fix: deliver prompt via stdin for non-Claude agents to avoid spawn ENAMETOOLONG on Windows (#15)

The composed prompt (system instructions + skill body + cwd hint + file
listing + user message) can easily exceed Windows' CreateProcess limit of
~32 KB when passed as a CLI argument via -p <string>.  This causes
spawn ENAMETOOLONG whenever Gemini CLI (or Codex, OpenCode, Cursor
Agent, Qwen) is selected on Windows — even for short user messages,
because the skill / design-system system prompt is folded in.

Fix: add promptViaStdin: true to every plain-text agent definition.
The daemon's /api/chat handler checks this flag, opens stdin as a pipe,
writes the composed prompt to it and closes the stream.  Claude Code is
unaffected — it still uses the -p argv path and a separate stream-json
parser.

docs/agent-adapters.md: update §5.5 Gemini CLI to document the stdin
delivery strategy, and update the Windows open-question note to reflect
the fix.

Co-authored-by: KNIGHTABDO <abdessamad.aabida-etu@etu.univh2c.ma>
Co-authored-by: pftom <1043269994@qq.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: lefarcen <20859779+lefarcen@users.noreply.github.com>

daemon/agents.js

@@ -154,9 +154,12 @@ export const AGENT_DEFS = [
       { id: 'medium', label: 'Medium' },
       { id: 'high', label: 'High' },
     ],
-    buildArgs: (prompt, _imagePaths, _extra, options = {}) => {
-      // Keep Codex in workspace-write sandbox while avoiding interactive
-      // permission prompts in terminal-less web UI.
+    // Prompt delivered via stdin (`codex exec -`) to avoid Windows
+    // `spawn ENAMETOOLONG` — CreateProcess caps argv at ~32 KB and the
+    // composed prompt easily exceeds that. `--full-auto` keeps Codex in
+    // its workspace-write sandbox while skipping interactive permission
+    // prompts in the no-TTY web UI.
+    buildArgs: (_prompt, _imagePaths, _extra, options = {}) => {
       const args = ['exec', '--full-auto'];
       if (options.model && options.model !== 'default') {
         args.push('--model', options.model);
@@ -166,9 +169,10 @@ export const AGENT_DEFS = [
         // is exposed as `model_reasoning_effort`.
         args.push('-c', `model_reasoning_effort="${options.reasoning}"`);
       }
-      args.push(prompt);
+      args.push('-');
       return args;
     },
+    promptViaStdin: true,
     streamFormat: 'plain',
   },
   {
@@ -181,14 +185,18 @@ export const AGENT_DEFS = [
       { id: 'gemini-2.5-pro', label: 'gemini-2.5-pro' },
       { id: 'gemini-2.5-flash', label: 'gemini-2.5-flash' },
     ],
-    buildArgs: (prompt, _imagePaths, _extra, options = {}) => {
+    // Gemini reads from stdin when `-p` is omitted and stdin is a pipe.
+    // Passing the full composed prompt as a CLI arg causes ENAMETOOLONG on
+    // Windows (CreateProcess limit ~32 KB) for any non-trivial prompt.
+    // `--yolo` skips interactive approval prompts in the no-TTY web UI.
+    buildArgs: (_prompt, _imagePaths, _extra, options = {}) => {
       const args = ['--yolo'];
       if (options.model && options.model !== 'default') {
         args.push('--model', options.model);
       }
-      args.push('-p', prompt);
       return args;
     },
+    promptViaStdin: true,
     streamFormat: 'plain',
   },
   {
@@ -208,14 +216,17 @@ export const AGENT_DEFS = [
       { id: 'openai/gpt-5', label: 'openai/gpt-5' },
       { id: 'google/gemini-2.5-pro', label: 'google/gemini-2.5-pro' },
     ],
-    buildArgs: (prompt, _imagePaths, _extra, options = {}) => {
+    // Prompt delivered via stdin (`opencode run -`) to avoid Windows
+    // `spawn ENAMETOOLONG` for large composed prompts.
+    buildArgs: (_prompt, _imagePaths, _extra, options = {}) => {
       const args = ['run'];
       if (options.model && options.model !== 'default') {
         args.push('--model', options.model);
       }
-      args.push(prompt);
+      args.push('-');
       return args;
     },
+    promptViaStdin: true,
     streamFormat: 'plain',
   },
   {
@@ -242,14 +253,18 @@ export const AGENT_DEFS = [
       { id: 'sonnet-4-thinking', label: 'sonnet-4-thinking' },
       { id: 'gpt-5', label: 'gpt-5' },
     ],
-    buildArgs: (prompt, _imagePaths, _extra, options = {}) => {
+    // Prompt delivered via stdin (`cursor-agent -`) to avoid Windows
+    // `spawn ENAMETOOLONG` for large composed prompts. `--force` skips
+    // interactive approval prompts in the no-TTY web UI.
+    buildArgs: (_prompt, _imagePaths, _extra, options = {}) => {
       const args = ['--force'];
       if (options.model && options.model !== 'default') {
         args.push('--model', options.model);
       }
-      args.push('-p', prompt);
+      args.push('-');
       return args;
     },
+    promptViaStdin: true,
     streamFormat: 'plain',
   },
   {
@@ -262,15 +277,18 @@ export const AGENT_DEFS = [
       { id: 'qwen3-coder-plus', label: 'qwen3-coder-plus' },
       { id: 'qwen3-coder-flash', label: 'qwen3-coder-flash' },
     ],
-    buildArgs: (prompt, _imagePaths, _extra, options = {}) => {
-      // Qwen Code is a Gemini-CLI fork and supports the same `--yolo` mode.
+    // Prompt delivered via stdin (`qwen -`) to avoid Windows
+    // `spawn ENAMETOOLONG` for large composed prompts. Qwen Code is a
+    // Gemini-CLI fork and supports the same `--yolo` non-interactive mode.
+    buildArgs: (_prompt, _imagePaths, _extra, options = {}) => {
       const args = ['--yolo'];
       if (options.model && options.model !== 'default') {
         args.push('--model', options.model);
       }
-      args.push('-p', prompt);
+      args.push('-');
       return args;
     },
+    promptViaStdin: true,
     streamFormat: 'plain',
   },
   {
@@ -322,7 +340,7 @@ export const AGENT_DEFS = [
   },
 ];
 
