diff --git a/packages/core/test/boot-work.test.ts b/packages/core/test/boot-work.test.ts
new file mode 100644
index 000000000000..1f2636c8fa08
--- /dev/null
+++ b/packages/core/test/boot-work.test.ts
@@ -0,0 +1,214 @@
+// Counts the work a cold Location boot and one session turn do, so a speedup can't silently regress:
+// catalog lookups per step (#58601), Azure model edits per model rebuild (#58600) and git spawns (#58602).
+// These are exact counts, not timings. Changing one is fine when the change is intended; update the expected
+// number here and say why in the PR.
+import { describe, expect } from "bun:test"
+import path from "path"
+import whichPkg from "which"
+import { Effect, Layer } from "effect"
+import { ChildProcessSpawner } from "effect/process"
+import { LanguageModel } from "@opencode/ai/effect"
+import { OpenAIChat } from "@opencode/ai/protocols/openai-chat"
+import { TestLLM } from "@opencode/ai/testing"
+import { Agent } from "@opencode/core/agent"
+import { AppNodeBuilder } from "@opencode/core/effect/app-node-builder"
+import { LayerNodePlatform } from "@opencode/core/effect/app-node-platform"
+import { Watcher } from "@opencode/core/filesystem/watcher"
+import { LocationServiceMap } from "@opencode/core/location-service-map"
+import { Model } from "@opencode/core/model"
+import { Plugin } from "@opencode/core/plugin"
+import { Provider } from "@opencode/core/provider"
+import { AbsolutePath } from "@opencode/core/schema"
+import { Session } from "@opencode/core/session"
+import { SessionRunnerModel } from "@opencode/core/session/runner/model"
+import { gitExecutable } from "@opencode/core/util/git-executable"
+import { CrossSpawnSpawner } from "@opencode/util/cross-spawn-spawner"
+import { LayerNode } from "@opencode/util/effect/layer-node"
+import { Global } from "@opencode/util/global"
+import { tempGlobalLayer } from "./fixture/global"
+import { initRepo } from "./fixture/git"
+import { offlineModels } from "./fixture/models"
+import { tmpdirScoped } from "./fixture/tmpdir"
+import { testEffect } from "./lib/effect"
+
+const counters = { catalogLookups: 0, azureModelEdits: 0, gitSpawns: [] as string[], gitLookups: 0 }
+
+const isGit = (command: string) => path.basename(command).replace(/\.exe$/i, "") === "git"
+
+const countingSpawner = CrossSpawnSpawner.node.replace(
+  CrossSpawnSpawner.node.mapLayer((base) =>
+    Layer.effect(
+      ChildProcessSpawner.ChildProcessSpawner,
+      Effect.map(ChildProcessSpawner.ChildProcessSpawner, (spawner) =>
+        ChildProcessSpawner.make((command) => {
+          if (command._tag === "StandardCommand" && isGit(command.command))
+            counters.gitSpawns.push([command.command, ...command.args].join(" "))
+          return spawner.spawn(command)
+        }),
+      ),
+    ).pipe(Layer.provide(base)),
+  ),
+)
+
+const isAzure = (provider: Provider.Info) =>
+  provider.id === Provider.ID.azure || provider.package.startsWith("@opencode/ai/providers/azure/")
+
+// `ctx.model.list()` reads `Model.available()`. Every editor list/get/update clones the models it touches.
+const countingModels = Model.node.replace(
+  Model.node.mapLayer((base) =>
+    Layer.effect(
+      Model.Service,
+      Effect.map(Model.Service, (models) =>
+        Model.Service.of({
+          ...models,
+          available: () =>
+            Effect.suspend(() => {
+              counters.catalogLookups++
+              return models.available()
+            }),
+          transform: (update) =>
+            models.transform((editor) => {
+              const azure = new Set(
+                editor.provider
+                  .list()
+                  .filter((item) => isAzure(item.provider))
+                  .map((item) => item.provider.id),
+              )
+              update({
+                ...editor,
+                list: (providerID) => {
+                  const result = editor.list(providerID)
+                  counters.azureModelEdits += result.filter((model) => azure.has(model.providerID)).length
+                  return result
+                },
+                get: (providerID, modelID) => {
+                  if (azure.has(providerID)) counters.azureModelEdits++
+                  return editor.get(providerID, modelID)
+                },
+                update: (providerID, modelID, edit) => {
+                  if (azure.has(providerID)) counters.azureModelEdits++
+                  editor.update(providerID, modelID, edit)
+                },
+              })
+            }),
+        }),
+      ),
+    ).pipe(Layer.provide(base)),
+  ),
+)
+
+// Counts PATH lookups of git: resolving the git executable (macOS and Windows) and finding Git Bash (Windows).
