refactor(mcp): rename start_claude_session to start_agent_session#1683
Merged
Conversation
…h multi-agent support Generalize the MCP tool to support multiple AI agents (claude, codex, gemini, opencode, copilot, cursor-agent) instead of being hardcoded to Claude Code.
Contributor
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
📝 WalkthroughWalkthroughRenames Claude-specific tooling to generic "agent" terminology, adds multi-agent support and presets, centralizes command generation in a new shared agent-command module, updates MCP start-session tooling and desktop UI to support selectable agents, and adds an opt-in local MCP setup step and docs updates. Changes
Sequence Diagram(s)sequenceDiagram
participant User
participant DesktopApp
participant SharedModule
participant MCP_Tool
participant Device
User->>DesktopApp: Click "Open in workspace" / start session
DesktopApp->>SharedModule: buildAgentCommand(task, randomId, agent)
SharedModule-->>DesktopApp: return agent command string
DesktopApp->>MCP_Tool: invoke start_agent_session with command payload
MCP_Tool->>Device: launch agent session with provided command
Device-->>MCP_Tool: session started / status
MCP_Tool-->>DesktopApp: tool execution result (session running)
DesktopApp-->>User: show "Agent session: running" status
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Poem
🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Contributor
🚀 Preview Deployment🔗 Preview Links
Preview updates automatically with new commits |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (2)
packages/mcp/src/tools/devices/start-agent-session/start-agent-session.ts (2)
99-104: Hardcoded agent list in the description string will drift ifAGENT_TYPESis updated.Line 103 enumerates agents as a literal string (
"claude", "codex", "gemini", ...). If a new agent is added toAGENT_TYPES, this description won't automatically reflect it. Consider generating the description from the constant, or leave a comment linking the two.♻️ Dynamic description
agent: z .enum(AGENT_TYPES) .optional() .describe( - 'AI agent to use: "claude", "codex", "gemini", "opencode", "copilot", or "cursor-agent". Defaults to "claude".', + `AI agent to use: ${AGENT_TYPES.map((t) => `"${t}"`).join(", ")}. Defaults to "claude".`, ),🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/mcp/src/tools/devices/start-agent-session/start-agent-session.ts` around lines 99 - 104, The description for the z.enum field `agent` is hardcoded and will drift from the `AGENT_TYPES` constant; change the `describe` call on the `agent: z.enum(AGENT_TYPES).optional()` schema to build its text from `AGENT_TYPES` (e.g., join the array into a string) or replace the literal list with a short reference comment linking to `AGENT_TYPES`, so the schema description always reflects the actual entries in `AGENT_TYPES`.
42-62: Narrow the return type ofvalidateArgsfromagent?: stringtoagent?: AgentType.
AgentTypeis properly exported from@superset/shared/agent-commandand the Zod schema already validatesagentas an enum member. Using the narrower type invalidateArgseliminates the need for the cast on line 112.Update the return type and the type assertion on line 53 to use
AgentType | undefined, then simplify line 112 from(validated.agent as (typeof AGENT_TYPES)[number]) ?? "claude"to justvalidated.agent ?? "claude".🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/mcp/src/tools/devices/start-agent-session/start-agent-session.ts` around lines 42 - 62, The validateArgs function currently types agent as string which forces a cast later; change its return type so agent is typed as AgentType | undefined (import AgentType from `@superset/shared/agent-command`), update the agent local extraction inside validateArgs to const agent = args.agent as AgentType | undefined, and return agent as AgentType when present; then remove the cast where AGENT_TYPES is used (replace `(validated.agent as (typeof AGENT_TYPES)[number]) ?? "claude"` with `validated.agent ?? "claude"`), keeping references to validateArgs, AgentType, and AGENT_TYPES.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@packages/shared/src/agent-command.ts`:
- Around line 54-64: The heredoc in buildHeredoc can be prematurely terminated
if the generated delimiter appears inside the prompt; update buildAgentCommand
(or the caller that generates the delimiter) to validate that the chosen
delimiter (e.g., SUPERSET_PROMPT_<id>) does not occur in the prompt/combined
task fields and, if it does, derive a new delimiter (for example by appending a
random suffix or regenerating the id) until it's unique, then pass that safe
delimiter into buildHeredoc so the heredoc body cannot be truncated; ensure the
check happens before calling buildHeredoc and use the same delimiter value for
both the opening and closing markers.
