feat: keep the working toolset available; write files instead of printing code

morganlinton · claude · morganlinton · commit d72f6df01e4c · 2026-06-04T08:47:20.000-07:00
Root cause of "it prints the whole HTML instead of writing files": the auto
tool-selector gated edit tools behind keyword guesses, so "build a bio site"
got shell but no file_write/file_edit — the model couldn't save anything.

- Replace the keyword heuristic: auto now sends the full configured pool for
  any real request, and no tools only for plain greetings (small talk).
- Add file_write to the core defaults so new files can be created from scratch.
- System prompt now tells the model to write code to files with the tools and
  reply with a short summary — never paste file contents into the chat.

Co-Authored-By: Claude Opus 4.8 &lt;noreply@anthropic.com&gt;
diff --git a/.env.example b/.env.example
@@ -28,8 +28,8 @@ APPROVAL_POLICY=always
 # of prompt eval per turn — meaningful on small local models.
 # Available: apply_patch, file_read, file_write, file_edit, glob, grep, list_dir,
 #   shell, repo_search, run_tests, batch_edit, ship_status, web_fetch, update_plan, task
-# Default (core): file_read, grep, list_dir, file_edit, shell, update_plan, task
-# AGENT_TOOLS=file_read,grep,list_dir,file_edit,shell,update_plan,task
+# Default (core): file_read, grep, list_dir, file_edit, file_write, shell, update_plan, task
+# AGENT_TOOLS=file_read,grep,list_dir,file_edit,file_write,shell,update_plan,task
 
 # Tool schema selection. auto sends only prompt-relevant enabled tools;
 # fixed sends every enabled tool schema every turn.
diff --git a/README.md b/README.md
@@ -241,11 +241,11 @@ All model-tuning lives under `/doctor`:
 | Workflow | `run_tests`, `ship_status`, `web_fetch`, `update_plan`, `task` |
 | MCP | anything an MCP server exposes, surfaced as `mcp__<server>__<tool>` |
 
-Each tool definition costs prompt-eval time on small local models. The
-default `toolSelection: "auto"` sends only the tools that look relevant to
-each prompt; switch to `fixed` to always send the full set. Toggle the
-active pool with `/tools file_read,grep,list_dir`, or persistently in
-`agent.config.json`.
+The default `toolSelection: "auto"` keeps the full working pool available for
+any real request (so "build me a site" writes files instead of dumping code
+into the chat) and sends no tools only for plain greetings. `fixed` always
+sends the pool. Set the pool with `/tools file_read,grep,list_dir`, or
+persistently in `agent.config.json`.
 
 ### Approval policies
 
@@ -440,7 +440,7 @@ OPENROUTER_API_KEY=sk-or-...                            # required for openroute
 OPENAI_BASE_URL=https://api.openai.com/v1               # point at a compatible proxy if needed
 
 APPROVAL_POLICY=always                                  # always | dangerous-only | never
-AGENT_TOOLS=file_read,grep,list_dir,file_edit,shell,update_plan,task
+AGENT_TOOLS=file_read,grep,list_dir,file_edit,file_write,shell,update_plan,task
 AGENT_TOOL_SELECTION=auto                               # auto | fixed
 
