1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
//! The `impl StdioMcpClient` request methods (spawn + JSON-RPC round-trips).
//!
//! Why: The spawn flow and per-method JSON-RPC request/response handling are
//! the bulk of the client; isolating them from the struct/types/Drop in
//! `mod.rs` keeps both files under the 500-line cap.
//! What: `StdioMcpClient::spawn`, `initialize`, `list_tools`, `call_tool`,
//! `ping`, `is_alive`, `respawn`, `ensure_alive` — and the low-level
//! `send`/`recv` helpers — dispatching through `build_initialize_request` /
//! `extract_result` defined in `mod.rs`.
//! Test: JSON-RPC framing is unit-tested in `stdio_mcp_client::tests`; the
//! full spawn flow has an `#[ignore]`d integration test.
use std::sync::atomic::{AtomicU64, Ordering};
use anyhow::{Context, Result, anyhow, bail};
use serde_json::{Value, json};
use tokio::io::{AsyncBufReadExt, AsyncWriteExt, BufReader, BufWriter};
use tokio::process::Command;
use tokio::time::timeout;
use tracing::{debug, warn};
use super::{
CALL_TIMEOUT, MCP_PROTOCOL_VERSION, McpTool, ServerInfo, StdioMcpClient,
build_initialize_request, extract_result, plugin_stderr_stdio,
};
impl StdioMcpClient {
/// Spawn `binary` with `args`, piping stdin/stdout (stderr is redirected
/// to a per-plugin log file so server logs don't pollute the parent's
/// terminal or MCP stdout stream).
///
/// Why: The MCP transport requires clean JSON on stdout, so plugin logs
/// MUST go to stderr. Redirecting stderr to a named file keeps the parent
/// console clean while still preserving logs for debugging. The
/// `client_name` parameter is caller-supplied so each consumer
/// (trusty-agents, trusty-console, etc.) advertises its own identity in
/// `clientInfo.name` during the `initialize` handshake — hard-coding the
/// library name here would mislead MCP server logs and any server-side
/// logic keyed on that field.
/// What: Returns an unconnected client with the handshake NOT yet sent.
/// Call `initialize` next to complete the MCP handshake.
/// Test: Indirectly via `#[ignore]`d e2e test; unit-test failure is
/// covered by `spawn_missing_binary_errors` in `mod.rs`. The
/// `initialize_envelope_is_well_formed` test verifies the supplied name
/// propagates to `clientInfo.name`.
pub async fn spawn(
binary: &str,
args: &[&str],
client_name: impl Into<String>,
) -> Result<Self> {
let mut child = Command::new(binary)
.args(args)
.stdin(std::process::Stdio::piped())
.stdout(std::process::Stdio::piped())
.stderr(plugin_stderr_stdio(binary))
.kill_on_drop(true)
.spawn()
.with_context(|| format!("failed to spawn MCP binary {binary}"))?;
let stdin = child
.stdin
.take()
.ok_or_else(|| anyhow!("child stdin not captured"))?;
let stdout = child
.stdout
.take()
.ok_or_else(|| anyhow!("child stdout not captured"))?;
Ok(Self {
child,
stdin: BufWriter::new(stdin),
stdout: BufReader::new(stdout),
next_id: AtomicU64::new(1),
binary: binary.to_string(),
args: args.iter().map(|s| s.to_string()).collect(),
client_name: client_name.into(),
})
}
/// Check whether the child process is still running.
///
/// Why: When the MCP child dies (crash, OOM, killed externally) and the
/// caller writes to its stdin, the write blocks until the 30s timeout —
/// causing 15-60s query latency. Probing `try_wait()` lets callers detect
/// the dead child cheaply and respawn before writing. (See issue #421.)
/// What: Returns `true` if `try_wait()` reports `Ok(None)` (still
/// running); `false` if the process has exited or `try_wait()` errored.
/// Test: `is_alive_returns_false_after_child_exits` verifies the false
/// path against a child that exits immediately; `ids_are_monotonic`
/// exercises the true path implicitly.
pub fn is_alive(&mut self) -> bool {
matches!(self.child.try_wait(), Ok(None))
}
/// Replace the dead child with a freshly spawned one and rerun the MCP
/// handshake. Used internally by `call_tool`/`list_tools`/`ping` when
/// `is_alive()` returns false.
///
/// Why: Avoids the 30s write-to-dead-stdin timeout (issue #421) by
/// transparently recovering the connection. Without this, every query
/// after a server crash stalls until timeout.
/// What: Spawns the same `binary` + `args` used at construction, swaps
/// in the new stdio handles, resets the request id counter, and runs
/// `initialize`. Returns Err if respawn or handshake fails.