- Around line 78-82: The mapping for "cursor-agent" in the agent registry uses
an unsupported CLI flag (--yolo) which may produce an invalid runtime command;
check whether cursor-agent supports that flag by running cursor-agent --help and
then either remove the "--yolo" token from the buildHeredoc call in the
"cursor-agent" entry, replace it with a supported flag, or gate its inclusion
behind a feature flag/config so buildHeredoc(prompt, delimiter, "cursor-agent
-p", ...) only receives valid args; update any tests or documentation that
assume "--yolo" is present.
---
Nitpick comments:
In `@packages/mcp/src/tools/devices/start-agent-session/start-agent-session.ts`:
- Around line 99-104: The description for the z.enum field `agent` is hardcoded
and will drift from the `AGENT_TYPES` constant; change the `describe` call on
the `agent: z.enum(AGENT_TYPES).optional()` schema to build its text from
`AGENT_TYPES` (e.g., join the array into a string) or replace the literal list
with a short reference comment linking to `AGENT_TYPES`, so the schema
description always reflects the actual entries in `AGENT_TYPES`.
- Around line 42-62: The validateArgs function currently types agent as string
which forces a cast later; change its return type so agent is typed as AgentType
| undefined (import AgentType from `@superset/shared/agent-command`), update the
agent local extraction inside validateArgs to const agent = args.agent as
AgentType | undefined, and return agent as AgentType when present; then remove
the cast where AGENT_TYPES is used (replace `(validated.agent as (typeof
AGENT_TYPES)[number]) ?? "claude"` with `validated.agent ?? "claude"`), keeping
references to validateArgs, AgentType, and AGENT_TYPES.
Comment on lines
+54
to
+64
| function buildHeredoc( | ||
| prompt: string, | ||
| delimiter: string, | ||
| command: string, | ||
| suffix?: string, | ||
| ): string { | ||
| const closing = suffix ? `)" ${suffix}` : ')"'; | ||
| return [`${command} "$(cat <<'${delimiter}'`, prompt, delimiter, closing].join( | ||
| "\n", | ||
| ); | ||
| } |
Contributor
There was a problem hiding this comment.
Heredoc body containing the delimiter would truncate the prompt.
If any task field (title, description, etc.) happens to contain the exact string SUPERSET_PROMPT_<randomId_without_hyphens>, the heredoc terminates early. The UUID-derived delimiter makes accidental collision astronomically unlikely, but it's not impossible with adversarial or machine-generated task content. Consider validating that the delimiter doesn't appear in the prompt, or adjusting the delimiter if it does.
🛡️ Suggested guard in buildAgentCommand
export function buildAgentCommand({
task,
randomId,
agent = "claude",
}: {
task: TaskInput;
randomId: string;
agent?: AgentType;
}): string {
const prompt = buildPrompt(task);
- const delimiter = `SUPERSET_PROMPT_${randomId.replaceAll("-", "")}`;
+ let delimiter = `SUPERSET_PROMPT_${randomId.replaceAll("-", "")}`;
+ // Ensure the delimiter doesn't appear in the prompt body
+ while (prompt.includes(delimiter)) {
+ delimiter += "_";
+ }
const builder = AGENT_COMMANDS[agent];
return builder(prompt, delimiter);
}🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@packages/shared/src/agent-command.ts` around lines 54 - 64, The heredoc in
buildHeredoc can be prematurely terminated if the generated delimiter appears
inside the prompt; update buildAgentCommand (or the caller that generates the
delimiter) to validate that the chosen delimiter (e.g., SUPERSET_PROMPT_<id>)
does not occur in the prompt/combined task fields and, if it does, derive a new
delimiter (for example by appending a random suffix or regenerating the id)
until it's unique, then pass that safe delimiter into buildHeredoc so the
heredoc body cannot be truncated; ensure the check happens before calling
buildHeredoc and use the same delimiter value for both the opening and closing
markers.
