Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

Repair native benchmark compatibility and record current performance

dee23d511e20742
diff --git a/desktop/tool/native_benchmark/NativeBenchmark.swift b/desktop/tool/native_benchmark/NativeBenchmark.swift
index 563b7fdd3..965c2cfcf 100644
--- a/desktop/tool/native_benchmark/NativeBenchmark.swift
+++ b/desktop/tool/native_benchmark/NativeBenchmark.swift
@@ -28,10 +28,19 @@ private enum NativeBenchmark {
"os": ProcessInfo.processInfo.operatingSystemVersionString,
])
case "begin":
- guard window.isKeyWindow, NSApp.isActive,
- let view = window.contentViewController?.view,
- window.makeFirstResponder(view) else {
- result(FlutterError(code: "inactive", message: "Benchmark content must own initial focus", details: nil))
+ guard window.isKeyWindow, NSApp.isActive else {
+ result(FlutterError(code: "inactive", message: "Benchmark window must be active and key", details: [
+ "key": window.isKeyWindow,
+ "active": NSApp.isActive,
+ "visible": window.isVisible,
+ ]))
+ return
+ }
+ // Match the production titlebar's responder. The Flutter controller
+ // accepts keyboard focus; its wrapper NSView may reject it.
+ guard let controller = window.contentViewController as? FlutterViewController,
+ window.makeFirstResponder(controller) else {
+ result(FlutterError(code: "focus", message: "Flutter controller must own initial focus", details: nil))
return
}
// Once, after the synthetic UI is mounted. Never reset native or Dart
diff --git a/desktop/tool/native_benchmark/README.md b/desktop/tool/native_benchmark/README.md
index 475a45ba0..58d33f099 100644
--- a/desktop/tool/native_benchmark/README.md
+++ b/desktop/tool/native_benchmark/README.md
@@ -1,13 +1,14 @@
# Native Release interaction benchmark
-**Handoff status, September 13, 2026:** this previously untracked fixture is saved
-for continuation, not a validated measurement result. Before building/running,
-update the old `Harness V2` product-name replacement in `prepare.py`, the process
-name check in `run.py`, and the initial native responder selection. See the
+**Status, September 14, 2026:** the builder supports the current Harness name,
+validates the copied product identity and verifies the resulting bundle. The
+runner identifies current/legacy previews by bundle ID, and initial native
+focus uses the same Flutter controller as the production titlebar. Six Python
+isolation checks cover current/legacy names, unknown/duplicate configuration,
+running previews and the built bundle identity. See the
[handoff](../../../docs/harness-v2-handoff.md) and
[failed calibration notes](../../../docs/harness-v2-performance.md#native-calibration-remains-unmeasured-2026-09-13).
-Keep isolation and foreground/key-window guards intact; no p50/p95/p99 result
-has been accepted.
+Foreground/key-window guards remain intact; no p50/p95/p99 result has been accepted.
This macOS fixture measures AppKit-queued input through the production Swarm
screen, terminal session parser and Flutter renderer. It runs as **Harness
@@ -22,10 +23,20 @@ From `desktop/`, build with a compatible Flutter SDK and Xcode:
python3 tool/native_benchmark/prepare.py --flutter /path/to/flutter
```
-Use the `BENCHMARK_APP` path printed by the build. Normally close V2 and finish
+Use the `BENCHMARK_APP` path printed by the build. Normally close the workspace preview and finish
other builds/tests first. Keep this fixture in the foreground during a run; it
exits on focus loss instead of reclaiming focus between observations. The runner
-refuses to start alongside another V2 or benchmark process and never quits them.
+refuses to start alongside another preview or benchmark process and never quits
+them. Its error names the exact bundle path: the installed Harness app and the
+workspace preview share the name Harness but have distinct bundle identities.
+The installed app may remain open; record other app activity and host load when
+reporting timings. That is not a guarantee of an otherwise idle workstation.
+
+Run preflight checks without launching an app:
+
+```sh
+PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tool/native_benchmark -p 'test_*.py' -v
+```
```sh
python3 tool/native_benchmark/run.py \
diff --git a/desktop/tool/native_benchmark/isolation.py b/desktop/tool/native_benchmark/isolation.py
new file mode 100644
index 000000000..09b06bfec
--- /dev/null
+++ b/desktop/tool/native_benchmark/isolation.py
@@ -0,0 +1,60 @@
+"""Identity checks shared by the disposable benchmark builder and runner."""
