feat(dev): fix dot-worktree resolution, unique task files, and --agent/-a UX

.claude-plugin/skills/agent-cli-dev/SKILL.md

@@ -56,7 +56,7 @@ agent-cli dev new <branch-name> --from HEAD --agent --prompt-file path/to/prompt
 This creates:
 1. A new git worktree with its own branch
 2. Runs project setup (installs dependencies)
-3. Saves your prompt to `.claude/TASK.md` in the worktree (for reference)
+3. Saves your prompt to a unique task file in `.claude/` in the worktree (for reference)
 4. Opens a new terminal tab with an AI coding agent
 5. Passes your prompt to the agent
 
@@ -144,11 +144,12 @@ agent-cli dev agent review-auth -m tmux --prompt-file .claude/review-tests.md
 
 Key rules for same-worktree launches:
 - Use `dev agent`, not `dev new`, after the worktree already exists
+- Use `dev agent -a <agent>` to select a specific agent for an existing worktree; `--with-agent` remains a deprecated alias on this subcommand
 - Use `-m tmux` for headless or scripted launching; it works even when not already inside tmux
 - Each launch joins the same deterministic repo-scoped tmux session, so related agents stay grouped together
 - Ask each agent to write to a unique report path such as `.claude/REPORT-security-<run-id>.md` or `.claude/REPORT-tests-<run-id>.md`
 - If you rerun the same prompt repeatedly, include a timestamp or other run id in the report filename so later runs do not overwrite earlier ones
-- Do not rely on `.claude/TASK.md` as per-agent state in shared worktrees; later launches overwrite it
+- Each agent launch gets its own unique task file in `.claude/` (e.g., `TASK-{timestamp}-{hex}.md`), so parallel launches do not overwrite each other
 
 ### Prompt guidance for shared worktrees
 

.claude-plugin/skills/agent-cli-dev/examples.md

@@ -585,7 +585,7 @@ Write findings to .claude/REPORT-tests-$run_id.md:
 - If you rerun the same prompt often, include a timestamp or run id in the filename so reports do not get replaced
 - `-m tmux` works even when the caller is not already inside tmux
 - All three agents land in the same deterministic tmux session for that repo
-- `.claude/TASK.md` is shared state and may be overwritten by later launches, so keep prompt files outside that convention
+- Each agent launch gets its own unique task file in `.claude/`, so parallel launches do not conflict
 
 ## Scenario 7: Parallel test validation
 

.claude/skills/agent-cli-dev/SKILL.md

@@ -56,7 +56,7 @@ agent-cli dev new <branch-name> --from HEAD --agent --prompt-file path/to/prompt
 This creates:
 1. A new git worktree with its own branch
 2. Runs project setup (installs dependencies)
-3. Saves your prompt to `.claude/TASK.md` in the worktree (for reference)
+3. Saves your prompt to a unique task file in `.claude/` in the worktree (for reference)
 4. Opens a new terminal tab with an AI coding agent
 5. Passes your prompt to the agent
 
@@ -144,11 +144,12 @@ agent-cli dev agent review-auth -m tmux --prompt-file .claude/review-tests.md
 
 Key rules for same-worktree launches:
 - Use `dev agent`, not `dev new`, after the worktree already exists
+- Use `dev agent -a <agent>` to select a specific agent for an existing worktree; `--with-agent` remains a deprecated alias on this subcommand
 - Use `-m tmux` for headless or scripted launching; it works even when not already inside tmux
 - Each launch joins the same deterministic repo-scoped tmux session, so related agents stay grouped together
 - Ask each agent to write to a unique report path such as `.claude/REPORT-security-<run-id>.md` or `.claude/REPORT-tests-<run-id>.md`
 - If you rerun the same prompt repeatedly, include a timestamp or other run id in the report filename so later runs do not overwrite earlier ones
-- Do not rely on `.claude/TASK.md` as per-agent state in shared worktrees; later launches overwrite it
+- Each agent launch gets its own unique task file in `.claude/` (e.g., `TASK-{timestamp}-{hex}.md`), so parallel launches do not overwrite each other
 
 ### Prompt guidance for shared worktrees
 

.claude/skills/agent-cli-dev/examples.md

@@ -585,7 +585,7 @@ Write findings to .claude/REPORT-tests-$run_id.md:
 - If you rerun the same prompt often, include a timestamp or run id in the filename so reports do not get replaced
 - `-m tmux` works even when the caller is not already inside tmux
 - All three agents land in the same deterministic tmux session for that repo
-- `.claude/TASK.md` is shared state and may be overwritten by later launches, so keep prompt files outside that convention
+- Each agent launch gets its own unique task file in `.claude/`, so parallel launches do not conflict
 
