@north-light/crouter 0.3.303 → 0.3.304

package/dist/pi-extensions/canvas-bash-valve.js

@@ -18,8 +18,9 @@
 // SIGPIPE, and its exit code is observable by nobody. So every command is
 // file-backed from birth via a tiny detached supervisor (a `bash -c` process
 // group leader) that redirects the real command's stdout+stderr to a log
-// file and records its exit code to a sentinel file — both readable, and
-// tail-able, independent of whether this pi process is even still alive.
+// file through a bounded collector and records its exit code to a sentinel
+// file — both readable and tail-able, independent of whether this pi process
+// is even still alive.
 //
 // Deviation from upstream: this backend pins `bash` as the shell (upstream
 // consults user shell config via non-exported helpers). Acceptable — crtr's
@@ -30,8 +31,9 @@ import { closeSync, existsSync, mkdirSync, openSync, readFileSync, readSync, rmS
 import { homedir } from 'node:os';
 import { dirname } from 'node:path';
 import { performance } from 'node:perf_hooks';
-import { backgroundBashJob, bashJobPaths, formatBashElapsed, newBashJobId, normalizeBashJobPurpose, readBashNiceness, writeJobDeadlineMs } from '../core/bash-jobs.js';
+import { backgroundBashJob, bashJobPaths, cancelBashJobProcessTree, formatBashElapsed, newBashJobId, normalizeBashJobPurpose, readBashNiceness, writeJobDeadlineMs } from '../core/bash-jobs.js';
 import { bashCommandRefusal } from '../core/bash-guard.js';
+import { capturePidIdentities } from '../core/canvas/pid.js';
 import { readConfig } from '../core/config.js';
 import { PREVIEW_RESULT_PATH_ENV, previewResultPath } from '../core/preview-result-path.js';
 import { resolveAdmittedHostCommand } from '../core/runtime/spawn-env.js';
@@ -43,18 +45,44 @@ const POLL_MS = 200;
 function resolveValveMs() {
     return envBashValveMs();
 }
-// The detached supervisor. Spawned with `detached: true` so it becomes its
-// own process-group leader (pgid === its pid) — killing `-pgid` reaches the
-// real command and anything it forked, even after this pi process exits.
-//
-// Positional argv (never interpolated into the script — the command text
-// lives only in cmd.sh):
+export const MAX_BACKGROUND_BASH_LOG_BYTES = 10 * 1024 * 1024;
+export const BACKGROUND_BASH_LOG_TRUNCATION_MARKER = '\n[background bash job output truncated at 10 MiB; remaining output discarded]\n';
+// This process owns one bounded prefix of the command stream. It keeps
+// draining after the cap so the command never sees a closed pipe/SIGPIPE, and
+// uses synchronous fixed-size stdin chunks plus direct file writes so neither
+// the collector nor the log can grow without bound.
+const LOG_COLLECTOR_SCRIPT = [
+    'const fs = require("node:fs");',
+    `const cap = ${MAX_BACKGROUND_BASH_LOG_BYTES};`,
+    `const marker = Buffer.from(${JSON.stringify(BACKGROUND_BASH_LOG_TRUNCATION_MARKER)});`,
+    'const prefixCap = cap - marker.length;',
+    'const fd = fs.openSync(process.argv[1], "w");',
+    'let written = 0;',
+    'let truncated = false;',
+    'process.stdin.on("data", (chunk) => {',
+    '  if (truncated) return;',
+    '  const available = prefixCap - written;',
+    '  if (available > 0) {',
+    '    const prefix = chunk.subarray(0, available);',
+    '    fs.writeSync(fd, prefix);',
+    '    written += prefix.length;',
+    '    if (prefix.length === chunk.length) return;',
+    '  }',
+    '  fs.writeSync(fd, marker);',
+    '  truncated = true;',
+    '});',
+    'process.stdin.on("end", () => fs.closeSync(fd));',
+].join(' ');
+// The detached supervisor owns the command and its bounded stream collector.
+// Positional argv (never interpolated into the script — the command text lives
+// only in cmd.sh):
 //   $1 cmd.sh     $2 job.log   $3 job.exit   $4 job.run
 //   $5 job.done   $6 job.bg    $7 jobId      $8 nodeId      $9 shell path
-//   ${10} niceness
+//   ${10} niceness ${11} node path
 //
 // `--` before the paths is a dummy $0 (a standard `bash -c script -- args`
 // idiom) so the real payload lands at $1 onward.
+//
 // The elapsed rendering mirrors formatBashElapsed() in core/bash-jobs.ts
 // (`12s` / `3m 40s` / `1h 04m`) so the completion message reads the same as the
 // handoff notice; it measures TOTAL runtime, foreground portion included.
@@ -64,8 +92,8 @@ function resolveValveMs() {
 const SUPERVISOR_SCRIPT = [
     'SECONDS=0',
     'n=$(command -v nice)',
-    '${n:+"$n" -n "${10}"} "${9}" "$1" > "$2" 2>&1',
-    'ec=$?',
+    '${n:+"$n" -n "${10}"} "${9}" "$1" 2>&1 | "${11}" -e ' + JSON.stringify(LOG_COLLECTOR_SCRIPT) + ' "$2"',
+    'ec=${PIPESTATUS[0]}',
     'd=$SECONDS',
     'echo "$ec" > "$3"',
     'if [ "$d" -lt 60 ]; then',
@@ -107,25 +135,9 @@ function installTerminationHandoff() {
     if (process.platform !== 'win32')
         process.prependListener('SIGHUP', backgroundLiveJobsForTermination);
 }
-function killGroup(pgid) {
-    try {
-        process.kill(-pgid, 'SIGTERM');
-    }
-    catch {
-        /* already gone */
-    }
-    const killer = setTimeout(() => {
-        try {
-            process.kill(-pgid, 'SIGKILL');
-        }
-        catch {
-            /* already gone */
-        }
-    }, 2000);
-    killer.unref?.();
-}
+const TAIL_CHUNK_BYTES = 64 * 1024;
 /** Read bytes appended to `path` since `offset`, feed them to onData, return
- *  the new offset. Best-effort — a not-yet-created log file is not an error. */
+ * the new offset. Best-effort — a not-yet-created log file is not an error. */
 function tailOnce(path, offset, onData) {
     let stat;
     try {
@@ -138,10 +150,16 @@ function tailOnce(path, offset, onData) {
