Merge pull request #1291 from corbitsdev/cl-9010-spawn-cold-start

perf(subagent): reuse parent setup across spawns

Author: TheGreatAxios

src/agent/environment.ts

@@ -1,6 +1,7 @@
 import { spawn } from "node:child_process";
 import { readdir } from "node:fs/promises";
 import { arch, release, type as osType } from "node:os";
+import { resolve } from "node:path";
 
 export interface EnvironmentInfo {
   cwd: string;
@@ -107,6 +108,57 @@ async function gatherTopLevel(cwd: string): Promise<string | undefined> {
   }
 }
 
+/** CL-9010: how long a cached environment snapshot stays fresh. */
+export const ENVIRONMENT_CACHE_TTL_MS = 5_000;
+
+type EnvironmentCacheEntry = {
+  at: number;
+  value: EnvironmentInfo;
+};
+
+// CL-9010: short-TTL environment cache keyed by resolved cwd. Spawned
+// workers share the dispatcher's git/top-level snapshot: the first spawn
+// pays the git calls and same-cwd spawns within the TTL reuse it, so
+// spawn_agent waves do not re-run `git status` per lane. Concurrent spawns
+// share one in-flight gather instead of stampeding git. Hits return a fresh
+// top-level copy with a current date so one worker cannot mutate another's
+// snapshot and the prompt never shows a stale clock.
+const environmentCache = new Map<string, EnvironmentCacheEntry>();
+const environmentInflight = new Map<string, Promise<EnvironmentInfo>>();
+
+export async function gatherEnvironmentCached(
+  cwd: string,
+  date = new Date(),
+  ttlMs = ENVIRONMENT_CACHE_TTL_MS,
+): Promise<EnvironmentInfo> {
+  const key = resolve(cwd);
+  const entry = environmentCache.get(key);
+  if (entry !== undefined && Date.now() - entry.at < ttlMs) {
+    return { ...entry.value, date };
+  }
+  const inflight = environmentInflight.get(key);
+  if (inflight !== undefined) {
+    return { ...(await inflight), date };
+  }
+  const gather = gatherEnvironment(cwd, date);
+  environmentInflight.set(key, gather);
+  try {
+    const settled = await gather;
+    environmentCache.set(key, { at: Date.now(), value: settled });
+    return { ...settled, date };
+  } finally {
+    // A gather failure caches nothing — the next spawn retries.
+    if (environmentInflight.get(key) === gather)
+      environmentInflight.delete(key);
+  }
+}
+
+/** Test hook: drop every cached environment snapshot. */
+export function resetEnvironmentCacheForTests(): void {
+  environmentCache.clear();
+  environmentInflight.clear();
+}
+
 export async function gatherEnvironment(
   cwd: string,
   date = new Date(),

src/agent/tools.ts

@@ -597,6 +597,10 @@ export async function createAgentToolset(
           ? { secretGuardExtraDeniedPaths }
           : {}),
         ...(skillDirs.length > 0 ? { skillDirs } : {}),
+        // CL-9010: the parent's already-discovered catalog. Shared-cwd lanes
+        // reuse it instead of rescanning; worktree lanes (different cwd)
+        // ignore it and rediscover for their own directory.
+        skillSnapshot: skills,
         ...(extraToolPlugins.length > 0 ? { extraToolPlugins } : {}),
         cwd,
         getWorkdirBase: sa.getWorkdirBase,

src/extensions/skills.ts

@@ -205,3 +205,41 @@ export async function discoverSkills(
   }
   return skills;
 }
+
+// CL-9010: process-lifetime skill catalog cache keyed by resolved cwd +
+// plugin dirs. Spawned workers share the dispatcher's catalog: the first
+// same-cwd spawn pays discovery and later spawns reuse the snapshot, so
+// spawn_agent dispatch stays non-blocking on repeated waves. The fallback
+// base dirs are cwd-relative, so the key must include the cwd — different
+// dirs (or a different cwd) always rediscover. Callers get a copy; the
+// cached canonical is never handed out, so one worker cannot mutate
+// another's catalog. Unbounded in practice the same way the inference
+// singleton is: one entry per distinct (cwd, dirs) pair per process.
+const skillSnapshotCache = new Map<string, readonly SkillSummary[]>();
+
+function skillSnapshotCacheKey(
+  cwd: string,
+  pluginDirs: readonly string[],
+): string {
+  return `${resolve(cwd)}\0${pluginDirs.join("\0")}`;
+}
+
+export async function discoverSkillsCached(
+  cwd: string,
+  pluginDirs: readonly string[] = [],
+): Promise<SkillSummary[]> {
+  const key = skillSnapshotCacheKey(cwd, pluginDirs);
+  const cached = skillSnapshotCache.get(key);
+  if (cached !== undefined) return cached.map((skill) => ({ ...skill }));
+  const fresh = await discoverSkills(cwd, [...pluginDirs]);
+  skillSnapshotCache.set(
+    key,
+    fresh.map((skill) => ({ ...skill })),
+  );
+  return fresh;
+}
+
+/** Test hook: drop every cached skill snapshot so suites start unpolluted. */
+export function resetSkillDiscoveryCacheForTests(): void {
+  skillSnapshotCache.clear();
+}

