http://programmer.ie/tags/runtime/ Recent content in Runtime on Programmer.ie: Modern AI programming Hugo en-us Wed, 23 Sep 2026 05:00:24 +0100 http://programmer.ie/books/context/24-chapter/ Wed, 23 Sep 2026 05:00:24 +0100 http://programmer.ie/books/context/24-chapter/ <!-- DRAFTING STATUS: PLACEHOLDER Replace this entire comment with the reader-facing chapter. Preserve the TOML frontmatter. ## Chapter problem Individual tricks do not make a context system. A real runtime must observe, classify, preserve, prune, compress, externalize, recover, budget, order, and explain context decisions without losing sight of behaviour. ## Working thesis The capstone is the smallest architecture justified by the book's measurements, not a maximal feature list. It should compile candidate information into a model-ready bundle and preserve an audit trail of every context decision. ## Mechanisms/concepts to earn - Context Lab instrumentation - typed retention policy - deduplication/pruning - compaction - progressive fidelity - external artifacts - just-in-time recall - cache-aware mutation - budget allocation - deterministic assembly - decision trace - evaluation harness ## Proposed book experiment Run the final matched ladder against real OpenCode coding tasks plus a controlled corpus. Require each capstone feature to justify itself through an ablation. Remove any feature that does not earn measurable value. ## Sources/research leads to verify - all prior chapter evidence - DCP, Magic Context, and WithContext as comparative implementations - OpenAI/Anthropic/DeepSeek/Kimi/Qwen runtime and architecture material - the external Context Lab implementation repo once created ## Boundary Implementation code lives outside this book repository. The chapter reports contracts, architecture, experiments, results, limitations, and what the results do not justify. ## Handoff to the next chapter End with the boundary: a context compiler controls what a model gets to know now; retrieval, memory, agents, evaluation, and inference remain distinct systems around it. ## Drafting prompt for OpenCode Draft Chapter 24, "The Context Compiler", for *Context From First Principles*. Before writing: 1. Read `README.md`, `AGENTS.md`, `planning/book-architecture.md`, `planning/capstone-contract.md`, and `research/source-map.md`. 2. Read Chapters 23-24 placeholders/drafts so this chapter has a clean incoming and outgoing boundary. 3. Research the chapter against current primary/first-party sources as of September 2026. Verify titles, dates, model/version claims, and URLs. Prefer papers, official documentation, official model repositories, and source code. Treat vendor benchmark claims as vendor-reported. Treat preprints as preprints. Writing requirements: - Write a strong initial reader-facing chapter, roughly 3,500-5,500 words unless the mechanism genuinely needs less. - British spelling. Concrete engineering prose. No hype, generic AI-history padding, or taxonomy for its own sake. - Begin from a concrete failure/question that makes the mechanism necessary. A coding-agent/OpenCode setting is useful where natural, but do not force every example into the same scenario. - Keep the strict definitions: context is what is made available to the model for this computation; the context window is capacity; available information/session state can exist outside the current context. - Separate established background, book hypotheses, and book results. At this scaffold stage, experiments are proposed unless committed run artifacts exist. Never write an unrun experiment as a result. - Introduce the smallest mechanism that answers the measured or clearly stated failure. Explain trade-offs and failure modes. - Connect to Memory, Embeddings, Retrieval, Agents, Tool Use, Evaluation, or Inference only at the boundary needed here; do not re-teach those books. - Use at most 1-2 tables or Mermaid diagrams, and only when they materially improve understanding. - Do not include capstone implementation code in this repository. Small illustrative pseudocode/functions are acceptable when they clarify the mechanism; the real Context Lab implementation belongs in its separate repository. - Include a clearly labelled proposed experiment or measurement design when evidence is needed. Define controls and measurements before the intervention. - Finish with a short References section containing only sources actually used and verified. - End by making the next step natural: End with the boundary: a context compiler controls what a model gets to know now; retrieval, memory, agents, evaluation, and inference remain distinct systems around it. When the draft is complete, remove this entire HTML comment so no drafting instructions remain in reader-facing prose. -->