         return offset;
     const fd = openSync(path, 'r');
     try {
-        const buf = Buffer.alloc(stat.size - offset);
-        readSync(fd, buf, 0, buf.length, offset);
-        onData(buf);
-        return stat.size;
+        while (offset < stat.size) {
+            const length = Math.min(TAIL_CHUNK_BYTES, stat.size - offset);
+            const buf = Buffer.allocUnsafe(length);
+            const bytesRead = readSync(fd, buf, 0, length, offset);
+            if (bytesRead === 0)
+                break;
+            onData(buf.subarray(0, bytesRead));
+            offset += bytesRead;
+        }
+        return offset;
     }
     finally {
         closeSync(fd);
@@ -211,6 +229,7 @@ export function createValveOperations(nodeId, contextDir, takePurpose = () => nu
                 let offset = 0;
                 let settled = false;
                 let pgid;
+                let pgidIdentity;
                 const startedAt = performance.now();
                 const valveMs = resolveValveMs();
                 let interval;
@@ -245,7 +264,7 @@ export function createValveOperations(nodeId, contextDir, takePurpose = () => nu
                     const notice = `\n[backgrounded as job ${paths.jobId} after ${elapsed} in the foreground${auto ? ' (auto)' : ''}. ` +
                         deadlineNotice +
                         `It will interrupt you with a message when it exits. ` +
-                        `Watch: tail -f ${paths.jobLog} ; stop: kill -- -${pgid}]`;
+                        `Watch: tail -f ${paths.jobLog} ; stop: crtr node bash kill ${paths.jobId}]`;
                     onData(Buffer.from(notice));
                     finish(() => resolve({ exitCode: 0 }));
                     // Backgrounded — job dir persists (inspectable); no cleanup here.
@@ -279,8 +298,8 @@ export function createValveOperations(nodeId, contextDir, takePurpose = () => nu
                     // finish the job.run → job.done transition before cleanup.
                     if (existsSync(paths.jobExit))
                         return;
-                    if (pgid)
-                        killGroup(pgid);
+                    if (pgid !== undefined && pgidIdentity !== undefined)
+                        void cancelBashJobProcessTree(pgid, pgidIdentity);
                     finish(() => reject(new Error('aborted')));
                     cleanupJobDir(paths.dir);
                 }
@@ -302,7 +321,7 @@ export function createValveOperations(nodeId, contextDir, takePurpose = () => nu
                     childEnv = { ...childEnv, [PREVIEW_RESULT_PATH_ENV]: previewPath };
                 }
                 const shellPath = resolveAdmittedHostCommand('bash', childEnv, execution.cwd);
-                const child = spawn(shellPath, ['-c', SUPERVISOR_SCRIPT, '--', paths.cmdSh, paths.jobLog, paths.jobExit, paths.jobRun, paths.jobDone, paths.jobBg, paths.jobId, nodeId, shellPath, String(readBashNiceness(contextDir))], {
+                const child = spawn(shellPath, ['-c', SUPERVISOR_SCRIPT, '--', paths.cmdSh, paths.jobLog, paths.jobExit, paths.jobRun, paths.jobDone, paths.jobBg, paths.jobId, nodeId, shellPath, String(readBashNiceness(contextDir)), process.execPath], {
                     cwd: execution.cwd,
                     detached: true,
                     stdio: 'ignore',
@@ -313,11 +332,22 @@ export function createValveOperations(nodeId, contextDir, takePurpose = () => nu
                     cleanupJobDir(paths.dir);
                 });
                 pgid = child.pid;
-                // The detached supervisor IS the group leader, so its pid is the pgid.
-                // Persisting it is what lets a human stop the job from the Inspector
-                // later; the agent's own handoff notice quotes the same number.
+                // The detached supervisor starts this process tree. Its launch-time
+                // identity lets cancellation reject a stale PID that now belongs to an
+                // unrelated process; it is not a containment guarantee.
                 if (pgid !== undefined) {
+                    pgidIdentity = capturePidIdentities([pgid])?.get(pgid);
+                    if (pgidIdentity === undefined) {
+                        try {
+                            process.kill(-pgid, 'SIGKILL');
+                        }
+                        catch { /* the just-started supervisor already exited */ }
+                        finish(() => reject(new Error(`bash job ${paths.jobId} started without a process identity`)));
+                        cleanupJobDir(paths.dir);
+                        return;
+                    }
                     writeFileSync(paths.jobPgid, String(pgid));
+                    writeFileSync(paths.jobPgidIdentity, pgidIdentity);
                     liveForegroundJobs.add(paths);
                 }
                 child.unref();
@@ -338,8 +368,8 @@ export function createValveOperations(nodeId, contextDir, takePurpose = () => nu
                     const elapsedMs = performance.now() - startedAt;
                     if (timeout !== undefined && timeout > 0) {
                         if (elapsedMs >= timeout * 1000) {
-                            if (pgid)
-                                killGroup(pgid);
+                            if (pgid !== undefined && pgidIdentity !== undefined)
+                                void cancelBashJobProcessTree(pgid, pgidIdentity);
                             finish(() => reject(new Error(`timeout:${timeout}`)));
                             cleanupJobDir(paths.dir);
                         }

