[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AFZPBKBJ68DMHF0CFK5CG7":3,"skill-same:01M7AFZPBKBJ68DMHF0CFK5CG7":46},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7AFZPBKBJ68DMHF0CFK5CG7","0.4.1-understand-context_window",null,"[0.4.1] The finite token sequence an LLM can process in a single forward pass; defines the maximum input + output capa","sancovp\u002Fdark-factory-live","sancovp","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F147458096?v=4","01M74F2EFE4Q0KF1Y5AY27HVNB","Isaac Wostrel-Rubin","01M7AC3EB8VZXS1SKHQW1EZ4N1",0,"MIT","https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live\u002Fblob\u002Fmain\u002Fkbworld\u002Fstate\u002Fmodules\u002Fllm_memory_and_context_engineering\u002Fskills\u002F0-4-1-understand-context-window\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: 0.4.1-understand-context_window\ndescription: \"[0.4.1] The finite token sequence an LLM can process in a single forward pass; defines the maximum input + output capa\"\n---\n\n# understand-context_window\n\n**CALL NUMBER:** `llm_memory_and_context_engineering.context_window : deep_context_window(13)`\n**DEFINITION:** The finite token sequence an LLM can process in a single forward pass; defines the maximum input + output capacity bounded by architecture or inference config\n\nInvoke this skill to understand `context_window` down to its primitives. The RELATIVE ROOT below is the least-fixed-point closure of everything it bundles from — the full import cone, grouped by the lib each prim comes from. Projected from a prover-typed KB (MAP\u002FSWI-Prolog consistency gate): every reference below resolves.\n\n## THE RELATIVE ROOT (the import cone, by lib)\n\n### from `deep_context_window`\n- **context_boundary** (d2): The hard edge position within context where context_overflow triggers if exceeded; exactly context_length minus one for the output region.\n- **context_resize_event** (d2): The runtime mechanism or policy invoked when context_length must accommodate inputs exceeding current capacity; involves reallocation or overflow handling.\n- **context_scaling_factor** (d2): The multiplier relating raw context_length to effective_context; models with improved attention patterns achieve factors closer to one.\n- **input_token_budget** (d2): The portion of context_length reserved and available for prompt tokens, retrieved context, and system instructions before generation begins.\n- **max_model_tokens** (d2): The absolute token ceiling of the model architecture; synonymous with context_length when no additional constraints apply.\n- **maximum_position** (d2): The highest valid integer position index representable by the position_embedding scheme within context_length; equal to max_model_tokens minus one for zero-indexed sequences.\n- **output_token_limit** (d2): The portion of context_length reserved for generated tokens; equals context_length minus input_token_budget minus overhead tokens.\n- **overflow_threshold** (d2): The position or token count at which the model behavior shifts from normal processing to truncation or error; often slightly below context_boundary.\n- **token_boundary_marker** (d2): The positional delimiter marking the transition from input context to output generation; tokens before this marker are provided context, tokens after are to be generated.\n- **context_capacity_ratio** (d3): The ratio of tokens actively processed to total available context_length; quantifies context_utilization as used_capacity divided by max_model_tokens.\n- **token_headroom** (d3): The arithmetic difference between current position or token count and context_boundary; the remaining room before overflow_threshold is reached.\n- **position_index** (d3): Ordinal integer from zero to context_window_boundary representing each token's location in the sequence; each index must be encodable by the position_embedding scheme in use\n- **sequence_position** (d3): The ordinal index of a token within the full context sequence; constrained to be less than or equal to maximum_position.\n\n### from `llm_memory_and_context_engineering`\n- **context_length** (d1): Total number of tokens (input + output) an LLM supports; the hard ceiling on how much text can be loaded into context_window at once\n- **context_overflow** (d1): The failure state where total input exceeds context_window, causing silent truncation or errors depending on the client library\n- **context_utilization** (d1): How effectively the model uses all available context tokens; metrics include attention entropy, retrieval accuracy at various positions, task performance variation with position.