DEV Community: FRANCISE LIANG

Build Time vs Edit Time — My Web Agent Can Publish But Can't Edit (An Honest Capability-Boundary Spike)

FRANCISE LIANG — Wed, 06 May 2026 00:55:53 +0000

TL;DR: My web-agent used itself to publish 2 blog posts to dev.to (full dogfooding story). But when I asked it to edit the already-published article to add a cross-link, it ran 7 steps before V0.5.0 anti-loop hard-aborted. The 5-action actuator (click/type/scroll/extract/done) lacks keyboard_shortcut / paste / textarea range API. Failure itself is data.

7 sections + 2 mermaid figures (capability-boundary flowchart + spike-and-decide flowchart).

Full version (with V0.16.31 dogfooding 4/5 = 80% E2E success rate cumulative table + V0.17+ trigger conditions):
https://github.com/franciseliang99-dot/web-agent/blob/main/docs/blog-drafts/2026-05-build-vs-edit-time-final-en.md

Project repo (MIT, ⭐ welcome):
https://github.com/franciseliang99-dot/web-agent

Series:

https://dev.to/francise_liang_e4544eadb9/50-compliance-not-0-how-a-logging-spike-almost-triggered-the-wrong-architecture-rewrite-1lna — measurement layer
https://dev.to/francise_liang_e4544eadb9/why-i-permanently-no-god-patchright-after-a-spike-and-the-anti-detection-decision-tree-3m11 — architecture layer
(this post) — capability-boundary layer

LLM thought self-reported 'Tool limitation (no keyboard shortcut like Ctrl+End)' — aware of boundary. Anti-loop fired = user data protected.

Why I Permanently NO-GO'd Patchright After a Spike (And the Anti-Detection Decision Tree)

FRANCISE LIANG — Tue, 05 May 2026 22:28:24 +0000

TL;DR: My web-agent runs sannysoft.com anti-bot tests. vanilla playwright passes 19/32, with playwright-stealth it's 21/32, but patchright (claimed stronger) — still 19/32, identical to vanilla. The reason: patchright's patches all sit in the launch phase, but I take over already-launched Chrome via connect_over_cdp — entirely bypassed.

8 chapters + 3 mermaid figures (xychart bar comparison + 2 decision flowcharts).

Full version (with reproducible spike commands):
https://github.com/franciseliang99-dot/web-agent/blob/main/docs/blog-drafts/2026-05-patchright-nogo-final-en.md

Project repo (MIT, ⭐ welcome):
https://github.com/franciseliang99-dot/web-agent

Anti-detection layered model: JS injection / CDP / TLS / IP reputation. Architecture decisions decide which layers are off the menu.

50% Compliance, Not 0%: How a Logging Spike Almost Triggered the Wrong Architecture Rewrite

FRANCISE LIANG — Tue, 05 May 2026 21:33:38 +0000

TL;DR: I thought my prompt augmentation track was completely broken (compliance=0%) and almost spent 27 hours rewriting plan-and-execute. Turns out the 0% was a regex bug — real value was 50%.

7 chapters + 4 figures (hero / decision matrix quadrantChart / 7-version timeline / data delta tables).

Full version (with mermaid quadrantChart + reproducible spike commands):
https://github.com/franciseliang99-dot/web-agent/blob/main/docs/blog-drafts/2026-05-w5c2-spike-story-final-en.md

Project repo (MIT, ⭐ welcome):
https://github.com/franciseliang99-dot/web-agent

LLM engineering + spike decision methodology + Chrome GPU deadlock diagnosis + regex false-negative lessons. Might save you 27h.

Related: Why I Permanently NO-GO'd Patchright After a Spike (And the Anti-Detection Decision Tree) — V0.16.14 anti-detection spike + 4-layer decision tree story.