package/dist/pi-extensions/canvas-inbox-watcher.js

@@ -344,8 +344,6 @@ class CanvasInboxWatcher {
      *  reports would render as bare refs instead of inlined bodies. One refresh
      *  per unknown node id keeps a genuinely missing node from stalling delivery. */
     refreshForUnknownReportRefs() {
-        if (this.reportNodesRefreshing)
-            return true;
         const unknown = [];
         for (const entry of this.buffer) {
             const refNodeId = reportRefNodeId(entry.ref);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@north-light/crouter",
-  "version": "0.3.303",
+  "version": "0.3.304",
   "description": "crtr — agent runtime with memory, plugins, and marketplaces",
   "type": "module",
   "main": "dist/index.js",

package/runtime.lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@north-light/crouter",
-  "version": "0.3.303",
+  "version": "0.3.304",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@north-light/crouter",
-      "version": "0.3.303",
+      "version": "0.3.304",
       "hasInstallScript": true,
       "license": "MIT",
       "workspaces": [
@@ -5206,17 +5206,17 @@
     },
     "packages/crouter-api": {
       "name": "@north-light/crouter-api",
-      "version": "0.3.303",
+      "version": "0.3.304",
       "license": "UNLICENSED"
     },
     "packages/crouter-env-docker": {
       "name": "@north-light/crouter-env-docker",
-      "version": "0.3.303",
+      "version": "0.3.304",
       "license": "UNLICENSED"
     },
     "packages/crouter-sdk": {
       "name": "@north-light/crouter-sdk",
-      "version": "0.3.303",
+      "version": "0.3.304",
       "license": "UNLICENSED",
       "dependencies": {
         "@north-light/crouter-api": "^0.3.295"