Show HN: Trueline – 해시 검증 편집으로 Claude의 출력 토큰 44% 절약

An MCP plugin that gives AI coding agents hash-verified file editing and targeted reads. Works with Claude Code, Gemini CLI, VS Code Copilot, OpenCode, and Codex CLI. Claude Code (recommended; hooks are automatic): /plugin marketplace add rjkaes/trueline-mcp /plugin install trueline-mcp@trueline-mcp Other platforms (Gemini CLI, VS Code Copilot, OpenCode, Codex CLI): See INSTALL.md for platform-specific setup. CLI (no MCP): For agents that use shell commands instead of MCP, install globally with npm i -g trueline-mcp and add configs/cli/instructions.md to your agent's system prompt. See INSTALL.md. AI coding agents read entire files to find one function, then echo back everything they're replacing. Both waste context on content the agent already knows or doesn't need. That context costs money, eats into the conversation window, and limits how much real work fits in a session. Worse, the built-in edit tools match by string content. If the agent hallucinates a line, works from stale context, or hits an ambiguous match, your code gets silently corrupted. trueline fixes both problems: it reads less, writes less, and rejects every edit that doesn't match the file's actual content. trueline provides six MCP tools organized around three workflows. trueline_outline returns an AST-based structural outline of any file: functions, classes, declarations, and their line ranges. For a typical source file, that's 10-20 lines instead of hundreds. 1-10: (10 imports) 12-12: const VERSION = pkg.version; 14-17: const server = new McpServer({ 25-45: async function resolveAllowedDirs(): Promise { 49-69: server.registerTool( 71-92: server.registerTool( (12 symbols, 139 source lines) The agent sees the full structure, then uses trueline_read to fetch only the ranges it needs. A 500-line file where the agent needs one 20-line function? It reads 20 lines, not 500. trueline_search finds lines by literal string or regex and returns them with edit-ready checksums, no outline or read step needed. For targeted edits where the agent knows what it's looking for, this is the fastest path. The built-in edit tool requires the agent to echo back the old text being replaced. trueline_edit replaces that with a compact line-range reference and a content hash. The agent outputs only the new content. The savings scale with the size of the replaced block. A one-line change saves little; replacing 30 lines of old code saves the agent from outputting all 30 of those lines again. Multiple edits can be batched in a single call and applied atomically. trueline_changes provides an AST-based summary of structural changes compared to a git ref. Instead of raw line diffs, it reports added/removed/renamed symbols, signature changes, and logic modifications with inline mini-diffs. Pass ["*"] to diff all changed files at once. Every line from trueline_read and trueline_search carries a content hash. Every edit must present those hashes back, proving the agent is editing what it thinks it's editing. If the file changed since the agent read it (concurrent edits, a build step, another tool), the edit is rejected. If the agent hallucinates content that doesn't match what's on disk, the edit is rejected. If the agent targets the wrong lines, the edit is rejected. Nothing hits disk unless the hashes match. trueline_verify checks whether held checksums are still valid without re-reading the file. When nothing changed (the common case), the response is a single line. trueline doesn't just register tools and hope the agent picks them up. On platforms that support hooks, it actively intercepts the agent's workflow: - SessionStart injects instructions telling the agent how and when to use each trueline tool, calibrated per platform. - PreToolUse intercepts calls to the built-in edit tool and blocks them, forcing the agent through hash-verified edits instead. With hooks, agent compliance is ~98%. Without hooks (instruction-only platforms like OpenCode and Codex CLI), compliance is ~60%. The instruction file still helps; hooks make it reliable. The instructions are not one-size-fits-all. They reference each platform's native tool names (Read /Edit on Claude Code, read_file /edit_file on Gemini CLI, view /edit on OpenCode) and adapt advice accordingly. trueline's overhead is an MCP round-trip per tool call. For small files (under ~200 lines), the built-in tools are perfectly fine, and the injected instructions tell the agent so. The payoff comes on larger files and multi-file editing sessions, where targeted reads and compact edits avoid sending hundreds or thousands of redundant lines through the context window. See DESIGN.md for the protocol specification, hash algorithm details, streaming architecture, and security model. Requires Bun ≥ 1.3. bun install # install dependencies bun test # run tests bun run build # build binary for the current platform Inspired by The Harness Problem by Can Boluk and the vscode-hashline-edit-tool by Seth Livingston.