/// Test: `call_tool_errors_when_respawn_unavailable` exercises the
/// failure path (binary no longer present).
async fn respawn(&mut self) -> Result<()> {
let args_ref: Vec<&str> = self.args.iter().map(String::as_str).collect();
let mut new_child = Command::new(&self.binary)
.args(&args_ref)
.stdin(std::process::Stdio::piped())
.stdout(std::process::Stdio::piped())
.stderr(plugin_stderr_stdio(&self.binary))
.kill_on_drop(true)
.spawn()
.with_context(|| format!("failed to respawn MCP binary {}", self.binary))?;
let stdin = new_child
.stdin
.take()
.ok_or_else(|| anyhow!("child stdin not captured on respawn"))?;
let stdout = new_child
.stdout
.take()
.ok_or_else(|| anyhow!("child stdout not captured on respawn"))?;
// Best-effort reap of the old child before replacing it.
let _ = self.child.start_kill();
self.child = new_child;
self.stdin = BufWriter::new(stdin);
self.stdout = BufReader::new(stdout);
// client_name is retained from the original spawn — no update needed.
self.next_id.store(1, Ordering::Relaxed);
self.initialize()
.await
.context("MCP initialize failed after respawn")?;
Ok(())
}
/// Ensure the child is alive, respawning it if necessary.
///
/// Why: Centralised pre-flight for every write path so callers don't have
/// to remember to probe before each request.
/// What: If `is_alive()` returns false, logs a warning and calls
/// `respawn()`. Returns Err if respawn fails.
/// Test: Covered indirectly by `call_tool_errors_when_respawn_unavailable`.
async fn ensure_alive(&mut self) -> Result<()> {
if !self.is_alive() {
warn!("StdioMcpClient: child process exited, attempting respawn");
self.respawn().await?;
}
Ok(())
}
/// Allocate the next JSON-RPC request id. Monotonic, starts at 1.
///
/// Why: JSON-RPC 2.0 requires each request to carry a unique id so
/// responses can be correlated. The atomic counter provides this without
/// locks.
/// What: Atomically increments and returns the previous value using
/// `Relaxed` ordering (ordering across threads is not required for id
/// uniqueness within a single connection).
/// Test: `ids_are_monotonic` verifies the strict ordering property.
pub(super) fn alloc_id(&self) -> u64 {
self.next_id.fetch_add(1, Ordering::Relaxed)
}
/// Send the MCP `initialize` request and follow up with the `initialized`
/// notification per the protocol spec.
///
/// Why: MCP requires this two-step handshake before any other method may
/// be called. Doing it here means callers get a ready-to-use client.
/// What: Sends `initialize` with our protocol version + minimal client
/// info, parses `serverInfo`/`protocolVersion`, then sends the
/// `initialized` notification (no response expected).
/// Test: Envelope construction covered in `build_initialize_request`;
/// e2e in the ignored integration test.
pub async fn initialize(&mut self) -> Result<ServerInfo> {
let id = self.alloc_id();
let req = build_initialize_request(id, &self.client_name);
let resp = self.request(&req).await?;
let result = extract_result(resp)?;
let server_info = result
.get("serverInfo")
.ok_or_else(|| anyhow!("initialize response missing serverInfo"))?;
let name = server_info
.get("name")
.and_then(|v| v.as_str())
.unwrap_or("")
.to_string();
let version = server_info
.get("version")
.and_then(|v| v.as_str())
.unwrap_or("")
.to_string();
let protocol_version = result
.get("protocolVersion")
.and_then(|v| v.as_str())
.unwrap_or(MCP_PROTOCOL_VERSION)
.to_string();
// Send `initialized` notification (no id, no response).
let notif = json!({
"jsonrpc": "2.0",
"method": "notifications/initialized",
});
self.write_line(¬if).await?;
Ok(ServerInfo {
name,
version,
protocol_version,
})
}
/// Call `tools/list` and return the advertised tool descriptors.
///
/// Why: The console poller must enumerate available tools to verify the
/// service exposes the expected metrics tool before polling.
/// What: Sends `tools/list`, parses the `tools` array, and returns a
/// `Vec<McpTool>`. Calls `ensure_alive` first to auto-respawn dead children.
/// Test: End-to-end in trusty-agents integration tests.
pub async fn list_tools(&mut self) -> Result<Vec<McpTool>> {
self.ensure_alive().await?;
let id = self.alloc_id();
let req = json!({
"jsonrpc": "2.0",
"id": id,
"method": "tools/list",
});
let resp = self.request(&req).await?;
let result = extract_result(resp)?;
let tools = result
.get("tools")
.and_then(|v| v.as_array())
.ok_or_else(|| anyhow!("tools/list response missing tools array"))?;
let mut out = Vec::with_capacity(tools.len());
for t in tools {
let name = t
.get("name")
.and_then(|v| v.as_str())
.ok_or_else(|| anyhow!("tool entry missing name"))?