 WARMUP=true                                             # pre-warm prompt cache at startup
@@ -460,7 +460,7 @@ root. Common shape:
   "backend": "ollama",
   "modelOverride": "qwen2.5-coder:14b",
   "approvalPolicy": "dangerous-only",
-  "tools": ["file_read", "grep", "list_dir", "file_edit", "shell", "update_plan", "task"],
+  "tools": ["file_read", "grep", "list_dir", "file_edit", "file_write", "shell", "update_plan", "task"],
   "toolSelection": "auto",
   "maxSteps": 20,
   "workspaceRoot": "/path/to/project",
diff --git a/src/config.rs b/src/config.rs
@@ -444,9 +444,17 @@ const SYSTEM_PROMPT: &str = concat!(
     "- Only call a tool when the user's request actually needs filesystem access.\n",
     "- When you do call a tool, emit a real tool call — not a JSON description in your text response.\n",
     "\n",
-    "When working with code:\n",
+    "When the user asks you to create or change code or files:\n",
+    "- DO write the changes to disk with the file tools: file_write to create a\n",
+    "  new file, file_edit to modify an existing one. Then briefly say what you\n",
+    "  did.\n",
+    "- DO NOT paste file contents or large code blocks into your reply. The user\n",
+    "  sees the files and diffs directly — repeating them wastes tokens and is\n",
+    "  not what they asked for. Only show code inline if they explicitly ask to\n",
+    "  see it, or for a tiny (1-3 line) illustration.\n",
+    "- After the edits, reply with a short summary: what you created or changed\n",
+    "  and the key files — a sentence or two, not the code.\n",
     "- Make minimal targeted edits consistent with existing style.\n",
-    "- Be concise. The user can read the diff.\n",
     "- For a task that takes 3 or more steps, call update_plan first with the\n",
     "  full plan, then call it again to mark each step done as you finish it.\n",
     "  Keep exactly one step in_progress. Skip the plan for trivial one-shot tasks.\n",
@@ -481,6 +489,7 @@ impl Default for AgentConfig {
                 "grep".into(),
                 "list_dir".into(),
                 "file_edit".into(),
+                "file_write".into(),
                 "shell".into(),
                 "update_plan".into(),
                 "task".into(),
diff --git a/src/tools/mod.rs b/src/tools/mod.rs
@@ -72,125 +72,69 @@ pub trait Tool: Send + Sync {
     }
 }
 
-fn enabled(config: &AgentConfig, name: &str) -> bool {
-    config.tools.iter().any(|t| t == name)
-}
-
-fn push_if_enabled(out: &mut Vec<String>, config: &AgentConfig, name: &str) {
-    if enabled(config, name) && !out.iter().any(|t| t == name) {
-        out.push(name.to_string());
-    }
-}
-
+/// Choose the tools to expose for a given prompt.
+///
+/// `Fixed` always sends the whole configured pool. `Auto` (the default) keeps
+/// the full working set available for any real request and sends *no* tools
+/// only for obvious small talk (greetings, thanks) — the model still decides
+/// whether to call them. This mirrors how modern harnesses work: the agent
+/// always has read/edit/write/shell on hand, so "build me a site" results in
+/// files written to disk instead of code dumped into the chat. The old
+/// keyword-bucket heuristic guessed wrong on phrasings like "build a bio site"
+/// (no edit tools surfaced), which forced the model to print code it couldn't
+/// save.
 pub fn select_tool_names(config: &AgentConfig, prompt: &str) -> Vec<String> {
-    if config.tool_selection == ToolSelection::Fixed {
+    if config.tool_selection == ToolSelection::Fixed || !is_small_talk(prompt) {
         return config.tools.clone();
     }
+    Vec::new()
+}
 
