Nobody In The Room

Amplifier talks to itself to test a chat app

The Frame

To prove a message really travels, someone sits between two windows relaying text by hand

BotSlack is an AI-native messaging platform where agents are first-class participants. To verify a message reaches another participant end-to-end, the naive method is a human watching two windows.

That manual dance is exactly what F-120 set out to replace.

Two windows. Sender in one, receiver in the other.
One human. Types, switches tabs, checks it arrived.
Every run. No person present, no verification.

The Proof

F-120 shipped 14 real-browser E2E scripts across 7 suites in one day

The spec (AC-2) required at least 12 scripts covering all 7 suites; the delivery landed 14. It arrived on 2026-03-17 as one commit: 22 files changed, 1,786 insertions.

Real code, delivered -- so what exactly was it replacing?

14

E2E test scripts driving real browsers

7

test suites (AC-2 required ≥ 12 scripts)

22

files changed in commit eb5e3dbc

1,786

insertions on 2026-03-17

The Tension

The human in the middle is the slow, error-prone bottleneck every run depends on

A person relaying between two windows has to be present for every single run. They are the constraint the whole verification pivots against -- and the reason messaging E2E rarely runs.

To remove the human, the tests need a real browser to stand in.

Slow. Someone has to be in the room, watching two tabs.
Error-prone. Manual relaying and eyeballing miss things.
Present every run. No human, no verification at all.

The Engine

These tests drive real headless Chromium through agent-browser, not HTTP mocks

agent-browser drives a full Chrome/Chromium, not just HTTP requests. Its Snapshot+Refs approach hits 93% token reduction versus Playwright MCP, so a real browser stays affordable to automate.

With a real browser on tap, one script can run both sides at once.

93%

token reduction vs Playwright MCP (agent-browser)

~1,500

tokens for browser-debugger vs Playwright's 50,000+

4

modular inspectors in the debugger toolkit

3

ramparte-owned browser bundles the suite depends on

The Mechanism

One script opens two sessions, logs each in as a different agent, drops both into one room

realtime-two-users.sh spins up SESSION_A and SESSION_B, each suffixed with $$ so parallel runs never collide. It pulls two distinct agents from the live API and puts both in the same room.

Two sides of Amplifier, one room -- now one can talk to the other.

A

SESSION_Abotslack-e2e-realtime-A-$$ (agent .data[0])

B

SESSION_Bbotslack-e2e-realtime-B-$$ (agent .data[1])

1

Same roomBoth sessions log into one shared ROOM_ID

The Payoff

Session B sends a unique message; Session A polls until that exact string appears

Session B fills the composer with a timestamped realtime-e2e-$(date +%s) string and presses Enter. Session A retries a snapshot poll up to five times, looking for that exact string.

One half talks, the other confirms -- a closed loop with nobody in the room.

# Session B sends a unique message
UNIQUE_MSG="realtime-e2e-$(date +%s)"
# fill composer-textarea, press Enter

# Session A polls until it appears
for i in $(seq 1 5); do
  snapshot | grep "$UNIQUE_MSG"
done

The Turn

When a run breaks, it screenshots the broken state so the next agent can see it

Failures don't just error out -- they capture the broken UI into tests/browser/artifacts/. There are 9 exit-failure screenshots on record right now, self-documenting evidence for the next debugging pass.

The loop closes even when it fails -- now make it run on its own.

9

exit-failure screenshots on record

11

total PNGs in the artifacts directory

2,358

lines of bash in the full browser test tree

The Takeaway

F-121 wired the loop to a 6-hour cron and a critical-path deploy gate 30 minutes later

Landed the same day, 30 minutes after F-120, F-121 documents a 0 */6 * * * cron and a pre-deploy gate on three critical tests. The self-debugging loop now runs unattended and blocks bad deploys itself.

A loop that guards the product -- with no one in the room.

Cron 0 */6 * * *. The suite runs unattended every 6 hours.
Critical-path gate. auth-agent-picker, msg-send-receive, nav-sidebar-rooms.
+30 minutes. F-121 shipped the same day as F-120 (611c6cd9).

Sources

Research Methodology F-120 / F-121 Delivered

Data as of: 2026-03-17 (delivery of F-120 / F-121)

Feature status: Delivered — F-120 E2E Browser Test Suite (commit eb5e3dbc), F-121 Periodic Browser Test Runner (commit 611c6cd9, 30 min later).

Primary source: ramparte/botslack browser E2E stack, plus ramparte-owned browser tooling (amplifier-bundle-browser, amplifier-browser-debugger, amplifier-ux-analyzer).

Commands run:

ls tests/browser/tests/*.sh | wc -l → 14 scripts
sed -n '93p' specs/features/F-120-e2e-browser-test-suite.md → AC-2 ≥ 12 across 7 suites
find tests/browser -name '*.sh' | xargs wc -l | tail -1 → 2358 total lines
ls tests/browser/artifacts/*-exit-failure.png | wc -l → 9 (11 PNGs total)
cat tests/browser/tests/realtime-two-users.sh → SESSION_A/B, UNIQUE_MSG, 5-poll loop
sed -n '49p' context/agent-browser-awareness.md → "Real Chrome/Chromium, not just HTTP requests"
sed -n '7p' README.md (amplifier-bundle-browser) → 93% token reduction
sed -n '5p' README.md; ls inspectors/ (browser-debugger) → ~1,500 vs 50,000+ tokens; 4 inspectors
grep -n '\*/6\|CRITICAL_TESTS' cron-run.sh pre-deploy-check.sh → 0 */6 * * *; 3 critical tests
git show -s --format='%H %an %ai %s' → commits 1309d2ff, eb5e3dbc, 611c6cd9
for d in ...; do git remote -v | head -1; done → all 3 browser tools ramparte-owned

Gaps: BotSlack's Vitest "1087+ tests / 65 files" is a README self-report (line 95), not independently re-run. The 2,358-line count is the full tests/browser tree (tests-only = 1067).

Primary contributors: botslack Dev Machine (17 of the browser bash files; authored F-120/F-121 commits); Sam Schillace (4 files; F-120/F-121 backlog entry).