perf(cmux-tui): bench interact, named budgets, record-only perf lane (IX0) by lawrencecchen · Pull Request #11688 · manaflow-ai/cmux

lawrencecchen · 2026-09-02T14:10:38Z

What this is

IX0 of the zero-wait interaction plan (hq plans/cmux-tui-zero-wait-interaction.md): measure and name before changing behavior. Three commits, no runtime behavior change.

Named budgets. cmux_tui_core::budgets holds every bounded wait the daemon, terminal hosts, and clients enforce, one value per budget, with the interaction stage it belongs to (accept, durable, settle, frame, client, planned) and the code site. The existing constant sites import these values instead of repeating the numbers. cmux diag budgets [--json] prints the table locally, no session needed. No value changed.
cmux bench interact. A client that drives a session over the raw control protocol and records the latencies a frontend or an agent feels per user intent. It sends only existing commands; it adds no protocol command or resource operation.
Record-only perf lane. A full-mode bench interact job runs the benchmark on Linux and macOS, prints the table, and uploads the JSON as cmux-tui-bench-interact-<os>. It is deliberately not in hosted-verification's needs, so it is never a required check. Plus docs/journal-operations.md.

What the bench measures

Per create (round-robin new-workspace / new-tab / split), all from one clock: request to response; request to the first tree_events:"deltas" delta that references the new surface on a separate subscriber (deltas may precede the response, so both time from the write); attach-surface to first render-state; view-only close and process-terminating close to response. Typing: one-byte send on a separate connection and on the create connection, so head-of-line blocking is visible. --clients N runs N concurrent create loops.

Baseline (Linux testbox, 32 vCPU, throwaway session; ms)

The convoy the plan predicts is already visible: at 8 concurrent create loops, create response p50 rises 8x and p99 to ~2.1 s, because every create serializes behind the daemon-wide creation fences, while typing to an existing terminal stays flat.

metric	1 client p50	1 client p99	8 clients p50	8 clients p99
create.response_ms	27.97	67.79	236.53	2128.44
create.visible_ms	27.13	67.86	229.00	2128.61
create.first_frame_ms	0.28	0.37	0.44	24.93
close.surface_response_ms	2.62	3.54	214.55	1536.98
close.terminal_response_ms	3.17	4.40	203.47	2536.94
typing.separate_conn_ms	0.02	0.25	0.02	0.06
typing.same_conn_ms	0.02	0.04	0.02	0.04

macOS arm64 baseline (from the full-run binary) is added below once the full hosted run finishes.

Verification

Focused hosted (--filter budgets) and full hosted: URLs below.
clippy clean for cmux-tui-core and cmux-tui on the testbox.
Unit tests: budgets table/value/stage checks, diag budgets JSON round-trip, bench percentile and delta-matching.

Design: hq plans/cmux-tui-zero-wait-interaction.md (IX0).

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

Summary by cubic

Adds IX0 of the zero-wait interaction plan: names every bounded wait as a budget and adds an interaction-latency benchmark with a record-only CI lane. No runtime behavior changes.

New Features

cmux_tui_core::budgets is the single source for every bounded wait; cmux diag budgets [--json] prints it with stage and code site.
cmux bench interact drives a session as a client, sending only existing commands, and records create, visibility, first-frame, close, and typing latencies.
A record-only CI job runs it on Linux and macOS, uploads the JSON as cmux-tui-bench-interact-<os>, and is never required; docs/journal-operations.md documents both.
Baseline on the testbox shows create response p50 rising 8x at 8 concurrent clients while typing stays flat.

Verification

Focused and full hosted verification passed; clippy clean; unit tests cover budgets, diag JSON, bench percentiles, and visibility matching.
A new test requires the public resource boundary scan to be clean.

^{Written for commit 738f47a. Summary will update on new commits.}

vercel · 2026-09-02T14:10:44Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
cmux166	Canceled		Sep 3, 2026 12:45pm UTC
cmux41	Canceled		Sep 3, 2026 12:45pm UTC

coderabbitai · 2026-09-02T14:10:49Z

Warning

Review limit reached

Next included review available in 5 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used all 10 included reviews currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Team

Run ID: e517fcbf-5106-41e4-8757-dfe12be14207

📥 Commits

Reviewing files that changed from the base of the PR and between f1c912b and 738f47a.

