Four repositories become one

.github/workflows/collector.yml

@@ -0,0 +1,57 @@
+name: collector
+
+# Path-filtered: a change to a phone runner should not spin up a Node matrix.
+#
+# The trap this creates is worth knowing about before you add a required check
+# in branch protection. A filtered workflow does not run on a PR that touches
+# nothing it watches, and GitHub reports a required check that never ran as
+# pending forever, so an iOS-only PR would block on the collector's suite. Keep
+# these advisory, or add a merge-gate job that always runs.
+on:
+  push:
+    branches: [main, develop]
+    paths: ["collector/**", ".github/workflows/collector.yml"]
+  pull_request:
+    paths: ["collector/**", ".github/workflows/collector.yml"]
+
+defaults:
+  run:
+    working-directory: collector
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      # Both versions report, so a failure on one does not hide the other.
+      fail-fast: false
+      matrix:
+        node: ["22", "24"]
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: ${{ matrix.node }}
+          cache: npm
+          cache-dependency-path: |
+            collector/package-lock.json
+            collector/dash/package-lock.json
+
+      - run: npm ci
+
+      # The dashboard has its own package.json, so `npm test` skips building it
+      # when its modules are absent rather than failing on a missing binary.
+      # This step is what makes that build actually happen here.
+      - name: Install dashboard dependencies
+        run: npm --prefix dash ci
+
+      # Playwright is a dependency but its browsers are not downloaded: nothing
+      # in this suite drives a browser. The web-* workloads that do are run by
+      # an executor on real hardware, not here.
+      #
+      # CI=1 also makes the metric-name check fail on a stale schemas/metrics.json
+      # rather than quietly regenerating it. That file is what an agent working in
+      # a runner directory reads, and a mirror that lags its schema states a
+      # contract confidently and wrongly.
+      - run: npm test

.github/workflows/runner-android.yml

@@ -0,0 +1,39 @@
+name: runner-android
+
+on:
+  push:
+    branches: [main, develop]
+    paths: ["runner-android/**", ".github/workflows/runner-android.yml"]
+  pull_request:
+    paths: ["runner-android/**", ".github/workflows/runner-android.yml"]
+
+defaults:
+  run:
+    working-directory: runner-android
+
+jobs:
+  unit-tests:
+    runs-on: ubuntu-latest
+
+    steps:
+      # Deliberately without submodules. -Pfleet.skipNative drops the native
+      # build, so llama.cpp is never compiled here and cloning it would cost
+      # more than the tests do.
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: "17"
+
+      - uses: gradle/actions/setup-gradle@v4
+        with:
+          build-root-directory: runner-android
+
+      # The JVM tests cover the protocol, the synthetic backend's scaling, the
+      # cancellation registry and the thermal curve's bookkeeping -- none of
+      # which call into the native library. The arm64 llama.cpp build takes
+      # about fifteen minutes and needs a pinned NDK, so it is verified on a
+      # developer machine (and by the phones that actually run the benchmarks)
+      # rather than on every push.
+      - run: ./gradlew :app:testDebugUnitTest -Pfleet.skipNative

.github/workflows/runner-ios.yml

@@ -0,0 +1,37 @@
+name: runner-ios
+
+on:
+  push:
+    branches: [main, develop]
+    paths: ["runner-ios/**", ".github/workflows/runner-ios.yml"]
+  pull_request:
+    paths: ["runner-ios/**", ".github/workflows/runner-ios.yml"]
+
+defaults:
+  run:
+    working-directory: runner-ios
+
+jobs:
+  build:
+    runs-on: macos-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - run: brew install xcodegen
+
+      # Picks the spec this checkout can build. There is no llama.xcframework
+      # here -- it is 850 MB and gitignored -- so this generates the project
+      # without the llama.cpp backend, which is exactly the state a stranger
+      # cloning the repo is in. If that ever stops building, this job is how
+      # we find out.
+      - run: ./generate.sh
+
+      - name: Build for the simulator
+        run: |
+          xcodebuild build \
+            -project FleetRunner.xcodeproj \
+            -scheme FleetRunner \
+            -destination 'generic/platform=iOS Simulator' \
+            -derivedDataPath build \
+            CODE_SIGNING_ALLOWED=NO

.github/workflows/runner-machine.yml

@@ -0,0 +1,37 @@
+name: runner-machine
+
+on:
+  push:
+    branches: [main, develop]
+    paths: ["runner-machine/**", ".github/workflows/runner-machine.yml"]
+  pull_request:
+    paths: ["runner-machine/**", ".github/workflows/runner-machine.yml"]
+
+defaults:
+  run:
+    working-directory: runner-machine
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        node: ["22", "24"]
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: ${{ matrix.node }}
+          cache: npm
+          cache-dependency-path: runner-machine/package-lock.json
+
+      - run: npm ci
+
+      # A GitHub runner has no GPU, no llama-bench and no mlx_lm, which is
+      # exactly the machine the capability probes exist for: the suite passes
+      # there by declaring `benchmark` and nothing else. If a probe ever starts
+      # guessing instead of answering, this is where it shows up.
+      - run: npm test

README.md

@@ -17,13 +17,35 @@ hardware?** It turned out to be yes, and the fleet is how I know.
 