 ## Scenario 7: Parallel test validation
 

agent_cli/dev/cli.py

@@ -373,7 +373,7 @@ def new(
         typer.Option(
             "--prompt",
             "-p",
-            help="Initial task for the AI agent. Saved to .claude/TASK.md. Implies --agent. Example: --prompt='Fix the login bug'",
+            help="Initial task for the AI agent. Saved to a unique file in .claude/ to avoid conflicts. Implies --agent. Example: --prompt='Fix the login bug'",
         ),
     ] = None,
     prompt_file: Annotated[
@@ -776,7 +776,9 @@ def status_cmd(  # noqa: PLR0915
 def remove(
     name: Annotated[
         str,
-        typer.Argument(help="Worktree to remove. Can be branch name or directory name"),
+        typer.Argument(
+            help="Worktree to remove. Can be branch name, directory name, or '.' for current"
+        ),
     ],
     force: Annotated[
         bool,
@@ -840,7 +842,9 @@ def remove(
 def path_cmd(
     name: Annotated[
         str,
-        typer.Argument(help="Worktree to get path for. Can be branch name or directory name"),
+        typer.Argument(
+            help="Worktree to get path for. Can be branch name, directory name, or '.' for current"
+        ),
     ],
 ) -> None:
     """Print the absolute path to a dev environment.
@@ -862,7 +866,9 @@ def path_cmd(
 def open_editor(
     name: Annotated[
         str,
-        typer.Argument(help="Worktree to open. Can be branch name or directory name"),
+        typer.Argument(
+            help="Worktree to open. Can be branch name, directory name, or '.' for current"
+        ),
     ],
     editor_name: Annotated[
         str | None,
@@ -911,7 +917,7 @@ def start_agent(
     name: Annotated[
         str,
         typer.Argument(
-            help="Worktree to start the agent in. Can be branch name or directory name",
+            help="Worktree to start the agent in. Can be branch name, directory name, or '.' for current",
         ),
     ],
     agent_name: Annotated[
@@ -922,6 +928,14 @@ def start_agent(
             help="Which agent: claude, codex, gemini, aider, copilot, cn, opencode, cursor-agent. Auto-detects if omitted",
         ),
     ] = None,
+    agent_name_deprecated: Annotated[
+        str | None,
+        typer.Option(
+            "--with-agent",
+            hidden=True,
+            help="[Deprecated: use --agent/-a] Which agent to start",
+        ),
+    ] = None,
     agent_args: Annotated[
         list[str] | None,
         typer.Option(
@@ -934,7 +948,7 @@ def start_agent(
         typer.Option(
             "--prompt",
             "-p",
-            help="Initial task for the agent. Saved to .claude/TASK.md. Example: --prompt='Add unit tests for auth'",
+            help="Initial task for the agent. Saved to a unique file in .claude/ to avoid conflicts. Example: --prompt='Add unit tests for auth'",
         ),
     ] = None,
     prompt_file: Annotated[
@@ -975,6 +989,11 @@ def start_agent(
     - `dev agent my-feature -a claude` — Start Claude specifically
     - `dev agent my-feature -p "Continue the auth refactor"` — Start with a task
     """
+    # Handle deprecated --with-agent alias
+    if agent_name_deprecated is not None:
+        warn("--with-agent is deprecated for 'dev agent', use --agent/-a instead")
+        agent_name = agent_name or agent_name_deprecated
+
     prompt = _resolve_prompt_text(prompt, prompt_file=prompt_file)
 
     repo_root = _ensure_git_repo()
@@ -1233,7 +1252,9 @@ def _doctor_check_git() -> None:
 def run_cmd(
     name: Annotated[
         str,
-        typer.Argument(help="Worktree to run command in. Can be branch name or directory name"),
+        typer.Argument(
+            help="Worktree to run command in. Can be branch name, directory name, or '.' for current"
+        ),
     ],
     command: Annotated[
         list[str],

agent_cli/dev/launch.py

@@ -168,14 +168,18 @@ def launch_editor(path: Path, editor: Editor) -> None:
 
 
 def write_prompt_to_worktree(worktree_path: Path, prompt: str) -> Path:
-    """Write the prompt to .claude/TASK.md in the worktree.
+    """Write the prompt to a unique file in .claude/ in the worktree.
 
