Sybil-2/ios/Tests/AgentPerformance.md

# Agent interaction benchmarks

Run the production Agent view in an app-hosted XCTest on the iPhone 16e simulator:

```sh
just generate
xcodebuild test -scheme SybilPresentation \
  -destination 'platform=iOS Simulator,name=iPhone 16e,OS=latest' \
  -derivedDataPath build/DerivedData -parallel-testing-enabled NO \
  -only-testing:SybilPresentationTests/AgentPerformanceTests/testInputAndEntranceResponsiveness \
  -resultBundlePath build/agent-input-performance.xcresult
```

`testInputAndEntranceResponsiveness` seeds 160 local mixed literal/Markdown
messages through the normal room reducer (100 initially displayed), then:

- Receives 12 more messages at 550 ms intervals through the production 100 ms
  publication batch and 240 ms entrance animation, checking bottom following.
- Presents and dismisses the actual software keyboard six times, 600 ms per
  transition. It asserts that the transcript viewport shrinks by over 100 points
  and remains at the bottom; a hardware-keyboard-only run is invalid.
- Calls UIKit `insertText` on the actual focused composer every 30 ms, including
  enough text to wrap and grow to its six-line limit. It verifies the draft,
  bottom following, and that keyboard/draft changes don't reproject the history.

The real composer, Markdown rows, avatars and entrance animations remain enabled.
No server, production credentials, notification delivery or network images are
needed. The log prints `AGENT_INPUT_PERFORMANCE_JSON`; the result bundle retains
JSON and screenshots of the actual hosted view before and after typing.
Screenshots capture the hosting view, not the separate system keyboard window.

Use the same test source, simulator, build configuration and host load for
before/after measurements. Don't run other builds or profilers during sampling.
Functional assertions about observation boundaries should fail against the old
implementation; use the same sampling procedure when collecting a baseline.

## Measurement limits

Delivery-gap p50/p95/p99/max and counts over 25/50 ms measure **CADisplayLink
main-run-loop callback delivery**, not compositor FPS, GPU hitches, touch-to-key
latency or physical-device performance. The hosted view excludes the surrounding
app shell. UIKit input exercises editing/layout without synthesizing key taps.
The first keyboard presentation includes system warmup; residual system and
simulator scheduling gaps remain. Native gestures, read settling, header fade,
attachments and reading anchors are covered separately by the loopback UI tests.

`AgentRenderDiagnostics` counters are Debug-only and opt-in. They explain work;
a drop in counts alone does not establish improved responsiveness.

The shipped transcript retains its eager stack and exact row geometry. A lazy
experiment reduced measured work but failed the existing 2–3 point prepend
anchor checks. Retaining a pending anchor across geometry updates also failed;
an indefinite correction loop is not acceptable. The eager fallback keeps the
existing one-shot correction, isolates composer/read observation, and removes
row frame observers while following the bottom.

The experimental lazy content height changed from 43869 to 35870 points because
unmaterialized rows use estimates, not because content was shortened. Its
after-arrival screenshot was byte-identical to the eager baseline. Row text,
spacing, attachments and entrance growth were unchanged; the anchor failure
was sufficient to reject the lazy variant.

## Scrolling harness

`testThinkingScrollResponsiveness` retains the earlier three approximately
12-second phases: idle scroll, stationary thinking, and scrolling with changing
queue/worker/tool activity every 500 ms. Its display link sweeps 2400 points at
600 points/second. This programmatic sweep does not reproduce native drag phases;
use the UI tests for reading-anchor geometry and interactive keyboard dismissal.

The eager-stack reference in `b9a6f5d` has exact heights of 39181.333
points idle and 39223.333 while thinking, retained as regression assertions. That earlier scrolling fix reduced busy-scroll p95 from 88.8 to
19.4 ms; it did not address the input/entrance paths measured here. Physical
phone validation remains necessary for hardware performance claims.

## 2026-10-16 focused comparison

Single Debug simulator runs on the same iPhone 16e / iOS 26.5, starting at
`7510603` (which includes `b9a6f5d`). Values are callback-gap p95 milliseconds:

| Path | Before | Shipped eager isolation | Rejected lazy isolation |
| --- | ---: | ---: | ---: |
| 12 message entrances | 40.8 | 33.2 | 16.8 |
| Six keyboard cycles | 33.2 | 33.1 | 29.2 |
| Multiline composer typing | 24.8 | 25.2 | 21.3 |

Final transcript projections were 12/0/0 versus 220/24/10, with no row-frame
preference callbacks in these pinned-bottom phases. Eager content height stayed
43869.333 points. Entrance p99 improved 83.4→57.1 ms; keyboard p99 55.9→44.5;
typing p99 33.2→26.7. Residual gaps remain: entrance/keyboard had 10/3 gaps over
50 ms; typing had 30 over 25 ms (baseline 19). Keyboard p95 was essentially flat
and typing p95 did not improve. These single runs establish removed transcript
work and a safer entrance improvement, not universally smooth input or device FPS.
The faster lazy option was rejected because strict reading/prepend anchors failed.