gm-skill 2.0.1972 → 2.0.1974

package/bin/plugkit.version

@@ -1 +1 @@
-0.1.903
+0.1.905

package/bin/plugkit.wasm.sha256

@@ -1 +1 @@
-23578c20939f8d090f7219836cb12ea187d4f184041bf0449e3eedab4840fafb  plugkit.wasm
+5e637cfefaddd54c2fdf06d6e718bc8834b6c026fed6e542ef1d18de2b5bebfc  plugkit.wasm

package/gm-plugkit/bootstrap.js

@@ -46,7 +46,7 @@ function obsEvent(subsystem, event, fields) {
 
 function writeBootstrapError(spec) {
   try {
-    const projectDir = process.env.CLAUDE_PROJECT_DIR || process.cwd();
+    const projectDir = resolveProjectRoot(process.env.CLAUDE_PROJECT_DIR || process.cwd());
     const spoolDir = path.join(projectDir, '.gm', 'exec-spool');
     fs.mkdirSync(spoolDir, { recursive: true });
     fs.writeFileSync(path.join(spoolDir, '.bootstrap-error.json'), JSON.stringify({ ts: new Date().toISOString(), ...spec }, null, 2));
@@ -55,7 +55,7 @@ function writeBootstrapError(spec) {
 
 function clearBootstrapError() {
   try {
-    const projectDir = process.env.CLAUDE_PROJECT_DIR || process.cwd();
+    const projectDir = resolveProjectRoot(process.env.CLAUDE_PROJECT_DIR || process.cwd());
     fs.unlinkSync(path.join(projectDir, '.gm', 'exec-spool', '.bootstrap-error.json'));
   } catch (_) {}
 }
@@ -223,6 +223,39 @@ function gmToolsDir() {
   return primary;
 }
 
+// Root a project-dir resolution at the git COMMON dir, not the raw cwd/
+// CLAUDE_PROJECT_DIR -- a worktree (e.g. Workflow's isolation:'worktree'
+// agents, each `git worktree add`-ing a fresh physical directory) shares the
+// SAME underlying repo as its main checkout but has its own separate
+// directory tree. Every cwd-derived project-dir computation in this file and
+// in supervisor.js/cli.js must funnel through this so a worktree-spawned
+// process resolves to the SAME .gm/exec-spool/ as its main-repo sibling --
+// otherwise the single-instance supervisor lock can never see a sibling
+// worktree's already-running watcher, and every worktree cold-boots its own
+// independent watcher (each loading the full embed model + running its own
+// cold reindex of what is, conceptually, the same project). Live-measured
+// this session under real concurrent multi-agent load: this was the actual
+// root cause of a user-flagged "memory grows to ~2GB then clears" churn
+// pattern -- N worktrees of one repo each independently paying full
+// cold-embed cost instead of sharing one already-warm watcher, and the
+// resulting CPU contention pushed genuinely-busy processes past their
+// heartbeat deadline into a restart cycle that produces the sawtooth memory
+// pattern. Mirrors the existing browserRootDir() resolution already used for
+// browser-session state in plugkit-wasm-wrapper.js (same fix, same reason,
+// applied one layer up at the process-boot-dedup level).
+function resolveProjectRoot(start) {
+  const resolved = path.resolve(start);
+  try {
+    const r = spawnSync('git', ['rev-parse', '--git-common-dir'], { cwd: resolved, encoding: 'utf-8', windowsHide: true, timeout: 1500 });
+    if (r.status === 0 && r.stdout && r.stdout.trim()) {
+      let commonDir = r.stdout.trim();
+      if (!path.isAbsolute(commonDir)) commonDir = path.resolve(resolved, commonDir);
+      if (/(^|[\\/])\.git$/.test(commonDir)) return path.dirname(commonDir);
+    }
+  } catch (_) {}
+  return resolved;
+}
+
 function readVersionFile() {
   const p = path.join(wrapperDir, 'plugkit.version');
   if (!fs.existsSync(p)) throw new Error(`plugkit.version not found at ${p}`);
@@ -1254,6 +1287,7 @@ module.exports = {
   ensureNextStepWiring,
   ensureInstructionsBundle,
   gmToolsDir,
+  resolveProjectRoot,
   getWasmPath,
   getBinaryPath,
   startSpoolDaemon,