-    This makes the task description available to the spawned agent
-    and provides a record of what was requested.
+    Uses a timestamp and random suffix to avoid overwrites when multiple
+    agents are launched in parallel on the same worktree.
     """
+    import time  # noqa: PLC0415
+
     claude_dir = worktree_path / ".claude"
     claude_dir.mkdir(parents=True, exist_ok=True)
-    task_file = claude_dir / "TASK.md"
+    timestamp = int(time.time())
+    suffix = os.urandom(2).hex()
+    task_file = claude_dir / f"TASK-{timestamp}-{suffix}.md"
     task_file.write_text(prompt + "\n")
     return task_file
 
@@ -201,8 +205,6 @@ def _create_prompt_wrapper_script(
     env: dict[str, str] | None = None,
 ) -> Path:
     """Create a wrapper script that reads prompt from file to avoid shell quoting issues."""
-    script_path = Path(tempfile.gettempdir()) / f"agent-cli-{worktree_path.name}.sh"
-
     # Build the agent command without the prompt
     exe = agent.get_executable()
     if exe is None:
@@ -222,7 +224,13 @@ def _create_prompt_wrapper_script(
 # Reads prompt from file to avoid shell parsing issues with special characters
 {env_prefix}exec {agent_cmd} "$(cat {shlex.quote(str(task_file_rel))})"
 """
-    script_path.write_text(script_content)
+    fd, script_path_str = tempfile.mkstemp(
+        prefix=f"agent-cli-{worktree_path.name}-",
+        suffix=".sh",
+    )
+    os.write(fd, script_content.encode())
+    os.close(fd)
+    script_path = Path(script_path_str)
     script_path.chmod(0o755)
     return script_path
 

agent_cli/dev/skill/SKILL.md

@@ -56,7 +56,7 @@ agent-cli dev new <branch-name> --from HEAD --agent --prompt-file path/to/prompt
 This creates:
 1. A new git worktree with its own branch
 2. Runs project setup (installs dependencies)
-3. Saves your prompt to `.claude/TASK.md` in the worktree (for reference)
+3. Saves your prompt to a unique task file in `.claude/` in the worktree (for reference)
 4. Opens a new terminal tab with an AI coding agent
 5. Passes your prompt to the agent
 
@@ -144,11 +144,12 @@ agent-cli dev agent review-auth -m tmux --prompt-file .claude/review-tests.md
 
 Key rules for same-worktree launches:
 - Use `dev agent`, not `dev new`, after the worktree already exists
+- Use `dev agent -a <agent>` to select a specific agent for an existing worktree; `--with-agent` remains a deprecated alias on this subcommand
 - Use `-m tmux` for headless or scripted launching; it works even when not already inside tmux
 - Each launch joins the same deterministic repo-scoped tmux session, so related agents stay grouped together
 - Ask each agent to write to a unique report path such as `.claude/REPORT-security-<run-id>.md` or `.claude/REPORT-tests-<run-id>.md`
 - If you rerun the same prompt repeatedly, include a timestamp or other run id in the report filename so later runs do not overwrite earlier ones
-- Do not rely on `.claude/TASK.md` as per-agent state in shared worktrees; later launches overwrite it
+- Each agent launch gets its own unique task file in `.claude/` (e.g., `TASK-{timestamp}-{hex}.md`), so parallel launches do not overwrite each other
 
 ### Prompt guidance for shared worktrees
 

agent_cli/dev/skill/examples.md

@@ -585,7 +585,7 @@ Write findings to .claude/REPORT-tests-$run_id.md:
 - If you rerun the same prompt often, include a timestamp or run id in the filename so reports do not get replaced
 - `-m tmux` works even when the caller is not already inside tmux
 - All three agents land in the same deterministic tmux session for that repo
-- `.claude/TASK.md` is shared state and may be overwritten by later launches, so keep prompt files outside that convention
+- Each agent launch gets its own unique task file in `.claude/`, so parallel launches do not conflict
 
 ## Scenario 7: Parallel test validation
 

agent_cli/dev/worktree.py

@@ -285,12 +285,39 @@ def resolve_worktree_base_dir(repo_root: Path) -> Path:
     return repo_root.parent / f"{repo_root.name}-worktrees"
 
