Latest
A source-level teardown of how five motion-design skills separate vocabulary, principles, review, planning, and execution—and where two important workflow exceptions keep the architecture honest.
CodexBar looks like a tiny macOS utility for AI coding limits. Read at v0.41.0, the real architecture is a curated provider catalog: compile-time provider IDs, descriptors, strategy pipelines, and separate app hooks that turn uneven vendor quota signals into one usage surface.
LazyCodex is marketed as an agent harness for complex codebases. Read at v4.15.1, the source says something narrower and more useful: the root package is a small Codex install alias, the concrete product is the bundled omo plugin, and the deeper harness ownership points back to OmO.
Vector search splits first by where it exists: Qdrant is a dedicated vector engine, pgvector is a PostgreSQL index access method, and LanceDB/Lance is an embedded file-backed stack. That operational home decides filtering, durability, consistency, and failure shape before benchmark tuning starts.
The AI coding agents split by one design choice — how they get your code into the model's context. I read the source of three: Aider ranks a static graph, Continue embeds and retrieves, Cline gives the model tools and lets it read on demand. That single choice decides determinism, token cost, privacy, and how each behaves on a big repo.
Aider is pitched as AI that reads your whole codebase. Read at v0.86.3.dev, the repo map is a ranking problem, not an AI one: tree-sitter pulls the symbols, a PageRank over the reference graph ranks them, and a binary search packs the top ones into a token budget. No embeddings, no vector store, no network call.
Continue is the embedding-first counterpart to Aider. Read at v2.1.0, the answer to the question everyone actually asks — does my proprietary code get shipped to an embeddings API? — is: not by default. It embeds locally with a bundled all-MiniLM-L6-v2, stores vectors in a local LanceDB, and only sends code out if you configure a remote embeddings provider.
LanceDB is a library and a file format, not a server: connect() opens a local path or object-store URI in-process, and only URIs starting with the literal prefix `db` route to the Cloud HTTP client. Storage is the Lance columnar format with append-only manifest versioning, and the index family is IVF-first with PQ compression — not an HNSW graph. Read at lance v8.0.0 / lancedb v0.30.0.
LiteLLM is sold as a smart load balancer across your model deployments. Read at v1.92.0-dev.1, the default is a weighted coin flip: simple-shuffle picks a random healthy deployment, weighted by rpm/tpm if you set it. The intelligence — least-busy, latency, usage-based routing — is opt-in, and it works only because LiteLLM meters every key. On failure it runs a retry-cooldown-fallback ladder.
Everyone assumes Ollama estimates your model's memory and computes how many layers to put on the GPU. Read at v0.31.1, that's not what happens: Ollama is a scheduler that spawns an upstream llama.cpp llama-server, lets it auto-detect the GPU split by default, and reads the offloaded layer count back from its logs. Ollama's own memory math is a deliberately rough proxy used only to decide what fits and what to evict.
pgvector is the 'just use Postgres' answer to vector search, and the contrast with a dedicated engine like Qdrant is the home, not the algorithm. Both are HNSW; pgvector implements it as a Postgres index access method. Read at v0.8.4 — the graph lives in WAL-logged Postgres pages, a search is plain SQL, and filtering is Postgres's job, patched by a new iterative scan.
Adding a metadata filter is where naive vector search breaks — post-filtering wrecks recall, pre-filtering can't use the graph. Read at v1.18.2, Qdrant's answer is two moves: it estimates the filter's cardinality to choose between an exact scan and a graph search, and at build time it adds payload-aware links so a filtered subgraph stays connected — a fix it justifies with percolation theory.
vLLM's throughput comes from two linked ideas, read here at v0.24.0rc2 (the v1 engine). PagedAttention pages each request's KV cache into fixed 16-token blocks from a shared pool, so memory is non-contiguous, tightly packed, and shareable across identical prefixes. Continuous batching rebuilds the batch every forward step from a token budget, mixing prefills and decodes and preempting when blocks run out.
The tools called LLM gateways split by what they route and who they're for. CLIProxyAPI pools your subscription logins. LiteLLM routes your API keys with metering and failover. Portkey is a stateless edge gateway driven by a declarative config. I read the source of all three, and the deciding axis is what state each one keeps.
Self-hosting an open model splits into three camps by the layer you operate at. llama.cpp is the raw engine — a self-contained binary that loads GGUF and offloads layers to whatever hardware you have. Ollama is the friendly manager that wraps it. vLLM is the throughput server that replaces it. I read the source of all three to compare portability, convenience, and throughput.
CLIProxyAPI is sold as an OpenAI-compatible proxy that round-robins across your Claude, ChatGPT/Codex, and Gemini subscriptions. Read the source and it's a reactive cooldown scheduler: it doesn't track your quota, it waits for a 429 — and only for Codex does it actually know when your limit resets.
By question
Each question splits into competing answers. We read the source of every option so the tradeoff is visible before you commit.
The decision axis is context ownership. Pre-compressing code, retrieving from a local index, or letting the model read files changes token cost and predictability.
A coding agent that compresses the codebase through symbol graphs and PageRank
How Aider's Repo Map Actually Reads Your Codebase — PageRank, Not EmbeddingsA coding tool that finds relevant code through local indexing and embedding retrieval
How Continue Indexes Your Codebase — and Whether Your Code Leaves Your MachineA coding agent that gives the model file tools and lets it read the codebase on demand
How AI Coding Agents Feed Your Codebase to the Model: Three CampsThe decision axis is the operating layer. The fork is whether you control the binary engine, use a model manager, or swap engines for server throughput.
A portability-first local inference engine that directly loads GGUF models
Running LLMs Locally: llama.cpp vs Ollama vs vLLM — Three CampsA local LLM manager that wraps the llama.cpp server and manages model lifecycle
How Ollama Loads a Model — and Why It Doesn't Actually Decide Your GPU LayersAn inference server that redesigns KV cache and batching for concurrent request throughput
How vLLM Serves So Many Requests at Once: PagedAttention and Continuous BatchingThe decision axis is state ownership. Whether the product pools login sessions, meters key usage, or only carries config at the edge defines its limits.
An account pooler that places multiple subscription accounts behind an OpenAI-compatible proxy
How CLIProxyAPI Load-Balances Your AI Subscriptions — and What Its Scheduler Reveals About Quota LimitsAn LLM proxy that meters and routes across multiple model deployments and API keys
How LiteLLM's Router Picks a Deployment — Weighted Random by Default, Smart Only If You AskA production-oriented gateway that controls provider routing through declarative config
LLM Gateways: Three Camps — Pool Subscriptions, Route Keys, or Run a Production GatewayThe decision axis is existence. Running a separate engine, living inside Postgres, or opening files from the application process changes consistency, deployment, and filtering strategy.
A dedicated vector server that solves filtered HNSW search inside its own engine
How Qdrant Does Filtered Vector Search: HNSW That Survives Your Payload FiltersA Postgres extension that implements HNSW as an index access method
How pgvector Puts HNSW Inside Postgres — Vector Search as an Index Access MethodA file-format-based vector database opened directly by the application, not a server
How LanceDB Puts a Vector Database in a File — IVF-PQ Over the Lance Columnar Format, No Server