Add regression test coverage for merged PRs #13-#20

CONFIG.md

@@ -3,13 +3,13 @@
 dirge reads an optional JSON config file named `config.json` from its config
 folder:
 
-- If `ZS_CONFIG_DIR` is set: `$ZS_CONFIG_DIR/config.json`
+- If `DIRGE_CONFIG_DIR` is set: `$DIRGE_CONFIG_DIR/config.json`
 - Otherwise: the platform config directory joined with `dirge/config.json`
   (for example `$XDG_CONFIG_HOME/dirge/config.json` on Linux)
 - Fallback: `$HOME/.config/dirge/config.json`
 
 All config keys are optional. CLI flags and their environment-backed values
-(such as `ZS_PROVIDER` and `ZS_MODEL`) take precedence where both exist.
+(such as `DIRGE_PROVIDER` and `DIRGE_MODEL`) take precedence where both exist.
 
 Example:
 
@@ -59,7 +59,7 @@ Accepted top-level keys:
 
 | Key                       | Type    | Description                                                                                                                                                                 |
 | ------------------------- | ------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `provider`                | string  | Provider name. Built-ins are `openrouter`, `openai`, `anthropic`, `gemini`/`google`, and `ollama`; custom provider aliases are also accepted. Default: `openrouter`.        |
+| `provider`                | string  | Provider name. Built-ins are `openrouter`, `openai`, `anthropic`, `gemini`/`google`, `deepseek`, `glm`/`zhipu`, and `ollama`; custom provider aliases are also accepted. Default: `openrouter`.        |
 | `model`                   | string  | Model name. Default: `deepseek/deepseek-v4-flash`.                                                                                                                          |
 | `max_tokens`              | integer | Maximum response tokens. Default: `8192`.                                                                                                                                   |
 | `max_agent_turns`         | integer | Maximum agent turns per response. Default: `100`.                                                                                                                           |

README.md

@@ -5,7 +5,7 @@ Minimal coding agent written in Rust, inspired by [pi](https://pi.dev/docs/lates
 ## Features
 
 - **Multi-provider**: OpenRouter, OpenAI, Anthropic, Gemini, DeepSeek, GLM, Ollama, plus custom providers
-- **Standard tools**: read, write, edit, bash, grep, find_files, list_dir, write_todo_list
+- **Standard tools**: read, write, edit, bash, grep, find_files, glob, list_dir, write_todo_list, apply_patch
 - **Line-numbered read output**: `read` tool prefixes each line with right-aligned line numbers (`123: content`)
 - **Environment-aware**: system prompt includes OS, shell, working directory, and git branch for context
 - **Semantic code tools** (tree-sitter): list_symbols, get_symbol_body, find_definition, find_callers, find_callees — supports TypeScript/TSX and Python
@@ -36,7 +36,7 @@ _dirge_ is one of the smallest and most performant coding agents on the market.
 
 ### Tool result caching
 
-Read-only tool calls (`read`, `grep`, `find_files`, `list_dir`) are cached per agent turn. Repeated calls with identical arguments within the same turn return cached results, avoiding redundant filesystem I/O. The cache clears automatically before each new prompt, and after `write`/`edit`/`bash` so a re-read sees fresh content.
+Most tool calls (`read`, `write`, `edit`, `bash`, `grep`, `find_files`, `list_dir`) are cached per agent turn. Repeated calls with identical arguments within the same turn return cached results, avoiding redundant filesystem I/O. The cache clears automatically before each new prompt, and after `write`/`edit`/`bash` so a re-read sees fresh content.
 
 ### Error recovery
 
@@ -107,11 +107,16 @@ dirge --provider glm       # defaults to glm-4
 | `/mode [mode]` | Set security mode (`standard`, `restrictive`, `accept`, `yolo`) |
 | `/reasoning` | Toggle reasoning visibility |
 | `/btw <question>` | Ask a quick question (no tools, doesn't affect session) |
-| `/session` | List/save/load sessions |
+| `/sessions` | List/save/load sessions |
 | `/loop [prompt]` | Start iterative coding loop |
 | `/worktree <name>` | Create a git worktree on branch |
 | `/wt-merge [branch]` | Merge worktree branch |
 | `/wt-exit` | Exit worktree |
+| `/toggle` | Toggle features on/off (currently todo tools) |
+| `/regen-prompts` | Restore built-in prompts |
+| `/mcp` | List MCP servers and tools |
+| `/quit` | Exit dirge |
+| `/retry` | Retry last prompt |
 | `/help` | Show all commands |
 
 ### Key bindings
@@ -179,10 +184,11 @@ Built-in prompts that change the agent's behavior and tone:
 | **`review-security`** | Security review mode — finds exploitable vulnerabilities |
 | **`simplify`** | Code simplification mode — refines for clarity without changing behavior |
 | **`write-prompt`** | Prompt writing mode — creates and optimizes agent prompts |
+| **`default`** | Default system prompt — the base built-in prompt |
 
 Custom prompts can be placed in `$XDG_CONFIG_HOME/dirge/prompts/` as `.md` files.
 
-The agent automatically loads `AGENTS.md` or `CLAUDE.md` from the project root or ancestor directories. Use `-n` / `--no-context-files` to disable.
+The agent automatically loads `AGENTS.md` or `CLAUDE.md` from the project root, ancestor directories, and `~/.config/dirge/agent/AGENTS.md` as a global fallback. Use `-n` / `--no-context-files` to disable.
 
