Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors
Rust
fix(llm): drain the cursor agent's pipes together and keep its workspace private
AkitaOnRailsf29e92a840f0
diff --git a/crates/ai-memory-llm/src/cursor_agent.rs b/crates/ai-memory-llm/src/cursor_agent.rsindex 2e9be51d..e6b3c1a4 100644--- a/crates/ai-memory-llm/src/cursor_agent.rs+++ b/crates/ai-memory-llm/src/cursor_agent.rs@@ -108,8 +108,15 @@ impl CursorAgentProvider {.ok_or_else(|| LlmError::UnexpectedShape("cursor agent stderr missing".into()))?;let collected = timeout(self.timeout, async {- let stdout = read_capped(stdout, MAX_STDOUT_BYTES).await?;- let stderr = read_capped(stderr, MAX_STDERR_BYTES).await?;+ // Both pipes drain at once: reading stdout to EOF first would+ // leave a chatty child blocked on a full stderr pipe until the+ // timeout. `try_join!` also stops at the first capped stream+ // instead of waiting on the other one (the child is killed on+ // drop).+ let (stdout, stderr) = tokio::try_join!(+ read_capped(stdout, MAX_STDOUT_BYTES),+ read_capped(stderr, MAX_STDERR_BYTES)+ )?;let status = child.wait().await.map_err(|err| {LlmError::UnexpectedShape(format!("waiting for cursor agent: {err}"))})?;@@ -176,7 +183,12 @@ struct TempWorkspace {impl TempWorkspace {fn create() -> LlmResult<Self> {let path = std::env::temp_dir().join(format!("ai-memory-cursor-{}", Uuid::now_v7()));- std::fs::create_dir(&path).map_err(|err| {+ // The prompt is captured session text: keep the workspace private+ // to the server's user on a shared temp directory.+ let mut builder = std::fs::DirBuilder::new();+ #[cfg(unix)]+ std::os::unix::fs::DirBuilderExt::mode(&mut builder, 0o700);+ builder.create(&path).map_err(|err| {LlmError::UnexpectedShape(format!("creating cursor workspace: {err}"))})?;Ok(Self { path })diff --git a/crates/ai-memory-llm/tests/suite/cursor_provider.rs b/crates/ai-memory-llm/tests/suite/cursor_provider.rsindex 99c54e66..fddfd1e3 100644--- a/crates/ai-memory-llm/tests/suite/cursor_provider.rs+++ b/crates/ai-memory-llm/tests/suite/cursor_provider.rs@@ -105,3 +105,69 @@ async fn fenced_json_is_parsed() {.unwrap();assert_eq!(value["answer"], 1);}++fn script_agent(dir: &Path, body: &str) -> PathBuf {+ let path = dir.join("agent");+ std::fs::write(&path, format!("#!/bin/sh\n{body}\n")).unwrap();+ let mut perms = std::fs::metadata(&path).unwrap().permissions();+ perms.set_mode(0o755);+ std::fs::set_permissions(&path, perms).unwrap();+ path+}++/// A child that floods stderr past the pipe buffer before writing stdout+/// must not hang the call: both pipes drain together, and a stream over its+/// cap ends the call at once instead of waiting out the timeout.+#[tokio::test]+async fn a_stderr_flood_neither_hangs_nor_blocks_stdout() {+ let dir = TempDir::new();+ // 40 KiB of stderr, then the answer: within the cap, so it succeeds.+ let ok = script_agent(+ &dir.0,+ "head -c 40960 /dev/zero | tr '\\0' e >&2\nprintf 'OK\\n'",+ );+ let response = tokio::time::timeout(+ std::time::Duration::from_secs(20),+ provider(ok, "default").complete(ChatRequest::user_prompt("hi")),+ )+ .await+ .expect("a 40 KiB stderr must not hang the call")+ .unwrap();+ assert_eq!(response.text, "OK");++ // 200 KiB of stderr while stdout stays open: over the cap, so it fails+ // fast instead of blocking on the full stderr pipe.+ let flood = script_agent(+ &dir.0,+ "head -c 204800 /dev/zero | tr '\\0' e >&2\nsleep 30\nprintf 'late\\n'",+ );+ let err = tokio::time::timeout(+ std::time::Duration::from_secs(20),+ provider(flood, "default").complete(ChatRequest::user_prompt("hi")),+ )+ .await+ .expect("a stderr flood must not hang the call")+ .unwrap_err();+ assert!(err.to_string().contains("size cap"), "{err}");+}++/// The per-call workspace holds the prompt (captured session text), so it is+/// created private to the server's user.+#[tokio::test]+async fn the_prompt_workspace_is_private() {+ let dir = TempDir::new();+ let sink = dir.0.join("mode");+ let agent = script_agent(+ &dir.0,+ &format!(+ "ls -ld . > {}\nprintf 'OK\\n'",+ sh_single(&sink.display().to_string())+ ),+ );+ provider(agent, "default")+ .complete(ChatRequest::user_prompt("hi"))+ .await+ .unwrap();+ let listing = std::fs::read_to_string(&sink).unwrap();+ assert!(listing.starts_with("drwx------"), "{listing}");+}diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.mdindex 9664eb36..7afbd317 100644--- a/docs/ARCHITECTURE.md+++ b/docs/ARCHITECTURE.md@@ -1067,7 +1067,7 @@ stopwords or adding a ranking signal.**LLM provider env** (opt-in):```AI_MEMORY_LLM_PROVIDER anthropic | anthropic-oauth | openai | openai-oauth | codex | copilot |- gemini | openai-compat | opencode+ cursor | gemini | openai-compat | opencodeAI_MEMORY_LLM_MODEL optional when the provider has a default; e.g. claude-haiku-4-5, gpt-5.4-miniANTHROPIC_API_KEY / OPENAI_API_KEY / GEMINI_API_KEY / LLM_API_KEYAI_MEMORY_LLM_BASE_URL required for openai-compat (Ollama, vLLM); optional override for@@ -1108,6 +1108,7 @@ AI_MEMORY_LLM_HEADERS optional extra HTTP headers on every chat request, asAI_MEMORY_RERANKER optional `llm`; reranks project/scopes query candidatesCOPILOT_GITHUB_TOKEN optional GitHub token for copilotAI_MEMORY_CODEX_EXECUTABLE optional Codex executable; defaults to codex on PATH+AI_MEMORY_CURSOR_AGENT optional Cursor Agent CLI executable for cursor; defaults to agent on PATHGITHUB_COPILOT_API_TOKEN optional pre-minted Copilot API tokenCOPILOT_API_URL optional Copilot API base URL override```