 
+def _find_worktree_for_cwd(worktrees: list[WorktreeInfo]) -> WorktreeInfo | None:
+    """Find the worktree containing the current working directory."""
+    cwd = Path.cwd().resolve()
+    # Prefer the deepest matching path so nested layouts like .worktrees/<name>
+    # resolve to the actual worktree instead of the main repo.
+    for wt in sorted(
+        worktrees,
+        key=lambda worktree: len(worktree.path.resolve().parts),
+        reverse=True,
+    ):
+        try:
+            cwd.relative_to(wt.path.resolve())
+            return wt
+        except ValueError:
+            continue
+    # Fallback: return main worktree
+    return next((wt for wt in worktrees if wt.is_main), None)
+
+
 def find_worktree_by_name(
     name: str,
     repo_path: Path | None = None,
 ) -> WorktreeInfo | None:
-    """Find a worktree by branch name or directory name."""
+    """Find a worktree by branch name or directory name.
+
+    Use '.' to match the worktree containing the current working directory,
+    or the main worktree if the CWD is not inside any worktree.
+    """
     worktrees = list_worktrees(repo_path)
+
+    if name == ".":
+        return _find_worktree_for_cwd(worktrees)
+
     sanitized = sanitize_branch_name(name)
 
     for wt in worktrees:

docs/commands/dev.md

@@ -84,7 +84,7 @@ agent-cli dev new [BRANCH] [OPTIONS]
 | `--branch-name-timeout` | `20.0` | Timeout in seconds for AI branch naming command |
 | `--direnv/--no-direnv` | - | Generate .envrc based on project type and run 'direnv allow'. Auto-enabled if direnv is installed |
 | `--agent-args` | - | Extra CLI args for the agent. Can be repeated. Example: --agent-args='--dangerously-skip-permissions' |
-| `--prompt, -p` | - | Initial task for the AI agent. Saved to .claude/TASK.md. Implies --agent. Example: --prompt='Fix the login bug' |
+| `--prompt, -p` | - | Initial task for the AI agent. Saved to a unique file in .claude/ to avoid conflicts. Implies --agent. Example: --prompt='Fix the login bug' |
 | `--prompt-file, -P` | - | Read the agent prompt from a file. Useful for long prompts to avoid shell quoting. Implies --agent |
 | `--multiplexer, -m` | - | Launch the agent in a specific multiplexer. Currently supported: tmux. When started outside tmux, creates or reuses a detached session and reports the pane handle |
 | `--hooks/--no-hooks` | `true` | Run built-in agent preparation (like Codex auto-trust) and configured pre-launch hooks before starting the agent |
@@ -292,8 +292,9 @@ agent-cli dev agent NAME [--agent/-a AGENT] [--agent-args ARGS] [--prompt/-p PRO
 | Option | Default | Description |
 |--------|---------|-------------|
 | `--agent, -a` | - | Which agent: claude, codex, gemini, aider, copilot, cn, opencode, cursor-agent. Auto-detects if omitted |
+| `--with-agent` | - | [Deprecated: use --agent/-a] Which agent to start |
 | `--agent-args` | - | Extra CLI args for the agent. Example: --agent-args='--dangerously-skip-permissions' |
-| `--prompt, -p` | - | Initial task for the agent. Saved to .claude/TASK.md. Example: --prompt='Add unit tests for auth' |
+| `--prompt, -p` | - | Initial task for the agent. Saved to a unique file in .claude/ to avoid conflicts. Example: --prompt='Add unit tests for auth' |
 | `--prompt-file, -P` | - | Read the agent prompt from a file instead of command line |
 | `--multiplexer, -m` | - | Launch the agent in a specific multiplexer instead of the current terminal. Currently supported: tmux |
 | `--hooks/--no-hooks` | `true` | Run built-in agent preparation (like Codex auto-trust) and configured pre-launch hooks before starting the agent |
@@ -775,7 +776,7 @@ for section in 1 2 3 4; do
 done
 ```
 
-If multiple agents share one worktree, do not have them all write to `.claude/REPORT.md` because they will overwrite each other. Instead, assign unique report paths such as `.claude/REPORT-security-<run-id>.md` and `.claude/REPORT-tests-<run-id>.md`. If you rerun the same prompt repeatedly, use a timestamp or other run id so later runs do not replace earlier results. The same applies to `.claude/TASK.md`: it reflects the most recent launch, not stable per-agent state.
+If multiple agents share one worktree, do not have them all write to `.claude/REPORT.md` because they will overwrite each other. Instead, assign unique report paths such as `.claude/REPORT-security-<run-id>.md` and `.claude/REPORT-tests-<run-id>.md`. If you rerun the same prompt repeatedly, use a timestamp or other run id so later runs do not replace earlier results. Each agent launch also gets its own `.claude/TASK-{timestamp}-{hex}.md` file, so prompt files no longer overwrite each other.
 
 ## Shell Integration