📒 Files selected for processing (19)

.github/workflows/cmux-tui.yml
cmux-tui/crates/cmux-tui-core/src/budgets.rs
cmux-tui/crates/cmux-tui-core/src/journal_ingress.rs
cmux-tui/crates/cmux-tui-core/src/lib.rs
cmux-tui/crates/cmux-tui-core/src/mux.rs
cmux-tui/crates/cmux-tui-core/src/server.rs
cmux-tui/crates/cmux-tui-core/src/terminal_host_runtime.rs
cmux-tui/crates/cmux-tui/src/app.rs
cmux-tui/crates/cmux-tui/src/cli.rs
cmux-tui/crates/cmux-tui/src/cli/command.rs
cmux-tui/crates/cmux-tui/src/cli/diag.rs
cmux-tui/crates/cmux-tui/src/cli/internal/bench.rs
cmux-tui/crates/cmux-tui/src/cli/internal/mod.rs
cmux-tui/crates/cmux-tui/src/local_owner.rs
cmux-tui/crates/cmux-tui/src/session/remote.rs
cmux-tui/docs/README.md
cmux-tui/docs/journal-operations.md
cmux-tui/scripts/test_check_resource_api_boundary.py
cmux-tui/spec/cli.md

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

github-actions · 2026-09-02T14:10:54Z

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

…ag budgets One module, cmux_tui_core::budgets, holds every bounded wait the daemon, terminal hosts, and clients enforce. The existing constant sites import these values instead of repeating the numbers, so there is exactly one value per budget. 'cmux diag budgets' prints the table locally with the stage each budget belongs to and the code site that enforces it. No value changes. IX0 of the zero-wait interaction plan.

…chmark cmux bench interact drives a session as an ordinary client over the raw control protocol and records the latencies an interactive frontend or an agent feels per user intent: create request to response, request to the tree delta that makes the resource visible on a separate deltas subscriber, attach to first render frame, close to response, and one-byte typing on both a separate connection and the create connection (so head-of-line blocking is visible). --clients N runs N concurrent create loops. With no socket or session it starts and stops a throwaway session. It sends only existing commands and adds no protocol command or resource operation. IX0 of the zero-wait interaction plan.

…ns doc A full-mode 'bench interact' job builds the server, runs the benchmark against a throwaway session on Linux and macOS runners, prints the table in the job log, and uploads the JSON as cmux-tui-bench-interact-<os>. It is deliberately not in hosted-verification's needs, so it is never a required check; it records the IX0 baseline, not a threshold. Adds docs/journal-operations.md describing the budgets verb and the bench metrics. IX0 of the zero-wait interaction plan.

lawrencecchen · 2026-09-02T19:01:42Z

This benchmark series is superseded by #11697, which contains the visibility and barrier follow-up. The current head also fails the live conformance compile and is stale against main. Closing the duplicate; #11697 remains held for a current-main safety and observability repair.

lawrencecchen · 2026-09-02T19:01:44Z

Closed as superseded by #11697.

lawrence703 and others added 5 commits September 2, 2026 07:48

test(cmux-tui): require public boundary scan to be clean

13b29fc

fix(cmux-tui): keep raw bench protocol code internal

738f47a

lawrencecchen force-pushed the feat-tui-bench-interact branch from b00bae9 to 738f47a Compare September 2, 2026 14:48

lawrencecchen mentioned this pull request Sep 2, 2026

fix(cmux-tui): measure bench probes during in-flight creates #11697

Open

lawrencecchen closed this Sep 2, 2026

vercel Bot deployed to Preview – cmux166 September 3, 2026 12:44 View deployment

vercel Bot deployed to Preview – cmux41 September 3, 2026 12:45 View deployment

Navigation Menu

Repository navigation

perf(cmux-tui): bench interact, named budgets, record-only perf lane (IX0) - #11688

perf(cmux-tui): bench interact, named budgets, record-only perf lane (IX0)#11688
lawrencecchen wants to merge 5 commits into
mainfrom
feat-tui-bench-interact

lawrencecchen commented Sep 2, 2026 •

edited by cubic-dev-ai Bot

Loading

Uh oh!

vercel Bot commented Sep 2, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Sep 2, 2026 •

edited

Loading

Review limit reached

Uh oh!

github-actions Bot commented Sep 2, 2026 •

edited

Loading

Uh oh!

lawrencecchen commented Sep 2, 2026

Uh oh!

lawrencecchen commented Sep 2, 2026

Uh oh!

This branch was successfully deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

lawrencecchen commented Sep 2, 2026 • edited by cubic-dev-ai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this is

What the bench measures

Baseline (Linux testbox, 32 vCPU, throwaway session; ms)

Verification

Summary by cubic

Uh oh!

vercel Bot commented Sep 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coderabbitai Bot commented Sep 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review limit reached

Uh oh!

github-actions Bot commented Sep 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lawrencecchen commented Sep 2, 2026

Uh oh!

lawrencecchen commented Sep 2, 2026

Uh oh!

This branch was successfully deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

lawrencecchen commented Sep 2, 2026 •

edited by cubic-dev-ai Bot

Loading

vercel Bot commented Sep 2, 2026 •

edited

Loading

coderabbitai Bot commented Sep 2, 2026 •

edited

Loading

github-actions Bot commented Sep 2, 2026 •

edited

Loading