+from pathlib import Path
+import plistlib
+import re
+
+
+BENCHMARK_NAME = 'Harness Benchmark'
+BENCHMARK_ID = 'ai.autonomous.harness.benchmark'
+PREVIEW_ID = 'ai.autonomous.harness.v2'
+RELEASE_ID = 'ai.autonomous.harness'
+KNOWN_NAMES = ('Harness', 'Harness V2', BENCHMARK_NAME)
+
+
+def benchmark_configuration(source):
+ """Replace exactly one known product/identifier in the copied config."""
+ replacements = (
+ ('PRODUCT_NAME', ('Harness', 'Harness V2'), BENCHMARK_NAME),
+ ('PRODUCT_BUNDLE_IDENTIFIER', (PREVIEW_ID,), BENCHMARK_ID),
+ )
+ for key, expected, replacement in replacements:
+ pattern = re.compile(rf'^{key}[ \t]*=[ \t]*(.*)$', re.MULTILINE)
+ matches = list(pattern.finditer(source))
+ if len(matches) != 1 or matches[0].group(1).strip() not in expected:
+ raise ValueError(f'Expected one known {key} assignment in the preview config')
+ source = pattern.sub(f'{key} = {replacement}', source)
+ return source
+
+
+def bundle_info(app):
+ with (app / 'Contents/Info.plist').open('rb') as source:
+ return plistlib.load(source)
+
+
+def validate_benchmark_bundle(app):
+ info = bundle_info(app)
+ if (app.name != f'{BENCHMARK_NAME}.app'
+ or info.get('CFBundleIdentifier') != BENCHMARK_ID
+ or info.get('CFBundleExecutable') != BENCHMARK_NAME):
+ raise ValueError('Expected the isolated Harness Benchmark product and identity')
+
+
+def conflicting_previews(executable_paths):
+ """Read bundle identities, since the installed app and preview share a name."""
+ conflicts = []
+ for line in executable_paths.splitlines():
+ executable = Path(line.strip())
+ if executable.parent.name != 'MacOS' or executable.parent.parent.name != 'Contents':
+ continue
+ app = executable.parent.parent.parent
+ if app.name not in {f'{name}.app' for name in KNOWN_NAMES}:
+ continue
+ try:
+ identifier = bundle_info(app).get('CFBundleIdentifier')
+ except (OSError, ValueError, plistlib.InvalidFileException) as error:
+ raise ValueError(f'Cannot identify running {app.name}: {error}') from error
+ if identifier in (PREVIEW_ID, BENCHMARK_ID):
+ conflicts.append(str(app))
+ elif identifier != RELEASE_ID:
+ raise ValueError(f'Unrecognized running Harness identity: {identifier}')
+ return conflicts
diff --git a/desktop/tool/native_benchmark/prepare.py b/desktop/tool/native_benchmark/prepare.py
index 804f1f6cc..2e4377b5d 100644
--- a/desktop/tool/native_benchmark/prepare.py
+++ b/desktop/tool/native_benchmark/prepare.py
@@ -7,6 +7,8 @@ import shutil
import subprocess
import tempfile
+from isolation import benchmark_configuration, validate_benchmark_bundle
+
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--flutter', required=True, type=Path)
args = parser.parse_args()
@@ -23,7 +25,7 @@ code = code.replace(marker, ' NativeBenchmark.install(window: self, messenger
code += '\n' + (source / 'tool/native_benchmark/NativeBenchmark.swift').read_text()
window.write_text(code)
info = desktop / 'macos/Runner/Configs/AppInfo.xcconfig'
-info.write_text(info.read_text().replace('PRODUCT_NAME = Harness V2', 'PRODUCT_NAME = Harness Benchmark').replace('ai.autonomous.harness.v2', 'ai.autonomous.harness.benchmark'))
+info.write_text(benchmark_configuration(info.read_text()))
env = dict(os.environ, XDG_CONFIG_HOME=str(root / 'tool-config'))
flutter = str(args.flutter / 'bin/flutter')
log = root / 'build.log'
@@ -38,4 +40,5 @@ with log.open('w') as output:
app = desktop / 'build/macos/Build/Products/Release/Harness Benchmark.app'
if not app.exists():
raise RuntimeError('Expected isolated bundle was not built')