\n- **effective_context** (d1): Portion of context_window the model actually leverages for any given generation; often less than the full window due to attention patterns\n- **flash_attention** (d1): A memory-efficient attention algorithm using tiling and recomputation to avoid materializing the full n×n attention matrix in HBM\n- **kv_cache** (d1): The key-value tensors cached during transformer self-attention to avoid recomputing activations for previously-seen tokens\n- **position_embedding** (d1): Encoding that injects positional information into token representations since attention is permutation-invariant. Choices: sinusoidal, RoPE, ALiBi. Affects how context scales.\n- **ring_attention** (d1): Distributed sliding_window_attention across multiple devices with communication-avoiding algorithms for very long sequences\n- **sliding_window** (d1): Fixed-size contiguous subset of recent tokens kept in working_memory as new tokens arrive and old ones are evicted\n- **truncation_strategy** (d1): Policy for removing old context when context_overflow occurs; options include FIFO, importance-weighted, and semantic summarization\n- **working_memory** (d1): The transient token budget allocated for active reasoning within a single model call; scratch space for chain_of_thought and intermediate results\n- **context_padding** (d2): Appending empty tokens to extend context to alignment boundaries; inefficient but sometimes necessary for batch processing\n- **token_accounting** (d2): Tracking token usage across context_building operations to ensure neither context_window nor max_tokens limits are exceeded\n- **context_caching** (d2): Storing computed key-value pairs from prior forward passes to avoid recomputation on repeated or overlapping context prefixes\n- **prefix_caching** (d2): Caching the KV tensors for shared prefixes (system prompt, few-shot examples) so they aren't recomputed per request; reduces latency and compute for repeated contexts.\n- **backward_recall** (d2): Retrieval of past context tokens; constrained by context_window for working_memory and retrieval_mechanism for long_term_memory\n- **context_horizon** (d2): Farthest retrievable or presentable token distance from current position; the effective reach of memory_access\n- **memory_window** (d2): The subset of episodic_memory available for retrieval at any time; analogous to context_window but for long_term_storage\n- **attention_mechanism** (d2): The core operation allowing each token to attend to all others; scaled dot-product attention is the standard. Memory and context are both outputs of attention computation.\n- **chain_of_thought** (d2): Producing intermediate reasoning steps before final answer; when demonstrated in context, subsequent reasoning follows the pattern. A form of contextual procedural memory.\n- **memory_efficiency** (d2): Ratio of useful information retained per unit of memory storage or context_token budget\n- **memory_hierarchy** (d2): The layered structure of LLM memory: immediate context (working), retrieved context (RAG), fine-tuned weights (parametric), external stores (vector DB). Each layer has different latency\u002Fcapacity tradeoffs.\n- **scratchpad** (d2): Explicit intermediate reasoning space written by the model itself; a form of ephemeral working_memory externalized to tokens\n- **attention_saturation** (d3): Condition where attention weights become uniform across long context; signals reduced context_utilization and effective_context shrinkage\n- **attention_sinks** (d3): Special tokens that accumulate disproportionate attention mass, enabling efficient long-context modeling without explicit positional encoding\n\n## CONSUMERS (what needs this)\n`attention_sinks`, `in_context_learning`, `kv_cache`, `working_memory`\n\n---\n*Projected from the `llm memory and context engineering` KB (438 concepts \u002F 213 relations) — consistency-typed by MAP; the facet list after the colon IS the cross-lib dependency web.*\n\n_(leaf — this is an actual skill.)_\n","# understand-context_window\n\n**CALL NUMBER:** `llm_memory_and_context_engineering.context_window : deep_context_window(13)`\n**DEFINITION:** The finite token sequence an LLM can process in a single forward pass; defines the maximum input + output capacity bounded by architecture or inference config\n\nInvoke this skill to understand `context_window` down to its primitives. The RELATIVE ROOT below is the least-fixed-point closure of everything it bundles from — the full import cone, grouped by the lib each prim comes from. Projected from a prover-typed KB (MAP\u002FSWI-Prolog consistency gate): every reference below resolves.\n\n## THE RELATIVE ROOT (the import cone, by lib)\n\n### from `deep_context_window`\n- **context_boundary** (d2): The hard edge position within context where context_overflow triggers if exceeded; exactly context_length minus one for the output region.\n- **context_resize_event** (d2): The runtime mechanism or policy invoked when context_length must accommodate inputs exceeding current capacity; involves reallocation or overflow handling.\n- **context_scaling_factor** (d2): The multiplier relating raw context_length to effective_context; models with improved attention patterns achieve factors closer to one.