Adds -m/--mcp flag to .superset/setup.sh that adds or updates a superset-local entry in .mcp.json pointing to the local API port.
…mplates - Add agent dropdown to OpenInWorkspace with icons and localStorage persistence - Extract AGENT_LABELS, AGENT_PRESET_COMMANDS, AGENT_PRESET_DESCRIPTIONS into packages/shared for reuse - Refactor PresetsSection to derive PRESET_TEMPLATES from shared agent metadata
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (1)
.superset/lib/setup/args.sh (1)
12-12: Align the description column with the other options.
--mcpis two characters shorter than--force, so the description starts two columns earlier, breaking the table layout.🎨 Alignment fix
- -m, --mcp Add superset-local MCP entry to .mcp.json + -m, --mcp Add superset-local MCP entry to .mcp.json🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In @.superset/lib/setup/args.sh at line 12, The "-m, --mcp" option line in args.sh is misaligned because the option text is shorter than others; update the "-m, --mcp Add superset-local MCP entry to .mcp.json" entry so its description column lines up with the other options (e.g., add the necessary spaces after "--mcp" to match the spacing used by "--force" and other entries), keeping the option flags and description text unchanged.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In @.superset/lib/setup/steps.sh:
- Around line 466-467: The code uses API_PORT with a default of 3001 which hides
the dynamically allocated workspace port because step_write_env sets API_PORT as
a local variable and it isn't exported; instead derive API_PORT from the
exported SUPERSET_PORT_BASE so non-first workspaces use the right port. Update
the logic in .superset/lib/setup/steps.sh (near step_load_env / step_write_env
usage) to compute API_PORT as $((SUPERSET_PORT_BASE + 1)) when
SUPERSET_PORT_BASE is set (fall back to existing API_PORT default only if
SUPERSET_PORT_BASE is unset), ensuring the MCP URL built from
local_url="http://localhost:${api_port}/api/agent/mcp" uses the correct
per-workspace port; reference functions/vars: allocate_port_base,
step_write_env, step_load_env, SUPERSET_PORT_BASE, and API_PORT.
- Around line 476-477: The mv of "$tmp_file" to "$mcp_file" is not checked so
success "Local MCP set to $local_url" can print even if mv failed; update the
block around the mv (the tmp_file -> mcp_file rename) to test mv's exit status
like allocate_port_base does: run mv "$tmp_file" "$mcp_file" and if it fails,
log an error (or call error) including the filenames, remove or preserve
"$tmp_file" as appropriate, and return a non-zero status instead of printing
success; ensure you reference the same variables tmp_file and mcp_file and
reproduce the error-handling pattern used in allocate_port_base.
---
Nitpick comments:
In @.superset/lib/setup/args.sh:
- Line 12: The "-m, --mcp" option line in args.sh is misaligned because the
option text is shorter than others; update the "-m, --mcp Add
superset-local MCP entry to .mcp.json" entry so its description column lines up
with the other options (e.g., add the necessary spaces after "--mcp" to match
the spacing used by "--force" and other entries), keeping the option flags and
description text unchanged.
Comment on lines
+466
to
+467
| local api_port="${API_PORT:-3001}" | ||
| local local_url="http://localhost:${api_port}/api/agent/mcp" |
Contributor
There was a problem hiding this comment.