+validate_benchmark_bundle(app)
print(f'BENCHMARK_APP={app}', flush=True)
diff --git a/desktop/tool/native_benchmark/run.py b/desktop/tool/native_benchmark/run.py
index 7f86c74a1..53a0375f7 100644
--- a/desktop/tool/native_benchmark/run.py
+++ b/desktop/tool/native_benchmark/run.py
@@ -3,10 +3,11 @@
import argparse
import json
from pathlib import Path
-import plistlib
import subprocess
import sys
+from isolation import conflicting_previews, validate_benchmark_bundle
+
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--app', required=True, type=Path)
@@ -21,17 +22,20 @@ if not app.is_relative_to(temporary) or not output.is_relative_to(temporary):
parser.error('The benchmark app and output must be under /private/tmp')
if not 1 <= args.samples <= 500:
parser.error('Use between 1 and 500 samples per operation/load')
-with (app / 'Contents/Info.plist').open('rb') as source:
- info = plistlib.load(source)
-if info.get('CFBundleIdentifier') != 'ai.autonomous.harness.benchmark':
- parser.error('Refusing to launch a bundle other than the isolated fixture')
+try:
+ validate_benchmark_bundle(app)
+except (OSError, ValueError) as error:
+ parser.error(str(error))
log = output.with_suffix('.log')
if output.exists() or log.exists():
parser.error('Choose fresh result and log paths')
-processes = subprocess.check_output(['ps', '-axo', 'command='], text=True)
-for name in ('Harness Benchmark', 'Harness V2'):
- if f'/{name}.app/Contents/MacOS/{name}' in processes:
- parser.error(f'Normally close {name} before measuring; this runner never quits other apps')
+processes = subprocess.check_output(['ps', '-axo', 'comm='], text=True)
+try:
+ conflicts = conflicting_previews(processes)
+except ValueError as error:
+ parser.error(str(error))
+if conflicts:
+ parser.error(f'Normally close {", ".join(conflicts)} before measuring; this runner never quits other apps')
output.parent.mkdir(parents=True, exist_ok=True)
command = [
'open', '-n', '-W', '-o', str(log), '--stderr', str(log),
diff --git a/desktop/tool/native_benchmark/test_isolation.py b/desktop/tool/native_benchmark/test_isolation.py
new file mode 100644
index 000000000..16d352c98
--- /dev/null
+++ b/desktop/tool/native_benchmark/test_isolation.py
@@ -0,0 +1,77 @@
+from pathlib import Path
+import plistlib
+import tempfile
+import unittest
+
+from isolation import (
+ BENCHMARK_ID, BENCHMARK_NAME, PREVIEW_ID, RELEASE_ID,
+ benchmark_configuration, conflicting_previews, validate_benchmark_bundle,
+)
+
+
+class IsolationTests(unittest.TestCase):
+ def setUp(self):
+ self.directory = tempfile.TemporaryDirectory(prefix='harness-benchmark-check-')
+ self.addCleanup(self.directory.cleanup)
+ self.root = Path(self.directory.name)
+
+ def app(self, folder, name, identifier, executable=None):
+ bundle = self.root / folder / f'{name}.app'
+ (bundle / 'Contents/MacOS').mkdir(parents=True)
+ binary = executable or name
+ with (bundle / 'Contents/Info.plist').open('wb') as output:
+ plistlib.dump({'CFBundleIdentifier': identifier, 'CFBundleExecutable': binary}, output)
+ return bundle, bundle / 'Contents/MacOS' / binary
+
+ def test_current_and_legacy_product_names_are_replaced(self):
+ for name in ('Harness', 'Harness V2'):
+ with self.subTest(name=name):
+ source = f'// product\nPRODUCT_NAME = {name}\nPRODUCT_BUNDLE_IDENTIFIER = {PREVIEW_ID}\n'
+ result = benchmark_configuration(source)
+ self.assertIn(f'PRODUCT_NAME = {BENCHMARK_NAME}\n', result)
+ self.assertIn(f'PRODUCT_BUNDLE_IDENTIFIER = {BENCHMARK_ID}\n', result)