-function resolveOnPath(bin) {
+export function resolveOnPath(bin) {
   const exts =
     process.platform === 'win32'
       ? (process.env.PATHEXT || '.EXE;.CMD;.BAT').split(';')
@@ -412,6 +430,7 @@ function stripFns(def) {
   return rest;
 }
 
+
 export async function detectAgents() {
   const results = await Promise.all(AGENT_DEFS.map(probe));
   // Refresh the validation cache from whatever we just surfaced to the UI

daemon/server.js

@@ -1010,12 +1010,29 @@ export async function startServer({ port = 7456 } = {}) {
 
     let child;
     try {
+      // When the agent definition sets `promptViaStdin`, pipe the composed
+      // prompt through stdin instead of embedding it in argv. Bypasses the
+      // OS command-line length limit (Windows CreateProcess caps at ~32 KB)
+      // which causes `spawn ENAMETOOLONG` for any non-trivial prompt.
+      const stdinMode = def.promptViaStdin ? 'pipe' : 'ignore';
       child = spawn(resolvedBin, args, {
         env: { ...process.env },
-        stdio: ['ignore', 'pipe', 'pipe'],
+        stdio: [stdinMode, 'pipe', 'pipe'],
         cwd: cwd || undefined,
         shell: useShell,
       });
+      if (def.promptViaStdin && child.stdin) {
+        // EPIPE from a fast-exiting CLI (bad auth, missing model, exit on
+        // launch) would otherwise surface as an unhandled stream error and
+        // crash the daemon. Swallow it — the regular exit/close handlers
+        // below already route the underlying failure to SSE via stderr.
+        child.stdin.on('error', (err) => {
+          if (err.code !== 'EPIPE') {
+            send('error', { message: `stdin: ${err.message}` });
+          }
+        });
+        child.stdin.end(composed, 'utf8');
+      }
     } catch (err) {
       send('error', { message: `spawn failed: ${err.message}` });
       return res.end();

docs/agent-adapters.md

@@ -155,11 +155,19 @@ The adapter declares which strategy to use via `capabilities().nativeSkillLoadin
 
 ### 5.5 Gemini CLI
 
-- Invocation: `gemini --prompt "<prompt>" --cwd <dir>`.
-- Streaming: yes, but less structured tool events. Expect fewer surgical-edit capabilities.
+- Invocation: `gemini` with the composed prompt delivered via **stdin** (no `-p` flag).
+  Gemini CLI enters headless mode automatically when stdin is a pipe and no `-p` flag is
+  supplied — verified with `gemini@0.1.x`.
+- Streaming: yes, plain text to stdout.
 - Skill loading: prompt injection only.
 - Surgical edits: regenerate whole file.
-- **Gotcha:** Gemini's tool-use format is distinct; we translate our file-write tool to its `file_tool` equivalent.
+- **Gotcha — `spawn ENAMETOOLONG` on Windows:** Passing the full composed prompt as a
+  `-p <string>` CLI argument hits Windows' `CreateProcess` hard limit of ~32 KB for the
+  entire command line. The fix is to set `promptViaStdin: true` in the agent definition
+  and write the prompt to `child.stdin` after spawning. The daemon's `/api/chat` handler
+  checks this flag and opens stdin as a pipe instead of `'ignore'`.
+- **Gotcha:** Gemini's tool-use format is distinct; we translate our file-write tool to its
+  `file_tool` equivalent when that feature is implemented.
 
 ### 5.6 OpenCode / OpenClaw
 
@@ -263,5 +271,8 @@ Each adapter is a separate module so community contributions can add new ones wi
 
 - **Nested agents.** What if Claude Code's agent itself spawns a subagent? We receive events from the outer process only. v1 policy: surface only top-level events; summarize subagent activity as "sub-task" placeholder.
 - **Billing awareness.** Some agents bill per message, some per token. OD doesn't track cost in MVP; v1 adds an optional "usage" event from adapters that expose it.
-- **Windows support.** PATH scanning and `spawn` semantics differ on Windows. v1 targets macOS and Linux; Windows is best-effort.
+- **Windows support.** PATH scanning and `spawn` semantics differ on Windows. v1 targets
+  macOS and Linux; Windows is best-effort. Known issue fixed: `spawn ENAMETOOLONG` when
+  running Gemini CLI (and other plain-text agents) on Windows — resolved by routing the
+  composed prompt through stdin instead of as a CLI argument (see §5.5).
 - **Docker-contained agents.** Some users run Claude Code in a container. Adapter needs a "remote" mode — probably same interface but talks over SSH. Phase 2+.