feat(bash): rework background shells to the Claude-Code model (unbounded + read/kill by id)

docs/features.md

@@ -20,7 +20,7 @@ headline differentiators, see the top-level [README](../README.md).
 - **Prompts system**: switch between system prompt modes at runtime (`code`, `plan`, `review`, `debug`, etc.). See [prompts.md](prompts.md).
 - **Per-prompt tool restrictions**: each prompt (`prompts/<name>.md`) can declare a `deny_tools` frontmatter list. The permission checker refuses those tools while the prompt is active — a real security boundary, not a prose gate. `plan`, `review`, and `review-security` ship with `edit`, `write`, `apply_patch`, `bash`, and `webfetch` denied. See [prompts.md](prompts.md).
 - **Subagent support**: `task` tool spawns a subagent for research or general analysis subtasks.
-- **Background shells**: the `bash` tool accepts `background: true` to run a command detached — it returns immediately with a shell id and the output is delivered automatically when the command finishes (the same completion-notification channel background subagents use; `task_status` can poll a shell id too). Useful for long-running builds, servers, and watchers that shouldn't block the turn. Tracked in the shared background registry and capped per kind. The status bar shows live counts when any are running: `agents:N` (background subagents) and `shells:N` (background shells).
+- **Background shells** (Claude-Code-style): the `bash` tool accepts `background: true` to run a command **detached and unbounded** (for dev servers, watch builds, long-running jobs) — it returns immediately with a shell id. The model reads accumulated output incrementally with the **`bash_output`** tool and stops the shell with **`kill_shell`** (both take the id); an optional `timeout` auto-kills after N seconds. Shells are tracked in a dedicated registry, capped at 8 concurrent, listed by `/tasks`, and killed when the session ends. The status bar shows live counts when any are running: `agents:N` (background subagents) and `shells:N` (background shells).
 - **Memory & self-improvement**: persistent per-project memory at `.dirge/memory/` (`MEMORY.md` facts + `PITFALLS.md`), injected into the system prompt as a frozen snapshot. The agent edits it via the `memory` tool. After an idle session, a unified post-session orchestrator runs (in order, fire-and-forget): a background review that extracts learnings into memory + skills, a skills curator and a memory curator (stale-detection + lifecycle + LLM consolidation, with audit reports under each store's `.curator_reports/`), and a cross-session pass that promotes sub-threshold patterns recurring across past sessions.
 - **MCP support**: connect MCP servers for extended tooling (optional compile-time feature).
 - **Git worktrees**: `/worktree` to create branch-per-task worktrees, `/wt-merge`, `/wt-exit`.

src/agent/agent_loop/integration.rs

@@ -1058,10 +1058,7 @@ mod tests {
         use crate::agent::tools::background::{BackgroundStore, TaskState};
 
         let store = BackgroundStore::new();
-        store.insert(
-            "sub-1".into(),
-            crate::agent::tools::background::TaskKind::Subagent,
-        );
+        store.insert("sub-1".into());
 
         let saw_reminder = Arc::new(Mutex::new(false));
         let saw_clone = saw_reminder.clone();

src/agent/builder.rs

@@ -376,8 +376,10 @@ pub async fn build_agent_inner<M: CompletionModel + 'static>(
                     sandbox.clone(),
                     cache.clone(),
                 )
-                .with_bg_store(bg_store.clone()),
+                .with_shell_store(Some(tools::bg_shell::global())),
             ),
+            Box::new(tools::BashOutputTool::new(tools::bg_shell::global())),
+            Box::new(tools::KillShellTool::new(tools::bg_shell::global())),
             Box::new(tools::GrepTool::with_cache(
                 permission.clone(),
                 ask_tx.clone(),
@@ -759,11 +761,13 @@ pub async fn build_loop_tools(
                 sandbox.clone(),
                 cache.clone(),
             )
-            .with_bg_store(bg_store.clone()),
+            .with_shell_store(Some(tools::bg_shell::global())),
             Some(ToolExecutionMode::Sequential),
         )
         .await,
     );
+    tools.push(wrap(tools::BashOutputTool::new(tools::bg_shell::global()), None).await);
+    tools.push(wrap(tools::KillShellTool::new(tools::bg_shell::global()), None).await);
 
     // Read-only batch.
     tools.push(