+ self.assertTrue(result.startswith('// product\n'))
+
+ def test_unknown_or_duplicate_configuration_fails_closed(self):
+ known = f'PRODUCT_NAME = Harness\nPRODUCT_BUNDLE_IDENTIFIER = {PREVIEW_ID}\n'
+ for source in (known.replace('Harness', 'Other'),
+ known.replace(PREVIEW_ID, RELEASE_ID),
+ known + 'PRODUCT_NAME = Harness\n',
+ known.replace('PRODUCT_NAME = Harness\n', '')):
+ with self.subTest(source=source), self.assertRaises(ValueError):
+ benchmark_configuration(source)
+
+ def test_current_and_legacy_preview_processes_are_blocked(self):
+ for name in ('Harness', 'Harness V2', BENCHMARK_NAME):
+ with self.subTest(name=name):
+ identifier = BENCHMARK_ID if name == BENCHMARK_NAME else PREVIEW_ID
+ bundle, binary = self.app(name, name, identifier)
+ self.assertEqual(conflicting_previews(str(binary)), [str(bundle)])
+
+ def test_installed_app_and_non_app_commands_are_not_preview_processes(self):
+ _, binary = self.app('installed', 'Harness', RELEASE_ID)
+ self.assertEqual(conflicting_previews(f'{binary}\n/usr/bin/python3\n/bin/zsh'), [])
+
+ def test_unknown_or_unreadable_running_harness_fails_closed(self):
+ bundle, binary = self.app('unknown', 'Harness', 'unexpected')
+ with self.assertRaises(ValueError):
+ conflicting_previews(str(binary))
+ (bundle / 'Contents/Info.plist').unlink()
+ with self.assertRaises(ValueError):
+ conflicting_previews(str(binary))
+
+ def test_built_bundle_requires_both_isolated_identity_and_executable(self):
+ bundle, _ = self.app('valid', BENCHMARK_NAME, BENCHMARK_ID)
+ validate_benchmark_bundle(bundle)
+ for folder, name, identifier, executable in (
+ ('wrong-name', 'Harness', BENCHMARK_ID, BENCHMARK_NAME),
+ ('wrong-id', BENCHMARK_NAME, PREVIEW_ID, BENCHMARK_NAME),
+ ('wrong-binary', BENCHMARK_NAME, BENCHMARK_ID, 'Harness'),
+ ):
+ wrong, _ = self.app(folder, name, identifier, executable)
+ with self.subTest(folder=folder), self.assertRaises(ValueError):
+ validate_benchmark_bundle(wrong)
+
+
+if __name__ == '__main__':
+ unittest.main()
diff --git a/docs/harness-v2-handoff.md b/docs/harness-v2-handoff.md
index becd2b6a4..b458c4f9f 100644
--- a/docs/harness-v2-handoff.md
+++ b/docs/harness-v2-handoff.md
@@ -28,8 +28,9 @@ Research must lead to justified improvements, not feature accumulation.
- Repository: <https://github.com/autonomous-ai/autonomous-harness>.
- Continue on **`swarm-onboarding`**, created from updated `origin/main`
- (`0b9ac72`) after PR #31 merged. Publish/checkpoint this branch; do not resume
- feature work on `app-v2`.
+ (`0b9ac72`) after PR #31 merged, tracking `origin/swarm-onboarding`.
+- Draft [PR #34 — Separate navigation from adding agents and simplify first use](https://github.com/autonomous-ai/autonomous-harness/pull/34)
+ is open against main. Keep it draft while qualification remains incomplete.
- [PR #31 — Harness: swarm workspaces for AI agents](https://github.com/autonomous-ai/autonomous-harness/pull/31)
merged at `cb24361` on September 14, 2026 (02:56 UTC), with `744c1fe` as its
app-v2 head. Its merge did **not** include the later Navigate/Add commits.
@@ -49,8 +50,9 @@ Research must lead to justified improvements, not feature accumulation.
without rewriting them, preserving newer main changes including device focus.
- The user explicitly authorized frequent public repository checkpoints and
requested deleting merged branches, then starting fresh from updated main.
- Preserve the unmerged continuation on the new remote branch before retiring
- app-v2. Keep the follow-up PR draft while qualification remains incomplete.