\n- **input_token_budget** (d2): The portion of context_length reserved and available for prompt tokens, retrieved context, and system instructions before generation begins.\n- **max_model_tokens** (d2): The absolute token ceiling of the model architecture; synonymous with context_length when no additional constraints apply.\n- **maximum_position** (d2): The highest valid integer position index representable by the position_embedding scheme within context_length; equal to max_model_tokens minus one for zero-indexed sequences.\n- **output_token_limit** (d2): The portion of context_length reserved for generated tokens; equals context_length minus input_token_budget minus overhead tokens.\n- **overflow_threshold** (d2): The position or token count at which the model behavior shifts from normal processing to truncation or error; often slightly below context_boundary.\n- **token_boundary_marker** (d2): The positional delimiter marking the transition from input context to output generation; tokens before this marker are provided context, tokens after are to be generated.\n- **context_capacity_ratio** (d3): The ratio of tokens actively processed to total available context_length; quantifies context_utilization as used_capacity divided by max_model_tokens.\n- **token_headroom** (d3): The arithmetic difference between current position or token count and context_boundary; the remaining room before overflow_threshold is reached.\n- **position_index** (d3): Ordinal integer from zero to context_window_boundary representing each token's location in the sequence; each index must be encodable by the position_embedding scheme in use\n- **sequence_position** (d3): The ordinal index of a token within the full context sequence; constrained to be less than or equal to maximum_position.\n\n### from `llm_memory_and_context_engineering`\n- **context_length** (d1): Total number of tokens (input + output) an LLM supports; the hard ceiling on how much text can be loaded into context_window at once\n- **context_overflow** (d1): The failure state where total input exceeds context_window, causing silent truncation or errors depending on the client library\n- **context_utilization** (d1): How effectively the model uses all available context tokens; metrics include attention entropy, retrieval accuracy at various positions, task performance variation with position.\n- **effective_context** (d1): Portion of context_window the model actually leverages for any given generation; often less than the full window due to attention patterns\n- **flash_attention** (d1): A memory-efficient attention algorithm using tiling and recomputation to avoid materializing the full n×n attention matrix in HBM\n- **kv_cache** (d1): The key-value tensors cached during transformer self-attention to avoid recomputing activations for previously-seen tokens\n- **position_embedding** (d1): Encoding that injects positional information into token representations since attention is permutation-invariant. Choices: sinusoidal, RoPE, ALiBi. Affects how context scales.\n- **ring_attention** (d1): Distributed sliding_window_attention across multiple devices with communication-avoiding algorithms for very long sequences\n- **sliding_window** (d1): Fixed-size contiguous subset of recent tokens kept in working_memory as new tokens arrive and old ones are evicted\n- **truncation_strategy** (d1): Policy for removing old context when context_overflow occurs; options include FIFO, importance-weighted, and semantic summarization\n- **working_memory** (d1): The transient token budget allocated for active reasoning within a single model call; scratch space for chain_of_thought and intermediate results\n- **context_padding** (d2): Appending empty tokens to extend context to alignment boundaries; inefficient but sometimes necessary for batch processing\n- **token_accounting** (d2): Tracking token usage across context_building operations to ensure neither context_window nor max_tokens limits are exceeded\n- **context_caching** (d2): Storing computed key-value pairs from prior forward passes to avoid recomputation on repeated or overlapping context prefixes\n- **prefix_caching** (d2): Caching the KV tensors for shared prefixes (system prompt, few-shot examples) so they aren't recomputed per request; reduces latency and compute for repeated contexts.\n- **backward_recall** (d2): Retrieval of past context tokens; constrained by context_window for working_memory and retrieval_mechanism for long_term_memory\n- **context_horizon** (d2): Farthest retrievable or presentable token distance from current position; the effective reach of memory_access\n- **memory_window** (d2): The subset of episodic_memory available for retrieval at any time; analogous to context_window but for long_term_storage\n- **attention_mechanism** (d2): The core operation allowing each token to attend to all others; scaled dot-product attention is the standard. Memory and context are both outputs of attention computation.