package/gm-plugkit/cli.js

@@ -5,7 +5,7 @@ const fs = require('fs');
 const os = require('os');
 const path = require('path');
 const cp = require('child_process');
-const { ensureReady, startSpoolDaemon, gmToolsDir, readVersionFile, ensureGmPlugkitVersionFresh, ensureSkillMdFresh, ensureWrapperFresh, isReady, getWasmPath, readPinnedGmPlugkitVersion, spawnPinnedBoot } = require('./bootstrap');
+const { ensureReady, startSpoolDaemon, gmToolsDir, readVersionFile, ensureGmPlugkitVersionFresh, ensureSkillMdFresh, ensureWrapperFresh, isReady, getWasmPath, readPinnedGmPlugkitVersion, spawnPinnedBoot, resolveProjectRoot } = require('./bootstrap');
 
 function getWasmPathSafe() {
   try { return getWasmPath(); } catch (_) { return null; }
@@ -201,7 +201,7 @@ function killStaleWatchers() {
 }
 
 function spoolDir() {
-  const projectDir = process.env.CLAUDE_PROJECT_DIR || process.cwd();
+  const projectDir = resolveProjectRoot(process.env.CLAUDE_PROJECT_DIR || process.cwd());
   return path.join(projectDir, '.gm', 'exec-spool');
 }
 
@@ -242,6 +242,29 @@ function writeCliError(phase, err) {
   } catch (_) {}
 }
 
+// gm-runner is a genuine, sha256-verified drop-in replacement for this
+// entire bun/node boot path -- when it's installed, delegate to it
+// immediately and exit, before any bun/node-specific bootstrap logic runs.
+// This is the actual code-level enforcement of what was previously only a
+// documented convention (SKILL.md telling an LLM agent to manually prefer
+// gm-runner) -- with this in place bun/node are only ever exercised on a
+// platform gm-runner-bin has no published binary for, or during the
+// one-time install before gm-runner itself lands on disk.
+function tryDelegateToGmRunner(args) {
+  if (process.env.GM_PLUGKIT_NO_RUNNER_DELEGATE === '1') return false;
+  const exeName = process.platform === 'win32' ? 'gm-runner.exe' : 'gm-runner';
+  const runnerPath = path.join(gmToolsDir(), exeName);
+  if (!fs.existsSync(runnerPath)) return false;
+  try {
+    const result = cp.spawnSync(runnerPath, args, { stdio: 'inherit', windowsHide: true });
+    if (result.error) return false; // exec genuinely failed to start -- fall through to bun/node path
+    process.exit(typeof result.status === 'number' ? result.status : 0);
+  } catch (_) {
+    return false;
+  }
+  return true;
+}
+
 (async () => {
   const args = process.argv.slice(2);
 
@@ -250,6 +273,8 @@ function writeCliError(phase, err) {
     process.exit(0);
   }
 
+  tryDelegateToGmRunner(args);
+
   if (args.includes('--kill-stale-watchers')) {
     process.exit(killStaleWatchers());
   }

package/gm-plugkit/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gm-plugkit",
-  "version": "2.0.1972",
+  "version": "2.0.1974",
   "description": "Bootstrap and daemon-spawn tool for gm plugkit binary. Downloads the correct platform binary, verifies SHA256, and starts the spool watcher daemon. Includes plugkit-wasm-wrapper for WASM-based spool watching.",
   "main": "index.js",
   "bin": {

package/gm-plugkit/plugkit-wasm-wrapper.js

@@ -4732,9 +4732,22 @@ async function runSpoolWatcher(instance, spoolDir) {
     } catch (_) {}
   }
   _writeStatusBusy = (ms) => { try { writeStatus(ms); } catch (_) {} };
-  setInterval(() => writeStatus(), 5000);
+  // Every regular heartbeat tick also carries a modest rolling busy_until
+  // (45s, comfortably above the supervisor's 30s STATUS_STALE_MS) -- not
+  // because this tick itself is doing long work, but because under heavy
+  // real multi-process contention (many concurrent gm-driven watchers on
+  // one machine, live-observed: 13 concurrent agent watchers each competing
+  // for CPU, all doing real embed/codesearch work) the OS scheduler can
+  // starve THIS process of any time slice for well over 30s between ticks
+  // even though the process is genuinely alive and will resume once
+  // scheduled -- no in-process yield/setImmediate can fix a starvation that
+  // happens entirely outside this process's own event loop. A missed tick
+  // or two under real contention should not itself be a kill signal; a
+  // GENUINELY dead process still gets caught because nothing here ever
+  // refreshes busy_until once the process actually exits.
+  setInterval(() => writeStatus(45000), 5000);
   setInterval(() => { try { scanStalledTurns(); } catch (_) {} }, 30000);
-  writeStatus();
+  writeStatus(45000);
 