+ The new remote branch preserves every app-v2 commit. After verifying that
+ ancestry and that no app-v2 PR remained open, app-v2 was deleted remotely with
+ an exact-tip lease and then deleted locally. Do not recreate it.
Start with a clean checkout of the remote branch and inspect its latest commit.
On an existing checkout, preserve local work before updating; use a fast-forward
@@ -205,24 +207,31 @@ the common page, without adding another first-tab-only component.
**Latest verification:** after integration with updated main, 1,271 desktop
tests passed with one existing skip. Analyzer reports zero errors/warnings and
14 informational diagnostics (12 vendored, two inherited from main). The last
-native checks passed 51 keymap decoder and 344 AppKit assertions; this onboarding
-pass changes no native source. Synthetic start-page/form captures with real
+native checks passed 51 keymap decoder and 344 AppKit assertions; the production
+native sources still match that checkpoint. Synthetic start-page/form captures with real
fonts were reviewed at 1280×800 and 880×560 with 2× text; they are not a real
native-session visual or latency measurement. Logs:
`/private/tmp/harness-onboarding-main-{tests,analyze}.log`.
Navigate/Add timings and their exact scope are in
[the performance record](harness-v2-performance.md). The new navigation catalog
-build measured 0.194 ms median / 0.264 ms p95 for 50 locations across 12 swarms.
-Logs: `/private/tmp/harness-two-pickers-{full-tests-final,analyze-clean,native}.log`.
+build measured 0.188 ms median / 0.259 ms p95 in the latest integrated run for
+50 locations across 12 swarms. Add's 2,000-agent query measured 0.959 ms median /
+1.005 ms p95. The five benchmark cases passed; native event-to-display remains
+unmeasured. Log: `/private/tmp/harness-onboarding-main-benchmark.log`.
**Also saved:** [native benchmark tooling](../desktop/tool/native_benchmark/README.md)
and [shared-swarm collaboration research](harness-collaboration-design.md), which
were existing uncommitted work. Collaboration is a separate proposal/prototype,
not implemented multiplayer or a prerequisite for finishing this desktop pass.
-The benchmark still needs compatibility fixes before another run: its prepare
-script assumes the old `Harness V2` product name, its process check uses that
-name, and its native initial-responder logic is the older version. Preserve its
-isolation/focus guards. The performance notes explain the failed calibration.
+The benchmark's product-name, process-identity and initial-responder compatibility
+issues are now repaired. Six isolated Python checks pass, and a read-only check
+recognizes the running renamed preview by its bundle ID. Preserve its isolation
+and foreground/key-window guards. The performance notes explain the earlier
+failed calibration; these tooling fixes provide no new latency measurements.
+The disposable Release build succeeded at
+`/private/tmp/harness-native-benchmark-x227eh35`. Its actual runner correctly
+refused to start while the workspace preview was running, writing no timing
+result. The [performance record](harness-v2-performance.md) has the artifacts.
## Previous terminal fix and verification
@@ -259,9 +268,10 @@ Historical checks for that terminal checkpoint:
51 keymap decoder + 343 AppKit checks before it was parked. Those results
apply to the old draft only. Its final visual treatment still needs review.
-The common onboarding continuation's macOS arm64 **Release build succeeded**
+The `swarm-onboarding` continuation's macOS arm64 **Release build succeeded**
+after integration with updated main,
with `FLUTTER_TARGET=lib/main.dart`. Log:
-`/private/tmp/harness-onboarding-release-build.log`. Computer Use still returns
+`/private/tmp/harness-onboarding-main-release-build.log`. Computer Use still returns
`cgWindowNotFound` for the exact workspace bundle path. The new binary is built;
it has not been confirmed loaded in the running preview or visually verified
through native capture. Preserve this distinction when resuming.
@@ -282,8 +292,8 @@ that its isolated window can become active/key.
## Plan for the next session
-1. **Confirm the latest source/build checkpoint.** The common onboarding
- continuation has a successful Release build; its running-window revision is
+1. **Confirm the latest source/build checkpoint.** The fresh-branch continuation
+ has a successful Release build; its running-window revision is
not confirmed. The last confirmed relaunch was `2d3c024`. Review
Navigate, Add, command mode, and the fresh terminal position in the native app
when computer-use access works. Preserve real running agents and input.
diff --git a/docs/harness-v2-performance.md b/docs/harness-v2-performance.md
index e9d0e131e..b82301eaf 100644
--- a/docs/harness-v2-performance.md
+++ b/docs/harness-v2-performance.md
@@ -10,6 +10,36 @@ flutter test --no-pub --reporter expanded test/benchmarks/swarm_benchmark.dart
It prints `SWARM_BENCH` JSON records. Its filename deliberately does not end in `_test.dart`, so ordinary correctness runs do not include timing measurements.