src/session/assemble-runtime.ts

@@ -138,11 +138,17 @@ import {
  * best-effort in long-lived entry points; a one-shot runner passes onError to
  * log and continue instead. Without onError a seed failure rejects, matching
  * the previous TUI boot behavior.
+ *
+ * CL-9010: fleet-spawned workers pass skipPricingSeed — the parent runtime
+ * already applied the process seed at boot (and the inference singleton is
+ * a no-op on repeat resolves), so the worker skips the cache re-read.
  */
 export async function assembleInferenceBase(
   onPricingError?: (err: unknown) => void,
+  opts?: { skipPricingSeed?: boolean },
 ): Promise<Awaited<ReturnType<typeof createInferenceDependencies>>> {
   const inferenceDeps = await createInferenceDependencies();
+  if (opts?.skipPricingSeed === true) return inferenceDeps;
   const seed = seedPricingMetadataFromCache({
     cachePath: defaultPricingCachePath(),
   });

src/subagent/agent-fleet.ts

@@ -1532,6 +1532,16 @@ export function createSpawnAgentTool(deps: AgentFleetDeps): AgentTool {
             ...(deps.skillDirs !== undefined
               ? { skillDirs: deps.skillDirs }
               : {}),
+            // CL-9010: shared-cwd lanes reuse the dispatcher's skill catalog
+            // instead of rescanning — the snapshot was discovered for this
+            // exact cwd. Worktree lanes run elsewhere, so they rediscover
+            // (through the cached discovery) for their own directory.
+            // Every fleet-spawned worker skips the pricing-cache seed read:
+            // the parent runtime already applied the process seed at boot.
+            ...(deps.skillSnapshot !== undefined && worktreeCwd === undefined
+              ? { skills: deps.skillSnapshot }
+              : {}),
+            skipPricingSeed: true,
             ...(resolved.systemPromptRole !== undefined
               ? { systemPromptRole: resolved.systemPromptRole }
               : {}),

src/subagent/run.ts

@@ -89,7 +89,7 @@ import {
 import { createShellCollectTool } from "../agent/background-shell-tool.js";
 import { createAttachmentRehydrateTransform } from "../session/attachment-store.js";
 import { tryReadPriorHandoffFile } from "../session/compaction-handoff.js";
-import { gatherEnvironment } from "../agent/environment.js";
+import { gatherEnvironmentCached } from "../agent/environment.js";
 import { generateSessionId } from "../session/index.js";
 import { consumeStream } from "../session/stream-consumer.js";
 import { createCycleTextRecorder } from "../session/stream-journal.js";
@@ -118,7 +118,7 @@ import {
   createUseSkillTool,
   workerUseSkillDefinition,
 } from "../agent/use-skill.js";
-import { discoverSkills } from "../extensions/skills.js";
+import { discoverSkillsCached } from "../extensions/skills.js";
 import { formatAttachedSkillConstraints } from "../agent/directors/attached-skills.js";
 import {
   createManageTasksRunner,
@@ -612,7 +612,11 @@ async function runSubAgentInner(
   telemetryRollup: SubAgentTelemetryRollup,
   settlementState: { latestModel: string },
 ): Promise<RunSubAgentResult> {
-  const inferenceDeps = await assembleInferenceBase();
+  // CL-9010: fleet-spawned workers skip the pricing-cache seed re-read —
+  // the parent runtime already applied the process seed at boot.
+  const inferenceDeps = await assembleInferenceBase(undefined, {
+    ...(params.skipPricingSeed === true ? { skipPricingSeed: true } : {}),
+  });
 
   // CL-9475: the fleet session id the parent observes (params.id); falls
   // back to the local session id below when run without a fleet caller.
@@ -774,7 +778,13 @@ async function runSubAgentInner(
       orchestrator: params.orchestrator === true,
     });
     const skillDirs = [...(params.skillDirs ?? [])];
-    const skillSnapshot = await discoverSkills(params.cwd, skillDirs);
+    // CL-9010: reuse the dispatcher's catalog when the lane shares its cwd;
+    // otherwise fall back to the cached discovery (same-cwd repeat spawns
+    // skip the rescan). Copies keep one worker from mutating another's list.
+    const skillSnapshot =
+      params.skills !== undefined
+        ? [...params.skills]
+        : await discoverSkillsCached(params.cwd, skillDirs);
     tools = [
       ...tools,
       createSkillSearchTool({
@@ -975,6 +985,9 @@ async function runSubAgentInner(
           ? { secretGuardExtraDeniedPaths: nd.secretGuardExtraDeniedPaths }
           : {}),
         ...(nd.skillDirs !== undefined ? { skillDirs: nd.skillDirs } : {}),
+        // CL-9010: nested shared-cwd lanes reuse this worker's catalog, which
+        // was discovered (or inherited) for exactly this cwd.
+        skillSnapshot,
         ...(nd.extraToolPlugins !== undefined
           ? { extraToolPlugins: nd.extraToolPlugins }
           : {}),
@@ -1059,7 +1072,9 @@ async function runSubAgentInner(
       getContextDir: () => childContextDir,
     });
 
-    const environment = await gatherEnvironment(params.cwd);
+    // CL-9010: same-cwd spawns within the TTL share the git/top-level
+    // snapshot instead of re-running git per lane.
+    const environment = await gatherEnvironmentCached(params.cwd);
     const attachedSection =
       params.attachedSkills !== undefined && params.attachedSkills.length > 0
         ? await formatAttachedSkillConstraints({