Agent Console production workload profile

Final status

Balanced six-run A/B/C audit complete on the clean measurementRef recorded in the committed JSON, with one canonical production build, exact visual-index boundaries, and strict correctness/performance gates. Each variant appears twice in every ordinal position across all six permutations: ABC/ACB/BAC/BCA/CAB/CBA.

Variants:

• A: copied/deep-reactive replay history + eager Markdown publication.
• B: shallow mutable replay backing + eager Markdown publication.
• C: shallow replay backing + lazy Markdown publication (final implementation).

The final implementation keeps all three measured application-level optimizations. It does not change Cell/text caches, TLogView architecture, long-text admission, renderer architecture, or virtual scrolling.

Findings and decision

Replay copying was a real application hotspot. Lazy Markdown publication removed a second hidden hotspot: Log mode previously materialized markdownSource.blocks after every invisible Markdown delta.

Workload	A median	B median	C median	C vs A
CLI framed burst	2,487 ms	1,122 ms	435 ms	82.5% faster
Browser framed burst	1,998 ms	1,095 ms	857 ms	57.1% faster
CLI single-task burst	1,912 ms	755 ms	87 ms	95.4% faster
Browser single-task burst	1,619 ms	687 ms	38 ms	97.6% faster

The real 12 ms product workloads remain stable: product-tail frame p95 is 1.70 → 2.34 ms CLI and 0.80 → 0.80 ms Browser; product Markdown frame p95 is 1.91 → 1.65 ms CLI and 1.00 → 0.90 ms Browser. Product stream + scroll input-to-commit p95 changes by a paired median of +0.02 ms CLI and −0.30 ms Browser; Browser DOM-flush and paint-opportunity p95 deltas are −0.45 ms and −0.40 ms.

In C's default Log burst, mergeGroups no longer dominates CPU samples. Scenario-specific preludes occur before counters reset and timing starts. Visible Markdown publication is rate-limited to a 32 ms minimum interval through one timer and one fixed-id low-priority frame task; mode exit and disposal cancel both stages. The controlled synthetic producer remains at 64 ms. Three separate product scenarios execute the real startStream() / stopStream() 12 ms timer for fixed tick counts. Formal benefits use the inner workload totalElapsedMs, paired by round; Playwright controller time remains diagnostic only. validate:agent-console:abc gates cadence, frames, interaction latency, Long Tasks, DOM/stdout amplification, CPU artifacts, correctness, and provenance. The committed schema-5 baseline retains paired frame, long-frame, Long Task, latency, and amplification evidence for independent checking.

For the canonical Agent Console workload measured here, no evidence justifies changing core Cell/text/wrap/provider caches, renderer architecture, long-text admission, or virtual scrolling. The current initiative closes with those areas unchanged.

Canonical workloads

All CLI and Browser runners use the same validated config: seed 6,000; append 1,000; steady 400; cadence 64 ms; batch size 10; six paired runs.

Controlled synthetic application workloads:

1. tail-stream-steady
2. tail-append-burst-framed
3. tail-append-burst-single-task
4. detached-append
5. search-large-history
6. stream-scroll-interaction
7. markdown-toggle-large-history
8. markdown-stream-steady

Real product-timer workloads:

9. product-tail-stream-12ms
10. product-markdown-stream-12ms
11. product-stream-scroll-interaction-12ms

Controller diagnostic (not represented as a complete application workload):

12. markdown-publication-burst-diagnostic

Synthetic append indices use an explicit post-seed cursor. Product scenarios call startStream() and stopStream(), verify the 12 ms product default, stop after exactly 400 timer ticks, preserve input, and account for the connected/paused status events. Prepared and final visual indices must be exact; the product Markdown scenario also verifies that the final paused marker was published. Detached append compares the complete visible viewport. Both concurrent interaction workloads record at least 100 fully correlated wheel samples per run; every recorded sample changes scrollTop, matches a reason: scroll frame, and—on Browser—matches a subsequent DOM flush. Wheel input runs independently at 16 ms and repeatedly reverses direction.

Production and diagnostics

• CLI loads built package exports in a fresh Node child for each run.
• Browser uses a Vite production build and the Vite preview JS API, with a fresh BrowserContext/Page per run and deterministic shutdown.
• CPU diagnostics are separately collected for typed registered scenarios; every expected .cpuprofile and hotspot summary is validated.
• CPU shares use all sampled time as the denominator (shareOfAllSamples); top-N-relative share is separately named.
• Memory values are explicitly marked includesProfilerBuffers: true and are diagnostic only, not clean retained-app evidence.
• Browser Long Task collection drains takeRecords() before disconnect.
• Environment, commit/dirty state, artifact hashes, corpus, correctness, frames, DOM/stdout, latency, memory, and CPU summaries are retained for A/B/C.
• CPU URLs in committed evidence are normalized to <repo>/....

Evidence and commands

The complete machine-readable summary is agent-console-profile-baseline.json. Raw frames and CPU profiles remain workflow artifacts.

pnpm run profile:agent-console:smoke
pnpm run profile:agent-console
pnpm run profile:agent-console:abc
pnpm run validate:agent-console:abc
pnpm run profile:agent-console:record
pnpm run check:agent-console-profile-baseline

The validator requires six passing runs, clean provenance, non-empty artifact hashes and CPU diagnostics, identical canonical corpus/config, exact prepared/final indices, and fully correlated input samples.