-    let lower = prompt.to_lowercase();
-    let fileish = [
-        "file",
-        "files",
-        "repo",
-        "repository",
-        "code",
-        "src",
-        "read",
-        "open",
-        "search",
-        "grep",
-        "find",
-        "list",
-        "directory",
-        "folder",
-        "where is",
-        "inspect",
-        "review",
-    ]
-    .iter()
-    .any(|needle| lower.contains(needle));
-    let editish = [
-        "edit",
-        "change",
-        "modify",
-        "update",
-        "fix",
-        "implement",
-        "add support",
-        "refactor",
-        "write",
-        "create",
-        "delete",
-        "patch",
-        "build it",
-    ]
-    .iter()
-    .any(|needle| lower.contains(needle));
-    let shellish = [
-        "run ", "execute", "command", "terminal", "shell", "test", "cargo ", "npm ", "git ",
-        "build", "lint", "check",
-    ]
-    .iter()
-    .any(|needle| lower.contains(needle));
-    let testish = ["failing", "verify", "unit test", "test suite"]
-        .iter()
-        .any(|needle| lower.contains(needle))
-        || (lower.contains("test") && !lower.contains("latest"));
-    let shipish = [
-        "ship",
-        "ready to commit",
-        "ready to ship",
-        "shipcheck",
-        "handoff",
-        "release",
-    ]
-    .iter()
-    .any(|needle| lower.contains(needle));
-    let batchish = [
-        "multi-file",
-        "multiple files",
-        "across files",
-        "batch edit",
-        "coordinated",
-    ]
-    .iter()
-    .any(|needle| lower.contains(needle));
-
-    let mut out = Vec::new();
-    if fileish || editish {
-        if config.project_memory.enabled {
-            push_if_enabled(&mut out, config, "repo_search");
-        }
-        push_if_enabled(&mut out, config, "file_read");
-        push_if_enabled(&mut out, config, "grep");
-        push_if_enabled(&mut out, config, "list_dir");
-        push_if_enabled(&mut out, config, "glob");
-        // Delegated read-only investigation keeps deep exploration out of the
-        // parent context.
-        push_if_enabled(&mut out, config, "task");
-    }
-    if editish {
-        push_if_enabled(&mut out, config, "file_edit");
-        push_if_enabled(&mut out, config, "apply_patch");
-        push_if_enabled(&mut out, config, "file_write");
-    }
-    // Multi-step work (editing or running things) benefits from a visible plan.
-    if editish || shellish {
-        push_if_enabled(&mut out, config, "update_plan");
+/// Conservatively detect throwaway social messages (greetings, thanks) so we
+/// can skip sending tool schemas for them. Deliberately narrow: a false
+/// negative just sends tools the model won't use, but a false positive would
+/// starve a real request of its tools, so we only match very short, clearly
+/// social inputs.
+fn is_small_talk(prompt: &str) -> bool {
+    let t = prompt.trim().to_lowercase();
+    let t = t.trim_end_matches(['.', '!', '?', ' ']);
+    if t.is_empty() {
+        return true;
     }
-    if shellish {
-        push_if_enabled(&mut out, config, "shell");
-    }
-    if testish || shellish {
-        push_if_enabled(&mut out, config, "run_tests");
-    }
-    if batchish && editish {
-        push_if_enabled(&mut out, config, "batch_edit");
-    }
-    if shipish {
-        push_if_enabled(&mut out, config, "ship_status");
-    }
-    out
+    const EXACT: &[&str] = &[
+        "hi",
+        "hii",
+        "hey",
+        "hello",
+        "yo",
+        "sup",
+        "thanks",
+        "thank you",
+        "thx",
+        "ty",
+        "ok",
+        "okay",
+        "k",
+        "cool",
+        "nice",
+        "great",
+        "awesome",
+        "hi there",
+        "hey there",
+        "hello there",
+        "good morning",
+        "good afternoon",
+        "good evening",
+        "how are you",
+        "how's it going",
+        "who are you",
+        "what are you",
+        "what can you do",
+        "gm",
+        "bye",
+        "goodbye",
+    ];
+    EXACT.contains(&t)
 }
 
 /// Returns true when a tool result indicates the workspace was mutated (for ship-loop hooks).
@@ -329,85 +273,73 @@ mod tests {
     use crate::config::ToolSelection;
 
     #[test]
-    fn auto_chat_uses_no_tools() {
+    fn small_talk_sends_no_tools() {
         let config = AgentConfig {
             tool_selection: ToolSelection::Auto,
             ..Default::default()
         };
-        assert!(select_tool_names(&config, "hello there").is_empty());
+        for greeting in [
+            "hi",
+            "hello there",
+            "thanks",
+            "ok",
+            "  hey!  ",
+            "Good morning",
+        ] {
+            assert!(
+                select_tool_names(&config, greeting).is_empty(),
+                "{greeting:?} should be treated as small talk"
+            );
+        }
     }
 