   setTimeout(async () => {
     try {

package/gm-plugkit/plugkit.version

@@ -1 +1 @@
-0.1.903
+0.1.905

package/gm-plugkit/supervisor.js

@@ -6,7 +6,7 @@ const path = require('path');
 const os = require('os');
 const crypto = require('crypto');
 const { spawn, spawnSync } = require('child_process');
-const { gmToolsDir } = require('./bootstrap');
+const { gmToolsDir, resolveProjectRoot } = require('./bootstrap');
 const { logEvent: _sharedLogEvent, GM_LOG_ROOT: _sharedGmLogRoot } = require('./gm-log');
 const { pidCommandLineForKillGuard: _sharedPidCommandLine } = require('./gm-process');
 
@@ -17,7 +17,11 @@ function wrapperSha12OnDisk() {
   } catch (_) { return null; }
 }
 
-const projectDir = process.env.CLAUDE_PROJECT_DIR || process.cwd();
+// Root the spool/lock state at the git common dir (see resolveProjectRoot's
+// own doc comment in bootstrap.js for the full worktree-dedup rationale) so
+// a worktree-spawned supervisor shares its lock/spoolDir with the main-repo
+// checkout instead of always cold-booting its own independent watcher.
+const projectDir = resolveProjectRoot(process.env.CLAUDE_PROJECT_DIR || process.cwd());
 const spoolDir = path.join(projectDir, '.gm', 'exec-spool');
 fs.mkdirSync(spoolDir, { recursive: true });
 
@@ -29,26 +33,7 @@ const LOG_PATH = path.join(spoolDir, '.watcher.log');
 const GM_LOG_ROOT = _sharedGmLogRoot;
 
 const POLL_INTERVAL_MS = 10_000;
-// Was 30s. Live-measured this session under real multi-agent concurrent load
-// (13 parallel gm-skill-driven agents, each booting its own watcher on the
-// same machine): 74 of 94 total watcher restarts logged in one session were
-// heartbeat-stale kills, the dominant restart cause by a wide margin -- each
-// restart drops the process's memory back to near-zero then it climbs again
-// through a fresh cold-embed pass, which is the actual mechanism behind the
-// "memory grows to ~2GB then clears" pattern a user observed and correctly
-// flagged as inefficient. The busy_until shielding mechanism (see writeStatus
-// call sites in plugkit-wasm-wrapper.js) already protects genuinely-busy
-// single dispatches, but under real CPU contention from many concurrent
-// processes, even the setInterval-based busy_until REFRESH callback itself
-// can be delayed past the old 30s window before it fires -- a starvation
-// problem one level above the per-dispatch busy-window fixes already shipped
-// this session. Raising the stale threshold gives real contention more slack
-// before concluding a process is dead, trading a slightly slower true-dead
-// detection for far fewer false-positive kills-and-restarts (each restart is
-// far more expensive than the extra wait -- a full cold embed pass observed
-// taking 90-120s+ under load) -- a clear net win under the contention this
-// session repeatedly measured.
-const STATUS_STALE_MS = 90_000;
+const STATUS_STALE_MS = 30_000;
 const MAX_RESTART_BURST = 5;
 const RESTART_WINDOW_MS = 60_000;
 const BURST_BACKOFF_MS = 60_000;

package/gm.json

@@ -1,6 +1,6 @@
 {
   "name": "gm",
-  "version": "2.0.1972",
+  "version": "2.0.1974",
   "description": "Spool-dispatch orchestration engine with unified state machine, skills, and automated git enforcement",
   "author": "AnEntrypoint",
   "license": "MIT",
@@ -17,5 +17,5 @@
   "publishConfig": {
     "access": "public"
   },
-  "plugkitVersion": "0.1.903"
+  "plugkitVersion": "0.1.905"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gm-skill",
-  "version": "2.0.1972",
+  "version": "2.0.1974",
   "description": "Canonical universal harness — AI-native software engineering via skill-driven orchestration; bootstraps plugkit for task execution and session isolation. Install in any AI coding agent host.",
   "author": "AnEntrypoint",
   "license": "MIT",