\n- **chain_of_thought** (d2): Producing intermediate reasoning steps before final answer; when demonstrated in context, subsequent reasoning follows the pattern. A form of contextual procedural memory.\n- **memory_efficiency** (d2): Ratio of useful information retained per unit of memory storage or context_token budget\n- **memory_hierarchy** (d2): The layered structure of LLM memory: immediate context (working), retrieved context (RAG), fine-tuned weights (parametric), external stores (vector DB). Each layer has different latency\u002Fcapacity tradeoffs.\n- **scratchpad** (d2): Explicit intermediate reasoning space written by the model itself; a form of ephemeral working_memory externalized to tokens\n- **attention_saturation** (d3): Condition where attention weights become uniform across long context; signals reduced context_utilization and effective_context shrinkage\n- **attention_sinks** (d3): Special tokens that accumulate disproportionate attention mass, enabling efficient long-context modeling without explicit positional encoding\n\n## CONSUMERS (what needs this)\n`attention_sinks`, `in_context_learning`, `kv_cache`, `working_memory`\n\n---\n*Projected from the `llm memory and context engineering` KB (438 concepts \u002F 213 relations) — consistency-typed by MAP; the facet list after the colon IS the cross-lib dependency web.*\n\n_(leaf — this is an actual skill.)_\n",{"name":5,"description":7},"main","dfed66814f223a9f42bec375a3c59d74eefdcad5","kbworld\u002Fstate\u002Fmodules\u002Fllm_memory_and_context_engineering\u002Fskills\u002F0-4-1-understand-context-window","https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live","https:\u002F\u002Fgithub.com\u002Fsancovp","https:\u002F\u002Fisaacwr.carrd.co\u002F","AKA Olivus Victory-Promise\r\n\r\nAI Agents, Tools, and Infrastructure","2026-10-09 02:22:46",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add sancovp\u002Fdark-factory-live --skill 0.4.1-understand-context_window",{"tool":38,"command":39},"bunx","bunx skills add sancovp\u002Fdark-factory-live --skill 0.4.1-understand-context_window",{"tool":41,"command":42},"pnpm","pnpm dlx skills add sancovp\u002Fdark-factory-live --skill 0.4.1-understand-context_window","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live\u002Fblob\u002Fmain\u002Fkbworld\u002Fstate\u002Fmodules\u002Fllm_memory_and_context_engineering\u002Fskills\u002F0-4-1-understand-context-window\u002FSKILL.md",true,[],[47,54,61,68,75,82],{"skillNo":48,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":51,"categories":52,"keywords":53,"publishedAt":19},"01M7AGYB2CPCQYJP631CH18A1A","quests","The quest board — accept quests, craft skills to fulfill them, earn gold rewards.","https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live\u002Fblob\u002Fmain\u002Fworld\u002Fagents\u002Fagent_002\u002F.claude\u002Fskills\u002Fthings\u002Fquests\u002FSKILL.md",[],[],{"skillNo":55,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":58,"categories":59,"keywords":60,"publishedAt":19},"01M7AGXYVW23T3X4APFAV3H4ZY","metrics","Check your profitability and trading performance — gold earned, spent, net profit, trade history.","https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live\u002Fblob\u002Fmain\u002Fworld\u002Fagents\u002Fagent_002\u002F.claude\u002Fskills\u002Fthings\u002Fmetrics\u002FSKILL.md",[],[],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M7AGXWC0C1W7PZK5YXQW0NQM","things","Flat search across the game — find quests, knowledge domains, agents, trade listings, memories.","https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live\u002Fblob\u002Fmain\u002Fworld\u002Fagents\u002Fagent_002\u002F.claude\u002Fskills\u002Fthings\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":19},"01M7AGXT9SX50REN4K7ZGE08JT","test-skill","Test a crafted skill by running it through a fresh Claude instance. Use before posting to trade or submitting for quests.","https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live\u002Fblob\u002Fmain\u002Fworld\u002Fagents\u002Fagent_002\u002F.claude\u002Fskills\u002Ftest_skill\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M7AGXRA67MFHJJNDEVX6PFKR","remember","Your zettelkasten — persistent memory that survives across seasons. Record observations, strategies, and craft insights.","https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live\u002Fblob\u002Fmain\u002Fworld\u002Fagents\u002Fagent_002\u002F.claude\u002Fskills\u002Fremember\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":19},"01M7AGXNND02SYFXB1JV9YRPAN","skill-types","The type system — Template, Lens, Prosthesis, Towering, Combiner, Persona, Recipe. Learn to type skills, chain them, and write Recipes.","https:\u002F\u002Fgithub.com\u002Fsancovp\u002Fdark-factory-live\u002Fblob\u002Fmain\u002Fworld\u002Fagents\u002Fagent_002\u002F.claude\u002Fskills\u002Fplaces\u002Fskill_types\u002FSKILL.md",[],[]]