API_PORT fallback of 3001 is incorrect for any non-first workspace.
step_write_env computes API_PORT as $((SUPERSET_PORT_BASE + 1)) but declares it local, so it is never exported to the calling shell. step_load_env sources the root .env, which either has a static default or no API_PORT at all. The result: this step always resolves to 3001 regardless of the dynamically allocated port base.
For workspace 1 (BASE=3000) this is coincidentally correct. For workspace 2 (BASE=3020, API=3021) the MCP URL will point to http://localhost:3001/api/agent/mcp — the wrong workspace's API — defeating the purpose of per-workspace port allocation.
SUPERSET_PORT_BASE is exported by allocate_port_base, so the fix is straightforward:
🐛 Proposed fix
- local api_port="${API_PORT:-3001}"
+ local api_port="${API_PORT:-$((${SUPERSET_PORT_BASE:-3000} + 1))}"📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| local api_port="${API_PORT:-3001}" | |
| local local_url="http://localhost:${api_port}/api/agent/mcp" | |
| local api_port="${API_PORT:-$((${SUPERSET_PORT_BASE:-3000} + 1))}" | |
| local local_url="http://localhost:${api_port}/api/agent/mcp" |
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In @.superset/lib/setup/steps.sh around lines 466 - 467, The code uses API_PORT
with a default of 3001 which hides the dynamically allocated workspace port
because step_write_env sets API_PORT as a local variable and it isn't exported;
instead derive API_PORT from the exported SUPERSET_PORT_BASE so non-first
workspaces use the right port. Update the logic in .superset/lib/setup/steps.sh
(near step_load_env / step_write_env usage) to compute API_PORT as
$((SUPERSET_PORT_BASE + 1)) when SUPERSET_PORT_BASE is set (fall back to
existing API_PORT default only if SUPERSET_PORT_BASE is unset), ensuring the MCP
URL built from local_url="http://localhost:${api_port}/api/agent/mcp" uses the
correct per-workspace port; reference functions/vars: allocate_port_base,
step_write_env, step_load_env, SUPERSET_PORT_BASE, and API_PORT.
Comment on lines
+476
to
+477
| mv "$tmp_file" "$mcp_file" | ||
| success "Local MCP set to $local_url" |
Contributor
There was a problem hiding this comment.
mv result is unchecked — success is reported even if the file was never updated.
If mv fails (permission error, cross-device rename, etc.), the original .mcp.json is unchanged, $tmp_file is left on disk, and the function still returns 0. Compare with the identical pattern in allocate_port_base (lines 290–295) which guards mv correctly.
🐛 Proposed fix
- mv "$tmp_file" "$mcp_file"
- success "Local MCP set to $local_url"
-
- return 0
+ if ! mv "$tmp_file" "$mcp_file"; then
+ error "Failed to update .mcp.json"
+ rm -f "$tmp_file"
+ return 1
+ fi
+ success "Local MCP set to $local_url"
+
+ return 0🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In @.superset/lib/setup/steps.sh around lines 476 - 477, The mv of "$tmp_file"
to "$mcp_file" is not checked so success "Local MCP set to $local_url" can print
even if mv failed; update the block around the mv (the tmp_file -> mcp_file
rename) to test mv's exit status like allocate_port_base does: run mv
"$tmp_file" "$mcp_file" and if it fails, log an error (or call error) including
the filenames, remove or preserve "$tmp_file" as appropriate, and return a
non-zero status instead of printing success; ensure you reference the same
variables tmp_file and mcp_file and reproduce the error-handling pattern used in
allocate_port_base.
Contributor
There was a problem hiding this comment.