.to_string();
let description = t
.get("description")
.and_then(|v| v.as_str())
.map(|s| s.to_string());
let input_schema = t.get("inputSchema").cloned().unwrap_or_else(|| json!({}));
out.push(McpTool {
name,
description,
input_schema,
});
}
Ok(out)
}
/// Invoke `tools/call` with `name` and the given `params` as `arguments`.
///
/// Why: The primary method for a console or agent to invoke a service's
/// tools (e.g., `console_metrics` for health data).
/// What: Sends `tools/call`, parses the result, and returns the raw JSON
/// value for the caller to interpret. Calls `ensure_alive` first.
/// Test: End-to-end in trusty-agents integration tests.
pub async fn call_tool(&mut self, name: &str, params: Value) -> Result<Value> {
self.ensure_alive().await?;
let id = self.alloc_id();
let req = json!({
"jsonrpc": "2.0",
"id": id,
"method": "tools/call",
"params": { "name": name, "arguments": params },
});
let resp = self.request(&req).await?;
extract_result(resp)
}
/// Send a `ping` request — useful for liveness checks.
///
/// Why: Operators and supervisors need a cheap way to verify the child is
/// alive and responsive without triggering side effects.
/// What: Sends a `ping` request and discards the response. Calls
/// `ensure_alive` first to respawn if needed.
/// Test: Liveness check in integration tests.
pub async fn ping(&mut self) -> Result<()> {
self.ensure_alive().await?;
let id = self.alloc_id();
let req = json!({
"jsonrpc": "2.0",
"id": id,
"method": "ping",
});
let _ = self.request(&req).await?;
Ok(())
}
/// Write one JSON value followed by a newline, then flush.
///
/// Why: MCP uses newline-delimited JSON framing; every write must end with
/// `\n` so the server's line reader sees a complete frame. Flushing
/// ensures the bytes leave the buffer immediately.
/// What: Serialises `value` to bytes, writes them and a `\n`, then flushes
/// the buffered writer.
/// Test: Covered indirectly by all request tests.
async fn write_line(&mut self, value: &Value) -> Result<()> {
let bytes = serde_json::to_vec(value).context("serializing JSON-RPC frame")?;
self.stdin.write_all(&bytes).await?;
self.stdin.write_all(b"\n").await?;
self.stdin.flush().await?;
Ok(())
}
/// Read one newline-terminated JSON object from the child.
///
/// Why: Some MCP servers (e.g., trusty-memory) print a human-readable
/// status banner to stdout before the JSON-RPC stream begins. Treating
/// such lines as JSON-RPC frames would break the handshake. We skip any
/// line that does not start with `{` and log it at debug level — these
/// are expected, not warnings. (See issue #425.)
/// What: Loops reading lines, discarding non-JSON prefix lines, until a
/// line starting with `{` is found, then parses it as JSON.
/// Test: `read_line_skips_non_json_prefix_lines` feeds banner-then-JSON
/// through the codec and asserts the JSON object is returned.
pub(super) async fn read_line(&mut self) -> Result<Value> {
loop {
let mut line = String::new();
let n = self.stdout.read_line(&mut line).await?;
if n == 0 {
bail!("MCP server closed stdout before responding");
}
let trimmed = line.trim_start();
if !trimmed.starts_with('{') {
debug!(
"StdioMcpClient: skipping non-JSON line from child stdout: {:?}",
line.trim_end()
);
continue;
}
let value: Value = serde_json::from_str(line.trim_end())
.with_context(|| format!("parsing JSON-RPC frame: {line:?}"))?;
return Ok(value);
}
}
/// Send `req` and read responses until one matches the expected id.
/// Server-initiated notifications (no `id`) are ignored.
///
/// Why: JSON-RPC allows servers to send notifications (no id) at any time.
/// A simple read-one-frame approach would misinterpret a notification as
/// a response. This loop discards notifications and out-of-order ids
/// (rare) until the matching response arrives.
/// What: Wraps the round-trip in a 30s timeout. Returns Err if the timeout
/// fires or if the frame cannot be parsed.
/// Test: Covered by every method that calls `request`.
async fn request(&mut self, req: &Value) -> Result<Value> {
let expected_id = req
.get("id")
.cloned()
.ok_or_else(|| anyhow!("request must carry an id"))?;
timeout(CALL_TIMEOUT, async {
self.write_line(req).await?;
loop {
let frame = self.read_line().await?;
// Skip server-initiated notifications.
if frame.get("id").is_none() {
continue;
}
if frame.get("id") == Some(&expected_id) {
return Ok::<Value, anyhow::Error>(frame);
}
// Out-of-order id (rare). Continue reading.
}
})
.await
.map_err(|_| anyhow!("MCP request timed out after {:?}", CALL_TIMEOUT))?
}
}