 ![The dashboard: llama.cpp benchmark numbers measured on a real phone](docs/img/results.png)
 
-## The repositories
+## What is in here
+
+Four projects in one repository. They ship independently and share no code —
+only a JSON protocol — but they are versioned together, because the protocol is
+the thing that breaks and a change to it touches three of them at once.
 
 | | What it is |
 |---|---|
-| **[fleet-collector](https://github.com/addisdev/fleet-collector)** | The brain. Device registry, job queue with leases, artifact store, results database, scheduler, alert engine, and the dashboard above. Node + Fastify + SQLite, no broker, no cloud. |
-| **[fleet-runner-android](https://github.com/addisdev/fleet-runner-android)** | The Android agent. A foreground service on anything back to Android 7, with llama.cpp (NDK/JNI) and LiteRT backends. |
-| **[fleet-runner-ios](https://github.com/addisdev/fleet-runner-ios)** | The iOS agent. SwiftUI, with llama.cpp and Core ML backends, speaking the same JSON protocol without sharing a line of code. |
+| **[collector/](collector)** | The brain. Device registry, job queue with leases, artifact store, results database, scheduler, alert engine, and the dashboard above. Node + Fastify + SQLite, no broker, no cloud. |
+| **[runner-android/](runner-android)** | The Android agent. A foreground service on anything back to Android 7, with llama.cpp (NDK/JNI) and LiteRT backends. |
+| **[runner-ios/](runner-ios)** | The iOS agent. SwiftUI, with llama.cpp and Core ML backends, speaking the same JSON protocol without sharing a line of code. |
+| **[runner-machine/](runner-machine)** | The desktop agent. A Node process that makes a laptop or desktop a fleet device, so a phone's tok/s and a laptop's land in the same table. |
+
+Each directory has its own README, its own tests and its own CI job, filtered by
+path so a change to a phone runner does not build the dashboard.
+
+### Why one repository
+
+These were four repositories until the protocol started changing. A metric name
+lives in `collector/schemas/result.schema.json` and is mirrored by hand in three
+runners; a capability list is declared by an agent and enforced by the queue.
+Every one of those is a change that has to land in several places at once, and
+across repositories it lands in several pull requests that can each merge alone.
+The drift is not hypothetical — an eval's accuracy once rode in a field named
+`decode_tok_s` because vision had no field of its own, and no query can
+reproduce that report's numbers today.
+
+One repository makes such a change one reviewable diff, and lets `npm test`
+in `collector/` fail when the schema and its mirror disagree.
 
 ## How it fits together
 
@@ -33,6 +55,7 @@ flowchart LR
         A["Android runner"]
         I["iOS runner"]
     end
+    M["machine runner<br/><i>laptop or desktop</i>"]
     subgraph host["a Mac with devices plugged in"]
         X["host executor<br/><i>adb · Maestro · simctl · Playwright</i>"]
     end
@@ -41,6 +64,7 @@ flowchart LR
 
     A -- "long-poll, claim, report" --> C
     I -- "long-poll, claim, report" --> C
+    M -- "long-poll, claim, report" --> C
     X -- "claims host jobs" --> C
     X -- "drives from outside" --> shelf
     C --- D
@@ -51,10 +75,15 @@ claimed by an executor on a Mac and drive a device from outside, because
 installing an APK or tapping through a UI test is not something an app can do
 to itself.
 
-The two runners share a protocol, not code — including a synthetic SHA-256
-benchmark that is identical on both platforms token for token. That is what
-lets a 2019 Android phone and a current iPhone produce numbers you can put in
-the same table, which is the difference between a fleet and a pile of phones.
+The runners share a protocol, not code — including a synthetic SHA-256
+benchmark that is identical on every platform token for token. That is what
+lets a 2019 Android phone, a current iPhone and a laptop produce numbers you can
+put in the same table, which is the difference between a fleet and a pile of
+phones.
+
+A runner also says what it can run. The queue routes on those declared
+capabilities rather than on a label someone applied, so adding a workload is
+something a runner can do without the collector shipping a release.
 
 ## What came out of it
 
@@ -87,7 +116,7 @@ Three findings worth the whole build:
    trustworthy.
 
 Full write-up, including the quantization scripts and the licensing of every
-model and dataset: **[the eval](https://github.com/addisdev/fleet-collector/blob/main/evals/greenfolio-plant-id.md)**.
+model and dataset: **[the eval](collector/evals/greenfolio-plant-id.md)**.
 
 ## Things I learned the hard way
 

collector/.gitignore

@@ -0,0 +1,11 @@
+node_modules/
+data/
+artifacts/store/
+*.log
+.DS_Store
+power.json
+dash/dist/
+
+# Playwright's per-run output. Traces and screenshots that matter are uploaded
+# to the artifact store by the executor; what lands here is scratch.
+test-results/

collector/.nvmrc

@@ -0,0 +1 @@
+22

collector/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Taylor Addison
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.