🧹 Nitpick comments (2)
packages/shared/src/agent-command.ts (1)
21-30: Clarify the distinction betweenAGENT_PRESET_COMMANDSandAGENT_COMMANDS.These two maps serve different purposes but their naming is very similar, which could confuse future contributors:
AGENT_PRESET_COMMANDS— interactive/preset commands shown in the UI (e.g., codex with detailed-cflags).AGENT_COMMANDS— autonomous task-execution builders using heredoc prompts (e.g., codex with--full-auto).For instance,
codexmaps to a long detailed invocation inAGENT_PRESET_COMMANDSbut tocodex --full-autoinAGENT_COMMANDS. Consider adding a brief doc comment above each to clarify their intent, or renaming one (e.g.,AGENT_INTERACTIVE_COMMANDSvsAGENT_TASK_COMMANDS).Also applies to: 98-114
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/shared/src/agent-command.ts` around lines 21 - 30, Add clear comments or rename the two maps to remove ambiguity: update AGENT_PRESET_COMMANDS and AGENT_COMMANDS (and the equivalent map around the other block at the 98-114 region) to either include brief doc comments above each describing their purpose ("interactive/preset UI commands" vs "autonomous/task-execution commands") or rename them to explicit identifiers like AGENT_INTERACTIVE_COMMANDS and AGENT_TASK_COMMANDS; ensure all references/usages in the module are updated to match any renamed symbol so callers still compile.apps/desktop/src/renderer/routes/_authenticated/_dashboard/tasks/$taskId/components/PropertiesSidebar/components/OpenInWorkspace/OpenInWorkspace.tsx (1)
100-114:addTabcorrectly acceptsinitialCommands— optional refactor recommended for better UX.The existing code properly passes
initialCommands: [command]toaddTab, which is fully supported by theAddTabOptionsinterface. For improved clarity, useAGENT_LABELS[selectedAgent]instead of the generic"Agent"string for the tab title:♻️ Use agent-specific tab title
- setTabAutoTitle(tabId, "Agent"); + setTabAutoTitle(tabId, AGENT_LABELS[selectedAgent]);🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@apps/desktop/src/renderer/routes/_authenticated/_dashboard/tasks/`$taskId/components/PropertiesSidebar/components/OpenInWorkspace/OpenInWorkspace.tsx around lines 100 - 114, Replace the hardcoded tab title "Agent" in both branches with the agent-specific label: call setTabAutoTitle(tabId, AGENT_LABELS[selectedAgent]) after addTab(result.workspace.id, { initialCommands: [command] }) and when creating a pending setup via useWorkspaceInitStore.addPendingTerminalSetup ensure you set the tab title similarly; if selectedAgent might be undefined, fallback to AGENT_LABELS[selectedAgent] ?? "Agent" to preserve behavior.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Duplicate comments:
In
`@apps/desktop/src/renderer/routes/_authenticated/_dashboard/tasks/`$taskId/components/PropertiesSidebar/components/OpenInWorkspace/OpenInWorkspace.tsx:
- Around line 50-52: The initialization of selectedAgent in the OpenInWorkspace
component unsafely casts localStorage.getItem("lastSelectedAgent") to AgentType
causing AGENT_COMMANDS[selectedAgent] to possibly be undefined; change the
initializer to read the stored string, validate it against the known agent
enum/keys (e.g., keys of AGENT_COMMANDS or a canonical AGENT_TYPES array) and
only use it if it matches a valid AgentType, otherwise default to "claude";
apply the same validation whenever you read/update lastSelectedAgent (e.g., in
setSelectedAgent or any effect that syncs with localStorage) so you never store
or use an invalid AgentType.
- Line 33: The import in the OpenInWorkspace component uses a fragile relative
path for buildAgentCommand; change the import to use the renderer/ alias (e.g.,
import { buildAgentCommand } from "renderer/utils/buildAgentCommand") so it
follows the project's tsconfig alias convention and matches the other imports in
this file.
In `@packages/shared/src/agent-command.ts`:
- Around line 50-51: The code currently does an unsafe cast of
localStorage.getItem("lastSelectedAgent") to AgentType and accesses
AGENT_COMMANDS[selectedAgent], which can be undefined for stale/invalid values;
update the logic that initializes selectedAgent (and anything using it inside
buildPrompt) to first read the raw value from localStorage, check that it exists
in AGENT_TYPES (or Object.keys(AGENT_COMMANDS)), and only then cast to
AgentType; if the check fails, fall back to the default "claude" so
AGENT_COMMANDS[selectedAgent] is always defined and you avoid runtime TypeError.
