Question 1: Caveman targets token bloat by conditioning coding agents (like Claude Code, Codex, or Gemini CLI) to strip grammatical fluff while preserving exact technical instructions. From a prompt-engineering perspective, how does the underlying SKILL.md system prompt enforce this high-compression syntax without causing the model to miss subtle edge cases in complex logic, or drop critical context during deep code modifications?
Question 2: To keep agent outputs structurally compact, Caveman features distinct operational intensities (such as lite, full, ultra, and wenyan). How does the system dynamically inject these constraints into the active generation session? Does it depend entirely on formatting rubrics in the system prompt, or does it leverage targeted Few-Shot user-assistant turn examples to align the target LLM with the desired compression density?
Question 3: For terminal setups like Claude Code, the repository sets up local lifecycle hooks to configure automatic session startup, active session flags, and an interactive status-line token counter (/caveman-stats). How does the installation script register these hooks into the client agent's terminal configuration directory, and how does it parse local execution logs to update the lifetime savings metric without blocking live input/output rendering?
Question 4: Beyond compressing immediate responses, the /caveman-compress sub-tool acts on local workspace context documents (like CLAUDE.md, documentation pages, or todo lists) to rewrite them into a high-density format. What precise parsing strategy does the compressor use to guarantee that critical identifiers, package scripts, syntax tokens, and source paths remain perfectly byte-preserved during the file translation?
Question 5: Recent ecosystem additions introduce specialized sub-agents (cavecrew-investigator, cavecrew-builder, and cavecrew-reviewer) that delegate task execution via a centralized decision-making guide. How does the core coordinator evaluate task complexity to decide when to spawn a scoped caveman-style worker subagent rather than attempting to process the code adjustments directly inside the main execution loop?
Question 6: To fully embed this token-saving behavior into model weights, the ecosystem introduces Cavegemma—a Gemma 4 31B fine-tune trained on baseline-to-caveman communication records. When evaluating the model via the three-arm evaluation harness (evals/), what structural metrics (e.g., semantic drift, AST matching, compilation success rates) are used to verify that the fine-tuned weights maintain exact functional alignment with standard verbose outputs?
Question 1: Caveman targets token bloat by conditioning coding agents (like Claude Code, Codex, or Gemini CLI) to strip grammatical fluff while preserving exact technical instructions. From a prompt-engineering perspective, how does the underlying SKILL.md system prompt enforce this high-compression syntax without causing the model to miss subtle edge cases in complex logic, or drop critical context during deep code modifications?
Question 2: To keep agent outputs structurally compact, Caveman features distinct operational intensities (such as lite, full, ultra, and wenyan). How does the system dynamically inject these constraints into the active generation session? Does it depend entirely on formatting rubrics in the system prompt, or does it leverage targeted Few-Shot user-assistant turn examples to align the target LLM with the desired compression density?
Question 3: For terminal setups like Claude Code, the repository sets up local lifecycle hooks to configure automatic session startup, active session flags, and an interactive status-line token counter (/caveman-stats). How does the installation script register these hooks into the client agent's terminal configuration directory, and how does it parse local execution logs to update the lifetime savings metric without blocking live input/output rendering?
Question 4: Beyond compressing immediate responses, the /caveman-compress sub-tool acts on local workspace context documents (like CLAUDE.md, documentation pages, or todo lists) to rewrite them into a high-density format. What precise parsing strategy does the compressor use to guarantee that critical identifiers, package scripts, syntax tokens, and source paths remain perfectly byte-preserved during the file translation?
Question 5: Recent ecosystem additions introduce specialized sub-agents (cavecrew-investigator, cavecrew-builder, and cavecrew-reviewer) that delegate task execution via a centralized decision-making guide. How does the core coordinator evaluate task complexity to decide when to spawn a scoped caveman-style worker subagent rather than attempting to process the code adjustments directly inside the main execution loop?
Question 6: To fully embed this token-saving behavior into model weights, the ecosystem introduces Cavegemma—a Gemma 4 31B fine-tune trained on baseline-to-caveman communication records. When evaluating the model via the three-arm evaluation harness (evals/), what structural metrics (e.g., semantic drift, AST matching, compilation success rates) are used to verify that the fine-tuned weights maintain exact functional alignment with standard verbose outputs?