[GSoC 2026] Proposal: Agentic Context Gathering & Evaluation - Md. Shamsul Alam

[shamsulalam1114]

Hi everyone,

My name is Md. Shamsul Alam, and I am highly interested in contributing to the Gemini CLI for GSoC 2026.

I bring a unique dual-background to this project: I am an AI/ML Researcher with published papers and experience building RAG-based LLM applications, while also having extensive full-stack API development experience using TypeScript and Node.js.

I am particularly interested in tackling challenges related to the CLI's reasoning, local context retrieval (RAG), and behavioral evaluation. Because I have built RAG chatbots and web-scraping context pipelines, I understand the friction points of feeding accurate context to an LLM. Furthermore, because I am highly proficient in TypeScript and Node.js, I can seamlessly integrate these AI features directly into the CLI's existing architecture without a steep learning curve.

I am also an active open-source contributor (Openclaw Moltbot) and thrive in complex, ambiguous environments (Bronze Medalist, International Astronomy and Astrophysics Competition).

I am currently exploring the local repository to see how the core agent interacts with the terminal. I would love to hear what the team considers the highest-priority bottleneck regarding the CLI's context-gathering or evaluation capabilities.

Looking forward to diving into the codebase!

[shamsulalam1114]

Update since my initial post — I've submitted PR #21902:

After spending time in the packages/core source, I've implemented a working prototype and submitted it as PR #21902. It includes:

1. ContextEngine service for semantic workspace file retrieval
2. EmbeddingStrategy + VectorStore interfaces (Strategy pattern, swappable)
3. RagHook wired into BeforeModel via HookSystem.registerHook()
4. Behavioral eval framework with LlmJudgeScorer and LikertJudgeScorer
5. 8 unit tests passing, 0 lint errors

One technical question for the team:
In ragHook.ts, I'm casting the HookInput to access llm_request.contents. Should I inject RAG context via hookSpecificOutput.llm_request (using the LLMRequest format from hookTranslator.ts) rather than via systemMessage? I want to make sure the injected context survives applyLLMRequestModifications() correctly.

Would love any architectural feedback before the CI runs!

[shamsulalam1114]

"I also noticed the evals/ directory is missing a tsconfig.json which causes the root build to fail. I've opened issue #21911 and am working on a fix."

[shamsulalam1114]

Progress update since my initial post:

PRs submitted:

#21971 — Added

evals/tsconfig.json
to fix root build failure (split into standalone PR per maintainer request)
#21979 — Fixed 2 TypeScript strict-null errors in integration tests (split into standalone PR per maintainer request)
#21902 — Main feature: RAG context engine + behavioral eval framework
What #21902 includes:

ContextEngine — semantic workspace file indexing and retrieval
EmbeddingStrategy + VectorStore interfaces (swappable via Strategy pattern)
RagHook — injects relevant context into every LLM request via BeforeModel hook
Eval scorers: ToolCallAccuracy, LlmJudge, LikertJudge
11 unit tests passing
Gemini Code Assist review feedback addressed:

Fixed prompt injection vulnerability in formatAsContent() — sanitize file content before embedding in prompt
Fixed score manipulation vulnerability in both judge scorers — sanitize agent responses
Replaced silent catch {} with debugLogger.warn()
Fixed division by zero in

evals/report.ts