- Around line 83-96: The heredoc in buildHeredoc can collide if prompt contains
the chosen delimiter; modify buildHeredoc to ensure the delimiter does not
appear inside prompt by testing prompt.includes(delimiter) and, if it does,
generate a new delimiter (e.g., append or re-roll a UUID fragment or suffix) in
a loop until unique before constructing the heredoc; keep the function signature
and the closing logic (closing variable and returned array) but replace the
single fixed delimiter selection with the uniqueness-check loop so the produced
heredoc delimiter is guaranteed not to occur in prompt.
---
Nitpick comments:
In
`@apps/desktop/src/renderer/routes/_authenticated/_dashboard/tasks/`$taskId/components/PropertiesSidebar/components/OpenInWorkspace/OpenInWorkspace.tsx:
- Around line 100-114: Replace the hardcoded tab title "Agent" in both branches
with the agent-specific label: call setTabAutoTitle(tabId,
AGENT_LABELS[selectedAgent]) after addTab(result.workspace.id, {
initialCommands: [command] }) and when creating a pending setup via
useWorkspaceInitStore.addPendingTerminalSetup ensure you set the tab title
similarly; if selectedAgent might be undefined, fallback to
AGENT_LABELS[selectedAgent] ?? "Agent" to preserve behavior.
In `@packages/shared/src/agent-command.ts`:
- Around line 21-30: Add clear comments or rename the two maps to remove
ambiguity: update AGENT_PRESET_COMMANDS and AGENT_COMMANDS (and the equivalent
map around the other block at the 98-114 region) to either include brief doc
comments above each describing their purpose ("interactive/preset UI commands"
vs "autonomous/task-execution commands") or rename them to explicit identifiers
like AGENT_INTERACTIVE_COMMANDS and AGENT_TASK_COMMANDS; ensure all
references/usages in the module are updated to match any renamed symbol so
callers still compile.
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@packages/shared/src/agent-command.ts`:
- Around line 21-30: AGENT_PRESET_COMMANDS entries diverge from the actual
runtime commands produced by buildAgentCommand (via AGENT_COMMANDS) for codex,
opencode, copilot, and cursor-agent; either update AGENT_PRESET_COMMANDS to
match the exact flags used by AGENT_COMMANDS/buildAgentCommand for those agents
or add a clear inline comment and UI-facing note that presets are templates and
additional flags (e.g., --full-auto, --prompt, -i --yolo, --yolo) are appended
at runtime by buildAgentCommand; locate AGENT_PRESET_COMMANDS and compare with
AGENT_COMMANDS and buildAgentCommand, then either synchronize the strings or add
the clarifying comment and ensure any UI text uses the clarified preset
description.
- Around line 108-109: The CLI invocation string passed to buildHeredoc is
incorrect: change the third argument in the opencode mapping from "opencode
--prompt" to the correct non-interactive command "opencode run" so that
opencode: (prompt, delimiter) => buildHeredoc(prompt, delimiter, "opencode run")
uses the documented CLI; locate the opencode property in
packages/shared/src/agent-command.ts and update the literal accordingly.
---
Duplicate comments:
In `@packages/shared/src/agent-command.ts`:
- Around line 83-96: The heredoc can break if the chosen delimiter appears
inside the prompt; update buildHeredoc to verify that the supplied delimiter
(and the generator used elsewhere, e.g.,
SUPERSET_PROMPT_${randomId.replaceAll("-", "")}) does not occur in prompt and,
if it does, generate a new unique delimiter (for example by appending a counter
or random suffix) until it is absent; then use that safe delimiter in the
returned heredoc string so prompt content cannot prematurely close the heredoc.
