pr_01m47d24b0e6n91zwymwxg0vpx/crates/runner/src/harness.rs
| 1 | //! The agent harness. g1t does not implement its own agent loop: it runs |
| 2 | //! Claude Code headless and translates its event stream into session |
| 3 | //! entries. |
| 4 | |
| 5 | use std::io::{BufRead, BufReader}; |
| 6 | use std::path::Path; |
| 7 | use std::process::{Command, Stdio}; |
| 8 | |
| 9 | use anyhow::{Context, Result, bail}; |
| 10 | use serde_json::Value; |
| 11 | |
| 12 | use crate::report::{Entry, Reporter}; |
| 13 | |
| 14 | const MAX_TURNS: &str = "80"; |
| 15 | |
| 16 | /// A compact, readable rendering of a tool's input. |
| 17 | fn describe_input(tool: &str, input: &Value) -> String { |
| 18 | let field = |name: &str| input.get(name).and_then(Value::as_str); |
| 19 | match tool { |
| 20 | "Bash" => field("command").unwrap_or_default().to_owned(), |
| 21 | "Read" | "Write" | "Edit" | "MultiEdit" | "NotebookEdit" => { |
| 22 | field("file_path").unwrap_or_default().to_owned() |
| 23 | } |
| 24 | "Glob" | "Grep" => field("pattern").unwrap_or_default().to_owned(), |
| 25 | _ => serde_json::to_string(input).unwrap_or_default(), |
| 26 | } |
| 27 | } |
| 28 | |
| 29 | /// The text of a tool result, which is either a string or content blocks. |
| 30 | fn result_text(content: &Value) -> String { |
| 31 | match content { |
| 32 | Value::String(text) => text.clone(), |
| 33 | Value::Array(blocks) => blocks |
| 34 | .iter() |
| 35 | .filter_map(|block| block.get("text").and_then(Value::as_str)) |
| 36 | .collect::<Vec<_>>() |
| 37 | .join("\n"), |
| 38 | _ => String::new(), |
| 39 | } |
| 40 | } |
| 41 | |
| 42 | /// Records one line of Claude Code's `stream-json` output. Returns the final |
| 43 | /// result when the line is the one that ends the run. |
| 44 | fn handle_event(event: &Value, reporter: &mut Reporter) -> Option<Result<String>> { |
| 45 | let blocks = || { |
| 46 | event["message"]["content"] |
| 47 | .as_array() |
| 48 | .map(Vec::as_slice) |
| 49 | .unwrap_or_default() |
| 50 | }; |
| 51 | match event["type"].as_str()? { |
| 52 | "assistant" => { |
| 53 | for block in blocks() { |
| 54 | match block["type"].as_str() { |
| 55 | Some("text") => { |
| 56 | let text = block["text"].as_str().unwrap_or_default().trim(); |
| 57 | if !text.is_empty() { |
| 58 | reporter.record(Entry::new("message", text)); |
| 59 | } |
| 60 | } |
| 61 | Some("tool_use") => { |
| 62 | let tool = block["name"].as_str().unwrap_or("tool"); |
| 63 | reporter.record(Entry::tool( |
| 64 | "tool_call", |
| 65 | tool, |
| 66 | &describe_input(tool, &block["input"]), |
| 67 | )); |
| 68 | } |
| 69 | _ => {} |
| 70 | } |
| 71 | } |
| 72 | None |
| 73 | } |
| 74 | "user" => { |
| 75 | for block in blocks() { |
| 76 | if block["type"] == "tool_result" { |
| 77 | let text = result_text(&block["content"]); |
| 78 | if !text.trim().is_empty() { |
| 79 | reporter.record(Entry::tool("tool_result", "result", text.trim())); |
| 80 | } |
| 81 | } |
| 82 | } |
| 83 | None |
| 84 | } |
| 85 | "result" => { |
| 86 | let text = event["result"].as_str().unwrap_or_default().to_owned(); |
| 87 | Some(if event["is_error"].as_bool().unwrap_or(false) { |
| 88 | Err(anyhow::anyhow!("the agent reported an error: {text}")) |
| 89 | } else { |
| 90 | Ok(text) |
| 91 | }) |
| 92 | } |
| 93 | _ => None, |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | /// Runs Claude Code on `prompt` in `workdir` and returns its closing |
| 98 | /// summary. |
| 99 | pub fn run_claude(workdir: &Path, prompt: &str, reporter: &mut Reporter) -> Result<String> { |
| 100 | let mut child = Command::new("claude") |
| 101 | .current_dir(workdir) |
| 102 | .args([ |
| 103 | "--print", |
| 104 | prompt, |
| 105 | "--output-format", |
| 106 | "stream-json", |
| 107 | "--verbose", |
| 108 | "--max-turns", |
| 109 | MAX_TURNS, |
| 110 | // The sandbox is the permission boundary: it holds one fork and |
| 111 | // one short-lived token, and nothing else. |
| 112 | "--dangerously-skip-permissions", |
| 113 | ]) |
| 114 | // The agent needs the model key and nothing else of ours. |
| 115 | .env_remove("G1T_TOKEN") |
| 116 | .stdin(Stdio::null()) |
| 117 | .stdout(Stdio::piped()) |
| 118 | .stderr(Stdio::inherit()) |
| 119 | .spawn() |
| 120 | .context("could not start Claude Code")?; |
| 121 | |
| 122 | let stdout = child.stdout.take().context("no output from Claude Code")?; |
| 123 | let mut outcome = None; |
| 124 | for line in BufReader::new(stdout).lines() { |
| 125 | let line = line?; |
| 126 | let Ok(event) = serde_json::from_str::<Value>(&line) else { |
| 127 | continue; |
| 128 | }; |
| 129 | if let Some(result) = handle_event(&event, reporter) { |
| 130 | outcome = Some(result); |
| 131 | } |
| 132 | } |
| 133 | let status = child.wait()?; |
| 134 | match outcome { |
| 135 | Some(result) => result, |
| 136 | None => bail!("Claude Code exited ({status}) without a result"), |
| 137 | } |
| 138 | } |