@waishnav/devspace 1.0.3 → 1.0.4

package/skills/subagent-delegation/SKILL.md

@@ -0,0 +1,132 @@
+---
+name: subagent-delegation
+description: Delegate coding tasks to user-configured DevSpace subagents.
+---
+
+# Subagent Delegation
+
+Use this skill when the user explicitly asks to delegate work to another coding
+agent, use a named subagent, get a second opinion, compare approaches, or run
+a subagent-like workflow.
+
+Do not use subagents silently. Tell the user when another subagent is
+being used.
+
+## Core commands
+
+Use only these commands for normal delegation:
+
+```bash
+devspace agents ls
+devspace agents run <profile-or-provider-or-id> "<prompt>"
+devspace agents show <id>
+```
+
+`ls` shows existing subagent sessions for the current workspace. DevSpace scopes
+it automatically from the shell environment injected by the workspace tool.
+
+`run <profile> "<prompt>"` starts a new configured profile and prints a
+DevSpace agent id.
+
+`run <provider> "<prompt>"` starts a raw built-in provider when no configured
+profile is needed. Built-in providers are listed by `open_workspace`.
+
+`run <id> "<prompt>"` sends a follow-up to an existing agent.
+
+`show <id>` prints status and the latest response. If the agent is still
+running, `show` waits briefly. If there is still no final response, call `show`
+again later.
+
+Do not run provider CLIs such as `codex`, `claude`, `opencode`, `pi`,
+`cursor-agent`, or `copilot` directly unless you are explicitly debugging
+DevSpace agent integration.
+
+## Choosing a profile
+
+Choose profiles from the compact subagent profile catalog returned by
+`open_workspace`. Use the profile name with `devspace agents run`. If no
+profile fits and delegation is still appropriate, use a built-in provider name
+from `open_workspace`.
+
+Profiles may declare a model and optional thinking level. To override the
+configured/default provider model or thinking level for a run, pass `--model`
+or `--thinking`:
+
+```bash
+devspace agents run <profile-or-provider> --model <model> "<prompt>"
+devspace agents run <profile-or-provider> --thinking <level> "<prompt>"
+```
+
+Use `--thinking` only when the user asks for a specific reasoning depth or when
+the task clearly needs a different effort than the configured profile default.
+Thinking values are provider-specific passthrough values. Use names supported by
+the selected local agent harness; DevSpace does not translate values between
+providers.
+
+Good delegation targets:
+
+- `reviewer`: second opinion, bug risk, security risk, test gaps.
+- `explorer`: read-only codebase investigation.
+- `implementer`: focused implementation when the user asked for delegation.
+
+Do not delegate ordinary coding work just because a profile exists. Use normal
+DevSpace tools unless the user asked for delegation, another agent's opinion,
+parallel work, or a named subagent.
+
+## Worker prompts
+
+Agents start with only the prompt you send plus their configured profile
+instructions. Make prompts self-contained.
+
+Implementation prompt shape:
+
+```text
+Goal:
+<clear goal>
+
+Context:
+<repo/module/user constraints>
+
+Relevant files:
+<paths and why they matter>
+
+Acceptance criteria:
+- <criterion>
+
+Rules:
+- Keep changes focused.
+- Do not perform unrelated refactors.
+- Report blockers clearly.
+```
+
+Read-only investigation prompt shape:
+
+```text
+Question:
+<specific question>
+
+Scope:
+<files/directories/modules to inspect>
+
+Rules:
+- Do not modify files.
+- Cite relevant file paths and symbols.
+- Separate facts from guesses.
+```
+
+## After the worker responds
+
+Always review the result before presenting it as verified.
+
+For write-capable tasks, inspect changed files and run or explain relevant
+tests. For read-only tasks, verify that important claims are supported by repo
+evidence.
+
+Be transparent in the final response:
+
+```text
+I used <profile>. It reported <summary>. I verified <checks>. Remaining risk:
+<risk or none>.
+```
+
+Never hide that a subagent was used.

package/docs/codex-tool-mode-qa.md

@@ -1,73 +0,0 @@
-# Codex Tool Mode Manual QA
-
-Run these checks against a disposable Git repository inside an allowed DevSpace
-root. Keep the DevSpace server logs visible during the test.
-
-## Setup
-
-1. Build the current branch with `npm ci && npm run build`.
-2. Start DevSpace with `DEVSPACE_TOOL_MODE=codex devspace serve`.
-3. Connect or refresh the DevSpace connector in ChatGPT.
-4. Open the disposable repository with `open_workspace`.
-5. Confirm the core tools are `open_workspace`, `read`, `apply_patch`,
-   `exec_command`, and `write_stdin`.
-6. Confirm `write`, `edit`, `bash`, `grep`, `glob`, and `ls` are absent.
-7. If `DEVSPACE_WIDGETS=changes`, also expect `show_changes`.
-
-## Apply Patch
-
-1. Add a text file containing multiple lines and a blank line.
-2. Update two separate regions of that file in one patch.
-3. Create a nested file, rename it, and then delete it.
-4. Patch an existing CRLF file and verify it remains CRLF.
-5. Verify executable permissions survive an update and a move.
-6. Try to add `../outside.txt`; confirm the tool rejects the path.
-7. Patch through a symlink targeting an external directory; confirm rejection.
-8. Submit a hunk whose context is absent; confirm no file from that patch changes.
-9. With changes widgets enabled, inspect the aggregate diff.
-
-## Foreground Commands
-
-1. Run `pwd` and confirm it reports the opened workspace.
-2. Run a command in a relative `workingDirectory` and confirm the directory.
-3. Write to stdout and stderr; confirm both appear.
-4. Exit nonzero; confirm `running=false` and the exit code.
-5. Use a small output budget on a noisy command; confirm truncation is reported.
-
-## Background Sessions
-
-1. Start a delayed command with a short yield time.
-2. Confirm `exec_command` returns `running=true` and a `sessionId`.
-3. Poll with empty `chars`; confirm output is not duplicated.
-4. Poll until completion; confirm the final exit code and no `sessionId`.
-5. Poll the completed session again; confirm it is unknown.
-6. Reconnect MCP without restarting DevSpace and confirm polling still works.
-7. Restart DevSpace and confirm old process session IDs are invalid.
-
-## Input, Interrupt, And PTY
-
-1. Start a program that reads stdin without a PTY and send it a line.
-2. Start a long-running process and send `\u0003`; confirm it stops.
-3. Start an interactive program with `tty=true`; confirm it detects a TTY.
-4. Resize a PTY from 80x24 to 120x30 and verify the observed dimensions.
-5. Omit optional dependencies; normal commands must work and `tty=true` must
-   return the explicit `node-pty` error.
-
-## Cleanup
-
-1. Start a non-PTY command that creates a long-running child process.
-2. Stop DevSpace with SIGINT and verify both shell and child exit.
-3. Repeat with a PTY command.
-4. Confirm no process remains after server exit.
-5. Repeat session cycles and check that memory use does not steadily increase.
-
-## Existing Mode Regression
-
-1. Start without `DEVSPACE_TOOL_MODE`; confirm `minimal` remains the default.
-2. Minimal must expose `read`, `write`, `edit`, and `bash`, but not Codex tools
-   or dedicated search tools.
-3. `DEVSPACE_TOOL_MODE=full` must add `grep`, `glob`, and `ls`.
-4. With no explicit mode, `DEVSPACE_MINIMAL_TOOLS=1` maps to minimal and `0`
-   maps to full.
-5. Set `DEVSPACE_TOOL_MODE=codex` with `DEVSPACE_MINIMAL_TOOLS=0`; confirm the
-   explicit Codex mode wins.