Comment on lines
+21
to
+30
| export const AGENT_PRESET_COMMANDS: Record<AgentType, string[]> = { | ||
| claude: ["claude --dangerously-skip-permissions"], | ||
| codex: [ | ||
| 'codex -c model_reasoning_effort="high" --ask-for-approval never --sandbox danger-full-access -c model_reasoning_summary="detailed" -c model_supports_reasoning_summaries=true', | ||
| ], | ||
| gemini: ["gemini --yolo"], | ||
| opencode: ["opencode"], | ||
| copilot: ["copilot"], | ||
| "cursor-agent": ["cursor-agent"], | ||
| }; |
Contributor
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
# First, locate and view the full agent-command.ts file to verify the divergence
fd -e ts -p "agent-command.ts" 2>/dev/null | head -5Repository: superset-sh/superset
Length of output: 100
🏁 Script executed:
# Search for AGENT_PRESET_COMMANDS usage across the codebase
rg "AGENT_PRESET_COMMANDS" --type ts -B 2 -A 3Repository: superset-sh/superset
Length of output: 2731
🏁 Script executed:
# Search for AGENT_COMMANDS and buildAgentCommand to understand the implementation
rg "AGENT_COMMANDS|buildAgentCommand" --type ts -B 2 -A 5 packages/shared/src/agent-command.tsRepository: superset-sh/superset
Length of output: 786
🏁 Script executed:
cat packages/shared/src/agent-command.tsRepository: superset-sh/superset
Length of output: 4043
AGENT_PRESET_COMMANDS and AGENT_COMMANDS diverge for codex, opencode, copilot, and cursor-agent.
AGENT_PRESET_COMMANDS is displayed as example commands in the Terminal Settings presets UI, but does not match what buildAgentCommand actually executes via AGENT_COMMANDS:
| Agent | AGENT_PRESET_COMMANDS |
AGENT_COMMANDS (actual runtime) |
|---|---|---|
codex |
codex -c model_reasoning_effort="high" --ask-for-approval never --sandbox danger-full-access … |
codex --full-auto |
opencode |
opencode |
opencode --prompt |
copilot |
copilot |
copilot -i --yolo |
cursor-agent |
cursor-agent |
cursor-agent --yolo |
The codex divergence is the most severe — these are fundamentally different invocation styles. Users copying the example from settings will not execute the same command that the system runs. Either align these constants, or add a clarifying comment explaining that presets are templates and the actual execution adds required flags.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@packages/shared/src/agent-command.ts` around lines 21 - 30,
AGENT_PRESET_COMMANDS entries diverge from the actual runtime commands produced
by buildAgentCommand (via AGENT_COMMANDS) for codex, opencode, copilot, and
cursor-agent; either update AGENT_PRESET_COMMANDS to match the exact flags used
by AGENT_COMMANDS/buildAgentCommand for those agents or add a clear inline
comment and UI-facing note that presets are templates and additional flags
(e.g., --full-auto, --prompt, -i --yolo, --yolo) are appended at runtime by
buildAgentCommand; locate AGENT_PRESET_COMMANDS and compare with AGENT_COMMANDS
and buildAgentCommand, then either synchronize the strings or add the clarifying
comment and ensure any UI text uses the clarified preset description.
Comment on lines
+108
to
+109
| opencode: (prompt, delimiter) => | ||
| buildHeredoc(prompt, delimiter, "opencode --prompt"), |
Contributor
There was a problem hiding this comment.
🧩 Analysis chain
🌐 Web query:
OpenCode CLI documentation "opencode run" non-interactive mode
💡 Result:
opencode run is the OpenCode CLI’s non-interactive mode: it runs a prompt directly (no TUI), which is intended for scripting/automation. (opencode.ai)
Basic usage
opencode run [message..]
