Skip to content
/
Phase D Batch 24

Phase D Batch 24 Notes — Config / Snapshot / Token Utility CLIs

Phase D Batch 24 Notes — Config / Snapshot / Token Utility CLIs

Batch theme: Single-purpose CLIs managing AI-context configs, snapshots, or token spend reduction.

Agent: Phase D batch agent #24 of 33 Frameworks analyzed: 10 Files written: 110 (11 per framework × 10 frameworks) Tier C stubs: 0 Duplicates encountered: none


Framework Summary

Slug Stars Language Primary pattern Key dimension
codesight 1,074 TypeScript codebase-to-LLM packer MCP server (13 tools) + HTML report
prpack 2 JavaScript PR review angles CLI Zero-dependency CLI
marmot 0 TypeScript Multi-provider agent CLI 35+ subcommands, 25+ provider adapters
codebase-pilot-cli 6 TypeScript 7-layer subagent system 19 skills, 2 hooks, system daemon
lynxprompt-geiserx 41 TypeScript Enterprise prompt platform Next.js + PostgreSQL + Helm + SAML
claude-snapshot 2 JavaScript Claude Code config backup 4 commands, tar.gz archives
claude-overlay 2 Shell Multi-provider config switcher Bash/Python CLI, surgical config management
agentify 28 TypeScript OpenAPI-to-agent compiler 9 output formats, AgentifyIR, security scan
rtk-token-killer 54,590 Rust Output compression proxy PreToolUse hook, SQLite analytics
ccs-switch 2,410 TypeScript Multi-runtime task router 1 skill, 1 command, CLIProxy, web dashboard

Key Findings

Stars Distribution

This batch has the widest star count range of any Phase D batch: from 0 (marmot, newly launched) to 54,590 (rtk-token-killer). The median is approximately 15 stars. rtk-token-killer is the most popular tool observed across all 33 batches in this catalog.

Architecture Patterns Observed

Output-side vs input-side context management — the most interesting architectural divide in this batch:

  • Input-side (traditional): codesight (packs code into context), marmot (structures prompts), lynxprompt (manages prompt library)
  • Output-side (novel): rtk-token-killer (compresses command output before Claude reads it)

rtk-token-killer is the only tool in the entire catalog that addresses token consumption at the tool output layer. All other tools address it at the prompt/context input layer.

Multi-provider routing — two distinct approaches:

  • Infrastructure-level: claude-overlay (rewrites ~/.claude/config to change the provider Claude Code uses)
  • Task-delegation-level: ccs-switch (routes entire tasks to different agent runtimes based on skill-driven profile matching)

These solve the same problem (use different providers) at very different abstraction levels.

Scope Outliers

codebase-pilot-cli (6 stars) punches far above its star count: 7-layer subagent system with haiku/sonnet/opus routing, 19 skills nearly identical to superpowers, SHA-256 incremental packing, 180 security patterns, SQLite FTS5 search, and a system daemon (launchd/systemd). This is the most feature-dense tool in the batch per star.

lynxprompt-geiserx (41 stars) is the most enterprise-complete: SAML/OIDC/LDAP SSO, Helm chart, VS Code extension, separate MCP server repo, and Chocolatey/AUR/Snap packaging. Scope exceeds the batch theme (it is not a "single-purpose CLI" at all).

marmot (0 stars) is a commercial product (marmot.sh) with a GitHub repo so new it has no community traction yet. 35+ subcommands and 25+ provider adapters suggest significant pre-launch investment. Worth monitoring for rapid star growth.

Hooks Usage Patterns

Three tools in this batch use Claude Code hooks:

  • rtk-token-killer: PreToolUse to intercept and compress Bash command output — the hook is the product
  • claude-overlay: Hooks for config injection/restoration around sessions
  • codebase-pilot-cli: UserPromptSubmit + SessionStart for 7-layer subagent routing

rtk's hook pattern is architecturally distinct: it rewrites commands rather than routing or triggering.

Security Standouts

agentify is the only code generation tool in the catalog that scans both input (OpenAPI spec for prompt injection) and output (generated TypeScript) for security issues before writing files. This supply-chain security concern is novel in the generator-tool category.

codebase-pilot-cli ships 180 security anti-patterns for code review — the most extensive static security pattern library observed in any tool in the catalog.

Generated Artifacts as Build Artifacts

agentify makes the strongest philosophical claim of any tool in the batch: that CLAUDE.md files, MCP servers, and skill files are build artifacts that should be generated from formal specifications (OpenAPI), not handwritten. This is analogous to how database clients are generated from schemas. No other tool in the catalog makes this claim.


Cross-Batch Connections

  • codebase-pilot-cli → superpowers: 19 skills in codebase-pilot are almost identical to superpowers' 14 skills (adds 5 codebase-specific skills). Likely inspired by superpowers.
  • claude-overlay → agentify: Both generate CLAUDE.md-like artifacts, but from opposite directions. Overlay modifies an existing CLAUDE.md; agentify generates one from an API spec.
  • ccs-switch → claude-code-router (musistudio): README credits claude-code-router as a related project. CCS adds account pool management and a web dashboard on top of the same routing concept.
  • rtk-token-killer has no close analog in any other batch — the output-compression-via-hook pattern is unique to this tool.

Most Surprising Find

rtk-token-killer has 54,590 GitHub stars — the most popular tool across all 33 batches in this catalog. A single Rust binary that intercepts Bash tool output via a PreToolUse hook achieved more community traction than any multi-agent orchestration framework, enterprise prompt platform, or AI-first development system in the catalog. The market validated a simple, transparent, invisible infrastructure layer over complex orchestration abstractions.