Function render_tool_output 

pub fn render_tool_output(v: &Value) -> String

A tool result’s output as a model should read it: outcome fields first (truncated, error, ok, status, exit_code, timed_out), one line each; then other scalars as key: value lines; then each long or multi-line string (a file’s content, a command’s output) as a raw block headed key (N bytes):, and each array of records as key (N records): with one line per record. A cut at the budget loses the tail of the bulk, never the verdict.

Journals, receipts and events keep the JSON; only the model’s copy changes. The model used to be handed the JSON too — so a file arrived as one line of \n, \t and \" escapes. On claude-haiku-4-5, 23% of native edit_file calls (42 of 182) missed with “old_text not found”, against 1 of 60 for Claude Code’s Edit on the same model and tasks; the misses carried escapes copied out of the JSON (\"\"\" for a docstring’s """). Claude Code’s Read returns plain numbered lines; this is that shape.

The text reads back exactly with parse_rendered_output, apart from the outcome fields (shown on one line) and a record list (read back as text): blocks carry their length, so no content line can be mistaken for a field, and a string that would read back as another JSON type ("123", "null") is written quoted.