+## Current continuation benchmark (2026-09-14)
+
+After integrating updated main, the five explicit benchmark cases passed on
+production source at `0b3b343`. They ran sequentially (`--concurrency=1`) after
+the builds and correctness checks finished, on the same shared M2 Max workstation.
+This is a fresh headless debug observation, not a paired before/after experiment
+or a native input-to-display measurement.
+
+| Operation | Median | p95 | p99 |
+| --- | ---: | ---: | ---: |
+| Navigate catalog: 50 locations / 12 swarms | 0.188 ms | 0.259 ms | 0.491 ms |
+| Navigate query | 0.020 ms | 0.028 ms | 0.134 ms |
+| Add query `agent`: 2,000 agents | 0.959 ms | 1.005 ms | 1.076 ms |
+| Add contextual query | 0.779 ms | 0.864 ms | 0.868 ms |
+| Add fuzzy query | 0.725 ms | 0.809 ms | 0.824 ms |
+| Decode/parse 16 KiB ASCII | 0.523 ms | 0.689 ms | 1.010 ms |
+| Decode/parse 16 KiB Unicode | 0.332 ms | 0.461 ms | 1.602 ms |
+| Switch/pump: 16 retained terminals | 13.056 ms | 20.293 ms | 30.776 ms |
+| Switch/pump: 48 retained terminals | 10.426 ms | 11.991 ms | 12.328 ms |
+| Focus/pump: 16 retained terminals | 5.802 ms | 6.949 ms | 7.493 ms |
+| Focus/pump: 48 retained terminals | 5.578 ms | 7.297 ms | 8.863 ms |
+
+Catalog/output operations have 100 measured samples; frame workloads have 60,
+with 1,000 retained lines per terminal at 1280×800. A p99 from 60 observations is
+effectively a maximum, not a well-established tail estimate. Host load and JIT
+remain sources of variation; the 48-terminal case being faster does not mean
+more terminals improve performance. The 16-terminal switch still has a much
+slower tail than its median. These numbers do not qualify native responsiveness.
+Raw log: `/private/tmp/harness-onboarding-main-benchmark.log`.
+
## Distinct Navigate and Add agent continuation
Navigate now indexes only existing swarm locations. The initial implementation
@@ -118,6 +148,32 @@ Reads traverse at most 32 ancestors and request at most 64 KiB plus one byte per
All three stale-cache regressions failed on the earlier code. The final focused run passed 38 tests, including directory and common-path changes, ended/failed/unsupported watches, missed events, deletion, same-path replacement, serialized slow reads and disposal. The preexisting real temporary-filesystem worktree watch also passes; synthetic watchers cover failure ordering without relying on OS timing. Logs: `/private/tmp/harness-v2-git-watch-{before,tests,analyze}.log`. This establishes recovery and bounded work; it is not a new native latency measurement.
+## Native benchmark tooling repair (2026-09-14)
+
+The saved benchmark assumed the old Harness V2 name and requested initial focus
+on a wrapper view. Its builder now replaces exactly one recognized product-name
+and bundle-ID assignment in the disposable copy, and verifies the built app's
+name, executable and isolated identity. The runner reads executable paths and
+bundle IDs so the renamed Harness preview is distinguished from the installed
+Harness app. Unknown/unreadable Harness identities stop preflight. Initial focus
+uses the Flutter controller, matching the production titlebar, while foreground
+and key-window requirements remain intact and report their actual state.
+
+Six isolated Python checks pass. A new macOS arm64 Release fixture built at
+`/private/tmp/harness-native-benchmark-x227eh35`, with the expected
+`ai.autonomous.harness.benchmark` bundle ID and `Harness Benchmark` executable.
+The actual runner then correctly refused to start alongside the known running
+workspace preview and wrote no timing result. That is preflight validation,
+not a new input-latency sample. The workspace preview remains running; Computer
+Use still cannot capture it by its exact path.
