI application. The goal is to test the complete request flow, from uploads to analysis, within the CI environment using a test double or a small cached model. This will replace the current nightly-only manual verification for the happy path.

This issue proposes to refactor the project's dependencies by separating GGUF-specific requirements into a new file (`requirements-gguf.txt`). The goal is to keep the base requirements clean and ensure the application can still boot and function with only GGUF metadata when these optional dependencies are not installed.

This issue proposes a visual enhancement to the UI's Timeline panel. It requires distinguishing between 'real' (instrumented) and 'proxy' (estimated) timings for per-layer performance data by adding distinct visual cues like badges or legends. The goal is to make it clearer to users which timing source they are observing.

This issue involves auditing and removing unused local variables and potentially entire React components within the `frontend/components/scenes/` directory. The goal is to clean up the codebase by pruning dead code, with a specific caution to avoid deleting components referenced by visual regression tests or the `TransformerOperationGraph`.