read_file output cap raised to ~256 KB** (64k tokens)

Author: alexylon

CHANGELOG.md

@@ -4,6 +4,14 @@ All notable changes to Sofos are documented in this file.
 
 ## [Unreleased]
 
+### Changed
+
+- **`read_file` output cap raised to ~256 KB** (64k tokens). Previously `read_file` shared the ~64 KB / 16k-token cap with `execute_bash` and `search_code`, which clipped mid-sized source files — generated code, JSON fixtures, long prompt templates — and forced the model into an extra range-reads round trip against the 200-iteration tool-loop budget. `execute_bash` stdout/stderr and `search_code` keep the 16k-token cap, since verbose test output and broad ripgrep patterns benefit from being forced to narrow rather than handing the model noise.
+
+### Fixed
+
+- Prompt glyph (> or :) now correctly reflects the normal/safe modes.
+
 ## [0.2.2] - 2026-04-21
 
 ### Security

src/tools/mod.rs

@@ -22,8 +22,8 @@ use tool_name::ToolName;
 
 use crate::tools::types::get_read_only_tools;
 use crate::tools::utils::{
-    MAX_DIFF_TOKENS, MAX_MCP_IMAGE_BYTES, MAX_MCP_IMAGE_COUNT, MAX_MCP_OUTPUT_TOKENS,
-    MAX_PATH_LIST_TOKENS, MAX_TOOL_OUTPUT_TOKENS, TruncationKind, confirm_destructive,
+    MAX_DIFF_TOKENS, MAX_FILE_READ_TOKENS, MAX_MCP_IMAGE_BYTES, MAX_MCP_IMAGE_COUNT,
+    MAX_MCP_OUTPUT_TOKENS, MAX_PATH_LIST_TOKENS, TruncationKind, confirm_destructive,
     is_absolute_or_tilde, truncate_for_context,
 };
 pub use types::{add_code_search_tool, get_all_tools, get_all_tools_with_morph};
@@ -659,7 +659,7 @@ impl ToolExecutor {
                         .read_file_with_outside_access(&resolved.canonical_str)?
                 };
                 let content =
-                    truncate_for_context(&raw, MAX_TOOL_OUTPUT_TOKENS, TruncationKind::File);
+                    truncate_for_context(&raw, MAX_FILE_READ_TOKENS, TruncationKind::File);
                 Ok(format!("File content of '{}':\n\n{}", path, content))
             }
             ToolName::WriteFile => {

src/tools/utils.rs

@@ -3,12 +3,25 @@ use std::io;
 use std::io::Write;
 use std::sync::OnceLock;
 
-/// Maximum tokens (≈ chars / 4, ≈ 64 KB) a single tool call is allowed
-/// to return before [`truncate_for_context`] clips it with an informational
-/// suffix. Keeps a single large bash output or file read from monopolising
-/// the model's context window.
+/// Maximum tokens (≈ chars / 4, ≈ 64 KB) returned from `execute_bash`
+/// (stdout / stderr) and `search_code` (ripgrep output) before
+/// [`truncate_for_context`] clips with an informational suffix. Verbose
+/// test runs and broad regex searches spiral quickly; capping here forces
+/// the model to narrow the query or redirect to a file rather than drown
+/// the context in noise. `read_file` uses a separate, larger cap —
+/// see [`MAX_FILE_READ_TOKENS`].
 pub const MAX_TOOL_OUTPUT_TOKENS: usize = 16_000;
 
+/// Larger cap (≈ 256 KB) for `read_file` output. A single mid-sized
+/// source file — generated code, JSON fixtures, long prompt templates —
+/// routinely exceeds the 16 KB bash/search budget, and clipping it
+/// forces the model into a range-reads round trip against the
+/// 200-iteration budget. Files are structured content the model
+/// typically needs to reason about in full, so the tradeoff favours a
+/// bigger budget here than for bash noise. Still far below OpenAI's
+/// 10 MB per-tool-output ceiling.
+pub const MAX_FILE_READ_TOKENS: usize = 64_000;
+
 /// Separate, more generous cap for path-list tools (`list_directory`,
 /// `glob_files`). Filenames are short and the model often needs to see
 /// the full listing to locate a specific file, so we apply roughly an