 ## Claude-compatible skills
 

src/agent/builder.rs

@@ -109,25 +109,11 @@ pub async fn build_agent_inner<M: CompletionModel + 'static>(
 
     // Inject mode-specific reminders
     if let Some(prompt_name) = &context.current_prompt_name {
-        match prompt_name.as_str() {
-            "plan" => {
-                preamble.push_str("\n\n---\n\nYou are now in PLAN mode. Create a detailed implementation plan. Save it to PLAN.md in the current directory. Analyze the task, break it into concrete steps, consider edge cases and trade-offs. Do NOT write any code or run any commands until the user reviews and approves the plan.");
-            }
-            "review" | "review-security" => {
-                preamble.push_str("\n\n---\n\nYou are now in REVIEW mode. Review the code or plan carefully. Identify bugs, security issues, performance problems, and design flaws. Be thorough and specific. Provide actionable feedback.");
-            }
-            "code" => {
-                let plan_path = std::env::current_dir()
-                    .unwrap_or_else(|_| ".".into())
-                    .join("PLAN.md");
-                if plan_path.exists() {
-                    preamble.push_str(
-                        "\n\n---\n\nA plan file exists at PLAN.md. Execute the plan step by step. Write and test code following the plan. Report progress after each step. The plan is your guide — follow it closely."
-                    );
-                }
-            }
-            _ => {}
-        }
+        let plan_exists = std::env::current_dir()
+            .unwrap_or_else(|_| ".".into())
+            .join("PLAN.md")
+            .exists();
+        append_mode_reminder(&mut preamble, prompt_name, plan_exists);
     }
 
     let mut builder = AgentBuilder::new(model).preamble(&preamble);
@@ -314,3 +300,80 @@ pub fn create_client(api_key: Option<&str>) -> anyhow::Result<openrouter::Client
         })?;
     Ok(openrouter::Client::new(String::from(key))?)
 }
+
+/// Append a mode-specific reminder to `preamble` based on the active prompt
+/// name. `plan_exists` reports whether `PLAN.md` is present in CWD — only
+/// consulted for the `code` mode reminder. Unknown prompt names produce no
+/// reminder so custom prompts don't accidentally pick up plan/review semantics.
+pub(crate) fn append_mode_reminder(preamble: &mut String, prompt_name: &str, plan_exists: bool) {
+    match prompt_name {
+        "plan" => {
+            preamble.push_str("\n\n---\n\nYou are now in PLAN mode. Create a detailed implementation plan. Save it to PLAN.md in the current directory. Analyze the task, break it into concrete steps, consider edge cases and trade-offs. Do NOT write any code or run any commands until the user reviews and approves the plan.");
+        }
+        "review" | "review-security" => {
+            preamble.push_str("\n\n---\n\nYou are now in REVIEW mode. Review the code or plan carefully. Identify bugs, security issues, performance problems, and design flaws. Be thorough and specific. Provide actionable feedback.");
+        }
+        "code" if plan_exists => {
+            preamble.push_str(
+                "\n\n---\n\nA plan file exists at PLAN.md. Execute the plan step by step. Write and test code following the plan. Report progress after each step. The plan is your guide — follow it closely.",
+            );
+        }
+        _ => {}
+    }
+}
+
+#[cfg(test)]
+mod reminder_tests {
+    use super::append_mode_reminder;
+
+    #[test]
+    fn plan_mode_injects_plan_reminder() {
+        let mut p = String::from("base");
+        append_mode_reminder(&mut p, "plan", false);
+        assert!(p.contains("PLAN mode"));
+        assert!(p.contains("PLAN.md"));
+        assert!(p.contains("Do NOT write any code"));
+    }
+
+    #[test]
+    fn review_modes_inject_review_reminder() {
+        for mode in &["review", "review-security"] {
+            let mut p = String::from("base");
+            append_mode_reminder(&mut p, mode, false);
+            assert!(p.contains("REVIEW mode"), "mode={mode}");
+            assert!(p.contains("Identify bugs"), "mode={mode}");
+        }
+    }
+
+    // Regression: the `code` reminder must only appear when PLAN.md exists.
+    // Without that guard every code-mode session would have a stale "execute
+    // the plan" instruction even with no plan written.
+    #[test]
+    fn regression_code_mode_reminder_requires_plan_md() {
+        let mut p_with = String::from("base");
+        append_mode_reminder(&mut p_with, "code", true);
+        assert!(p_with.contains("plan file exists"));
+
+        let mut p_without = String::from("base");
+        append_mode_reminder(&mut p_without, "code", false);
+        assert_eq!(p_without, "base", "no reminder must be added");
+    }
+
+    // Unknown prompts (custom user prompts) must produce no reminder so the
+    // plan/review semantics don't bleed into other modes.
+    #[test]
+    fn unknown_prompt_name_appends_nothing() {
+        let mut p = String::from("base");
+        append_mode_reminder(&mut p, "my-custom-prompt", true);
+        assert_eq!(p, "base");
+    }
+
+    // Each reminder is prefixed by the section separator so it visually
+    // detaches from the prior prompt — regression-guards the leading "\n\n---".
+    #[test]
+    fn reminders_use_section_separator() {
+        let mut p = String::new();
+        append_mode_reminder(&mut p, "plan", false);
+        assert!(p.starts_with("\n\n---\n\n"), "got: {p:?}");
+    }
+}