Browser Automation
For applicable interfaces, use WCAG 2.2, WAI-ARIA Authoring Practices, ISO 9241-110/210, current ISO 9241-112:2025 information-presentation principles, and ISO 9241-171:2025 software-accessibility guidance. Apply ISO/IEC 23859:2023, ISO 21801-1:2020, and ISO/IEC 29138-1/-4 when UI text, cognition, or user accessibility needs can block the task. Prefer native semantics before custom ARIA.
- Confirm the authorized site, account, environment, intended users, user promise, journeys, and allowed side effects. For a material accessibility barrier, trace
user accessibility need → barrier → journey or requirement → evidence.
- Use an isolated profile or clean context by default. Use a real signed-in profile only when required and authorized; do not copy cookies, tokens, or unrelated session data.
- Establish the real runtime path. Start permitted local services, then inspect the rendered page, accessibility tree, or screenshot before choosing selectors. Wait for a meaningful ready condition, not a fixed sleep.
- Start critical journeys from known state. Test reload, reopen, stale state, session expiry, and interruption when persistence or recovery matters. In multistep flows, verify that a returning user can identify completed, current, and pending work and that important input is preserved.
- Use semantic or stable locators. Keep actions small and verify navigation or mutation through visible state, DOM, accessibility tree, network, console, saved files, screenshots, or target-system read-back.
- Exercise the few states most likely to fail across representative viewports or window sizes: loading, empty, error, disabled, focus, resize, invalid input, slow or failed network, cancellation, and recovery. For interactive accessibility scope, verify keyboard operation, focus order, accessible name, role, state, and visible result. For important UI text or errors, verify that the user can identify the purpose, next action, expected result, and whether work or data was preserved.
- Retry only known transient failures. Record intent before a consequential submission; after an uncertain result, inspect state before retrying.
- Capture concise reproducible evidence without secrets. For reference-driven interface work, compare rendered output side by side or with an overlay and fix the largest meaningful mismatch first. Source inspection is not proof that a user journey works.
- Convert durable manually observed behavior into the smallest regression test with semantic locators and an outcome assertion.
- Report
PASS, FAIL, or BLOCKED per independent journey, with environment and evidence.
Do not bypass access controls, anti-abuse systems, CAPTCHA, or consent. Do not purchase, publish, send, delete, or mutate production without explicit authorization.
User-facing: Apply the global outcome-first delivery overlay. State supported conclusions directly; avoid litotes and rhetorical hedging that obscure status or responsibility. Preserve genuine uncertainty, evidence scope and degree, logical negation, quotations, and requested artifact voice. Own actual agent errors without inventing blame; give the correction or next action within existing permissions. Match reply length and structure to the weight of the ask. Investigate enough internally to be right, but report only the useful outcome, fresh verification, material uncertainty, and remaining user action; do not replay routine tool calls or internal process. Simple turns stay short. For substantive chat, use Summary and TL;DR when required by the active user or host contract or when they improve navigation; each MUST add distinct value and MUST NOT repeat the same conclusion. Apply ASD-STE100, ISO 24495-1, and W3C COGA proportionally. Add Feynman, Diátaxis, or BCP 14 only when their function applies. Use truthful named 20-cell progress separate from verdict. Preserve machine and artifact formats. Be considerate, avoid surprise scope, and leave the result ready to use or resume.
1---2name: browser-automation3description: Build or run authorized browser automation, real-user QA, data entry, or extraction with stable locators, explicit state, bounded retries, and evidence of the user-visible result.4---56# Browser Automation78For applicable interfaces, use **WCAG 2.2**, **WAI-ARIA Authoring Practices**, **ISO 9241-110/210**, current **ISO 9241-112:2025** information-presentation principles, and **ISO 9241-171:2025** software-accessibility guidance. Apply **ISO/IEC 23859:2023**, **ISO 21801-1:2020**, and **ISO/IEC 29138-1/-4** when UI text, cognition, or user accessibility needs can block the task. Prefer native semantics before custom ARIA.9101. Confirm the authorized site, account, environment, intended users, user promise, journeys, and allowed side effects. For a material accessibility barrier, trace `user accessibility need → barrier → journey or requirement → evidence`.112. Use an isolated profile or clean context by default. Use a real signed-in profile only when required and authorized; do not copy cookies, tokens, or unrelated session data.123. Establish the real runtime path. Start permitted local services, then inspect the rendered page, accessibility tree, or screenshot before choosing selectors. Wait for a meaningful ready condition, not a fixed sleep.134. Start critical journeys from known state. Test reload, reopen, stale state, session expiry, and interruption when persistence or recovery matters. In multistep flows, verify that a returning user can identify completed, current, and pending work and that important input is preserved.145. Use semantic or stable locators. Keep actions small and verify navigation or mutation through visible state, DOM, accessibility tree, network, console, saved files, screenshots, or target-system read-back.156. Exercise the few states most likely to fail across representative viewports or window sizes: loading, empty, error, disabled, focus, resize, invalid input, slow or failed network, cancellation, and recovery. For interactive accessibility scope, verify keyboard operation, focus order, accessible name, role, state, and visible result. For important UI text or errors, verify that the user can identify the purpose, next action, expected result, and whether work or data was preserved.167. Retry only known transient failures. Record intent before a consequential submission; after an uncertain result, inspect state before retrying.178. Capture concise reproducible evidence without secrets. For reference-driven interface work, compare rendered output side by side or with an overlay and fix the largest meaningful mismatch first. Source inspection is not proof that a user journey works.189. Convert durable manually observed behavior into the smallest regression test with semantic locators and an outcome assertion.1910. Report `PASS`, `FAIL`, or `BLOCKED` per independent journey, with environment and evidence.2021Do not bypass access controls, anti-abuse systems, CAPTCHA, or consent. Do not purchase, publish, send, delete, or mutate production without explicit authorization.222324**User-facing:** Apply the global outcome-first delivery overlay. State supported conclusions directly; avoid litotes and rhetorical hedging that obscure status or responsibility. Preserve genuine uncertainty, evidence scope and degree, logical negation, quotations, and requested artifact voice. Own actual agent errors without inventing blame; give the correction or next action within existing permissions. Match reply length and structure to the weight of the ask. Investigate enough internally to be right, but report only the useful outcome, fresh verification, material uncertainty, and remaining user action; do not replay routine tool calls or internal process. Simple turns stay short. For substantive chat, use **Summary** and **TL;DR** when required by the active user or host contract or when they improve navigation; each MUST add distinct value and MUST NOT repeat the same conclusion. Apply **ASD-STE100**, **ISO 24495-1**, and **W3C COGA** proportionally. Add Feynman, Diátaxis, or BCP 14 only when their function applies. Use truthful named 20-cell progress separate from verdict. Preserve machine and artifact formats. Be considerate, avoid surprise scope, and leave the result ready to use or resume.
zznabil