Example:opencode run "Explain the use of context in Go"(opencode.ai)
Useful flags for non-interactive automation
--format jsonto output raw JSON events (machine-friendly). (opencode.ai)--file, -fattach file(s) to the message. (opencode.ai)--model, -mselect a model (provider/model). (opencode.ai)- Session controls:
--continue/-c,--session/-s,--fork,--title,--share. (opencode.ai)
Avoid cold starts (attach to a running server)
- Start a headless server:
opencode serve(opencode.ai) - Run with:
opencode run --attach http://localhost:4096 "Your prompt"(opencode.ai)
Citations:
- 1: https://opencode.ai/docs/cli/
- 2: https://opencode.ai/docs/cli/
- 3: https://opencode.ai/docs/cli/
- 4: https://opencode.ai/docs/cli/
- 5: https://opencode.ai/docs/cli/
- 6: https://opencode.ai/docs/cli/
- 7: https://opencode.ai/docs/cli/
- 8: https://opencode.ai/docs/cli/
opencode --prompt is not a valid CLI invocation — change to opencode run.
OpenCode's official CLI documentation specifies non-interactive mode as opencode run [message..], intended for scripting and automation. There is no --prompt top-level flag; the current invocation will fail with an unknown flag error at runtime.
Proposed fix
- opencode: (prompt, delimiter) =>
- buildHeredoc(prompt, delimiter, "opencode --prompt"),
+ opencode: (prompt, delimiter) =>
+ buildHeredoc(prompt, delimiter, "opencode run"),📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| opencode: (prompt, delimiter) => | |
| buildHeredoc(prompt, delimiter, "opencode --prompt"), | |
| opencode: (prompt, delimiter) => | |
| buildHeredoc(prompt, delimiter, "opencode run"), |
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@packages/shared/src/agent-command.ts` around lines 108 - 109, The CLI
invocation string passed to buildHeredoc is incorrect: change the third argument
in the opencode mapping from "opencode --prompt" to the correct non-interactive
command "opencode run" so that opencode: (prompt, delimiter) =>
buildHeredoc(prompt, delimiter, "opencode run") uses the documented CLI; locate
the opencode property in packages/shared/src/agent-command.ts and update the
literal accordingly.
- Guard against heredoc delimiter collision with prompt content - Derive API_PORT from SUPERSET_PORT_BASE for non-first workspaces - Add error handling for mv in local MCP setup - Validate localStorage agent type against AGENT_TYPES before use
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
start_claude_sessionMCP tool intostart_agent_sessionwith support for multiple AI agents: Claude, Codex, Gemini, OpenCode, Copilot, and Cursor Agentagentparameter (defaults to"claude") so existing callers are unaffectedpackages/shared/src/agent-command.tswith per-agent CLI buildersclaude-command.tsChanges
packages/sharedagent-command.ts—AGENT_TYPES,AgentType,TaskInput,buildAgentCommand(), deprecatedbuildClaudeCommand()wrapperclaude-command.ts— replaced with re-exports fromagent-command.tspackage.json— added"./agent-command"exportpackages/mcpstart-agent-session/tool — renamed fromstart-claude-session, addedagentparam validated againstAGENT_TYPESstart-claude-session/directoryindex.tsimport andget-workspace-detailsdescription referenceapps/desktopstart-claude-session.ts→start-agent-session.ts(tool + export)buildClaudeCommand.ts→buildAgentCommand.ts(re-exports from shared)OpenInWorkspace.tsxto usebuildAgentCommandindex.tsimportsDocs & commands
.claude/commands/task-run.md— referencesstart_agent_sessionapps/docs/content/docs/mcp.mdx— updated tool table, descriptions, and examples for multi-agent supportTest Plan
bun run typecheck— all 18 packages passbun run lint:fix— no lint errorsstart_claude_session,startClaudeSession,buildClaudeCommandonly appear in deprecated re-export pathsSummary by CodeRabbit
New Features
User Interface
Documentation