+const countingGitLookups = Effect.acquireRelease(
+  Effect.sync(() => {
+    const sync = whichPkg.sync
+    whichPkg.sync = Object.assign((cmd: string, options?: Parameters<typeof sync>[1]) => {
+      if (cmd === "git") counters.gitLookups++
+      return sync(cmd, options as never)
+    }, sync) as typeof sync
+    return sync
+  }),
+  (sync) => Effect.sync(() => void (whichPkg.sync = sync)),
+)
+
+const llmLayer = TestLLM.testLayer({ fallback: TestLLM.text("Done", "done") })
+const it = testEffect(
+  Layer.merge(
+    llmLayer,
+    AppNodeBuilder.build(LayerNode.group([Session.node, LocationServiceMap.node]), [
+      Global.node.replace(tempGlobalLayer),
+      offlineModels,
+      Watcher.node.replace(Watcher.configured({ enabled: false })),
+      LayerNodePlatform.llmClient.replace(llmLayer),
+      SessionRunnerModel.node.replace(
+        Layer.succeed(SessionRunnerModel.Service, {
+          resolve: () =>
+            Effect.succeed(
+              SessionRunnerModel.resolved(
+                LanguageModel.make({ id: "boot", provider: "test", route: OpenAIChat.route }),
+                {
+                  capabilities: { tools: true, input: ["text"], output: ["text"] },
+                  cost: [],
+                  limit: { context: 200_000, output: 32_000 },
+                },
+              ),
+            ),
+        }),
+      ),
+      countingSpawner,
+      countingModels,
+    ]),
+  ),
+)
+
+const expected = {
+  // Folder startup: project identity, VCS branch/remote reads and the Location's repository probe.
+  "boot: git spawns": 8,
+  // The step's snapshot: repository probe, snapshot init, diff, untracked files, ignore check, index and tree.
+  "turn: git spawns": 9,
+  // Git is resolved once per process, before this test, and every spawn uses that resolution. On Windows the shell
+  // tool's ShellSelect.gitbash() looks up git once to find Git Bash (cached per process), so one lookup is allowed.
+  "boot + turn: git lookups beyond one": 0,
+  "boot + turn: git spawns not using the resolved executable": 0,
+  // One step resolves the prompt hook model once, shared by the context, compaction and generate hooks.
+  "turn: catalog lookups per step": 1,
+  // Azure is enabled without a resource name, so its model transform has nothing to rewrite.
+  "model rebuild: Azure model edits": 0,
+}
+
+describe("boot work", () => {
+  it.live("counts catalog lookups, Azure model edits and git spawns over a cold boot and one turn", () =>
+    Effect.gen(function* () {
+      const tmp = yield* tmpdirScoped()
+      yield* Effect.promise(() => initRepo(tmp.path))
+      yield* Effect.promise(() =>
+        Bun.write(path.join(tmp.path, "opencode.json"), JSON.stringify({ providers: { azure: {} } })),
+      )
+      const executable = gitExecutable()
+      counters.gitSpawns = []
+      counters.gitLookups = 0
+      yield* countingGitLookups
+
+      const sessions = yield* Session.Service
+      const locations = yield* LocationServiceMap.Service
+      const session = yield* sessions.create({
+        location: { directory: AbsolutePath.make(tmp.path) },
+        agent: Agent.ID.make("build"),
+        model: Model.Ref.make({ id: Model.ID.make("boot"), providerID: Provider.ID.make("test") }),
+      })
+      const location = locations.get(session.location)
+      yield* Plugin.Service.use((plugins) => plugins.awaitActivation).pipe(Effect.provide(location))
+      const boot = counters.gitSpawns.length
+
+      counters.catalogLookups = 0
+      yield* sessions.prompt({ sessionID: session.id, text: "hello" })
+      yield* sessions.wait(session.id)
+      const steps = (yield* sessions.context(session.id)).filter((message) => message.type === "assistant").length
+      const lookups = counters.catalogLookups
+      const spawns = [...counters.gitSpawns]
+
+      counters.azureModelEdits = 0
+      yield* Model.Service.use((models) => models.reload().pipe(Effect.andThen(models.available()))).pipe(
+        Effect.provide(location),
+      )
+
+      const measured: Record<keyof typeof expected, number> = {
+        "boot: git spawns": boot,
+        "turn: git spawns": spawns.length - boot,
+        "boot + turn: git lookups beyond one": Math.max(0, counters.gitLookups - 1),
+        "boot + turn: git spawns not using the resolved executable": spawns.filter(
+          (spawn) => !spawn.startsWith(`${executable} `),
+        ).length,
+        "turn: catalog lookups per step": steps === 0 ? lookups : lookups / steps,
+        "model rebuild: Azure model edits": counters.azureModelEdits,
+      }
+      const drift = Object.entries(expected).flatMap(([name, want]) => {
+        const got = measured[name as keyof typeof expected]
+        if (got === want) return []
+        return [`${name}: expected ${want}, got ${got} (${got > want ? "+" : ""}${got - want})`]
+      })
+      const report = drift.length === 0 ? "" : [...drift, "", "git spawns, boot then turn:", ...spawns].join("\n")
+      expect(steps).toBe(1)
+      expect(report).toBe("")
+    }),
+  )
+})