     #[test]
-    fn auto_file_question_uses_read_tools() {
+    fn real_request_sends_the_full_pool() {
         let config = AgentConfig {
             tool_selection: ToolSelection::Auto,
             ..Default::default()
         };
-        let names = select_tool_names(&config, "search the repo for config");
-        assert!(names.contains(&"file_read".to_string()));
-        assert!(names.contains(&"grep".to_string()));
-        assert!(names.contains(&"list_dir".to_string()));
+        assert_eq!(
+            select_tool_names(&config, "what files are in src?"),
+            config.tools
+        );
     }
 
     #[test]
-    fn repo_search_is_opt_in_not_default() {
-        // repo_search is no longer in the default pool; auto-selection can't
-        // surface it unless the user enables it.
+    fn build_request_includes_write_tools() {
+        // Regression: "build a bio site" used to surface no edit tools, forcing
+        // the model to dump code into the chat. It must now get file_write +
+        // file_edit + shell so it can actually create files.
         let config = AgentConfig {
             tool_selection: ToolSelection::Auto,
             ..Default::default()
         };
-        assert!(!config.tools.contains(&"repo_search".to_string()));
-        let names = select_tool_names(&config, "search the repo for config");
-        assert!(!names.contains(&"repo_search".to_string()));
+        let names = select_tool_names(&config, "build a bio site");
+        assert!(names.contains(&"file_write".to_string()));
+        assert!(names.contains(&"file_edit".to_string()));
+        assert!(names.contains(&"shell".to_string()));
     }
 
     #[test]
-    fn auto_repo_search_requires_memory_enabled() {
-        // When repo_search IS enabled, auto-selection still gates it on memory.
-        let base_tools = vec![
-            "repo_search".to_string(),
-            "file_read".to_string(),
-            "grep".to_string(),
-            "list_dir".to_string(),
-        ];
-        let enabled = AgentConfig {
-            tools: base_tools.clone(),
-            ..Default::default()
-        };
-        assert!(select_tool_names(&enabled, "search the repo for config")
-            .contains(&"repo_search".to_string()));
-
-        let disabled = AgentConfig {
-            tools: base_tools,
-            project_memory: crate::config::ProjectMemoryConfig {
-                enabled: false,
-                ..Default::default()
-            },
-            ..Default::default()
-        };
-        assert!(!select_tool_names(&disabled, "search the repo for config")
-            .contains(&"repo_search".to_string()));
-    }
-
-    #[test]
-    fn auto_verify_prompt_includes_run_tests() {
+    fn auto_only_sends_configured_tools() {
         let config = AgentConfig {
-            tools: vec!["run_tests".into(), "file_read".into(), "file_edit".into()],
+            tool_selection: ToolSelection::Auto,
+            tools: vec!["file_read".into(), "shell".into()],
             ..Default::default()
         };
-        let names = select_tool_names(&config, "verify the failing unit test");
-        assert!(names.contains(&"run_tests".to_string()));
+        assert_eq!(
+            select_tool_names(&config, "do some real work"),
+            vec!["file_read".to_string(), "shell".to_string()]
+        );
     }
 
     #[test]
-    fn auto_ship_prompt_includes_ship_status() {
+    fn fixed_mode_sends_full_pool_even_for_greetings() {
         let config = AgentConfig {
-            tools: vec!["ship_status".into(), "file_read".into()],
+            tool_selection: ToolSelection::Fixed,
             ..Default::default()
         };
-        let names = select_tool_names(&config, "is this ready to ship?");
-        assert!(names.contains(&"ship_status".to_string()));
+        assert_eq!(select_tool_names(&config, "hi"), config.tools);
     }
 
     #[test]