+
+Artifacts: `/private/tmp/harness-native-isolation-tests.log`,
+`/private/tmp/harness-native-prepare-current.log`, the disposable `build.log`,
+and `/private/tmp/harness-native-current-preflight-check.log`. Use the updated
+[benchmark instructions](../desktop/tool/native_benchmark/README.md) once the
+preview can be normally closed and the isolated window can acquire valid
+foreground/key status. No p50/p95/p99 native result is accepted yet.
+
## Native calibration remains unmeasured (2026-09-13)
A fresh disposable Release copy of the existing native benchmark built at `/private/tmp/harness-native-benchmark-4t74nfhn`. Its separate bundle identity, temporary store, blocked networking and synthetic retained terminals isolate it from real agents. AppKit events are posted only to that fixture's own queue. The planned metric joins each observed input to Flutter's matching frame number and raster-finish timestamp; it excludes network and physical display latency. [Flutter's FrameTiming reference](https://api.flutter.dev/flutter/dart-ui/FrameTiming-class.html) describes those fields and recommends profile/release performance collection. The distinction between input latency under load and throughput follows the measurement discussion in [Dan Luu's terminal study](https://danluu.com/term-latency/); its old terminal rankings are not treated as current comparisons.
diff --git a/docs/harness-v2-progress.md b/docs/harness-v2-progress.md
index 3ed543bbb..4f5a9269b 100644
--- a/docs/harness-v2-progress.md
+++ b/docs/harness-v2-progress.md
@@ -6,15 +6,31 @@ Updated 2026-09-14 after the common empty-swarm onboarding continuation and bran
## Common empty-swarm onboarding continuation
+- Repaired the saved native benchmark for current Harness branding, bundle-ID
+ preflight and the production Flutter-controller focus target. Six Python
+ isolation checks pass and the disposable arm64 Release fixture builds.
+ The real runner refuses the still-running workspace preview and writes no
+ timing result. Native latency remains unmeasured; details are in the
+ [performance record](harness-v2-performance.md).
- GitHub verification found PR #31 merged at `cb24361`, with `744c1fe` as its
app-v2 head. The later Navigate/Add commits were not in that merge. Saved the
onboarding work as `e122b63`, created **swarm-onboarding** from updated main
(`0b9ac72`), and carried the full unmerged continuation forward in `a2888ef`.
Newer main/device changes are retained. Continue on the fresh branch.
+- Pushed the continuation and opened draft [PR #34](https://github.com/autonomous-ai/autonomous-harness/pull/34).
+ Confirmed the pushed branch contains all old app-v2 commits and no old-branch
+ PR remains open, then deleted app-v2 remotely with an exact-tip lease and
+ locally. The branch lifecycle follows the user's requested practice.
- After integration with updated main, **1,271 desktop tests passed with one
existing skip**. Analyzer reports zero errors/warnings and 14 informational
diagnostics (12 vendored, two inherited from main). Logs:
`/private/tmp/harness-onboarding-main-{tests,analyze}.log`.
+- The current five-case CPU benchmark passed after all builds/tests finished.
+ Navigate catalog: 0.188 ms median / 0.259 ms p95; Add's 2,000-agent query:
+ 0.959 / 1.005 ms. Tab/frame pump: 13.056 / 20.293 ms for 16 terminals and
+ 10.426 / 11.991 ms for 48. The 16-terminal p99 was 30.776 ms; retain that
+ slower tail. These headless observations do not measure native display
+ latency. Full scope/raw log are in the performance record.
- One start page serves every empty swarm. With no existing work it leads with
**Start with one agent**, a static side-by-side example, **Choose folder…** and
separate **Clone repository…**. With existing work it offers shared Add search,
@@ -42,6 +58,8 @@ Updated 2026-09-14 after the common empty-swarm onboarding continuation and bran
target (`/private/tmp/harness-onboarding-release-build.log`). Exact-path native
capture still returns `cgWindowNotFound`; the new binary has not been confirmed
loaded in the running preview. No native visual or latency result is claimed.
+- The integrated fresh branch also has a successful Release build:
+ `/private/tmp/harness-onboarding-main-release-build.log`.
## Navigate and Add agent continuation (previous checkpoint)