Diagnose a live webpage with DeepSeek Harness

Use only native Lexmount tools already loaded in this DSH process for webpage navigation and interaction. DSH's built-in read-only file and Git inspection may be used only to determine whether the current workspace is related to the target source.

Compatibility constraint: the current DeepSeek model does not accept image input. Do not call lexmount_browser_screenshot or any tool that returns image content to the model, and do not generate PNG, JPEG, or GIF files. All evidence must come from text/DOM results returned by lexmount_browser_snapshot, values returned by lexmount_browser_evaluate, Recording metadata, and text/JSON files. Recording is retained for human replay only; do not read it, extract frames, or attach its content to the conversation.

Preflight: if lexmount_doctor is unavailable, stop with SETUP_REQUIRED and tell the user to run npx --yes @deepseek-ai/dsh@0.1.0-rc.6 plugin --profile web add @lexmount/dsh-browser@next, restart DSH Web, and resend this prompt. Do not install the Bundle into the running process or switch browser providers.

If the tool exists, call lexmount_doctor first. When credentials are missing, call lexmount_auth_login and wait for browser authorization, then call lexmount_doctor again. Continue only when it is ready.

Task: validate a dynamic UI performance report against Selenium's official test page at https://www.selenium.dev/selenium/web/dynamic.html . After clicking “Reveal a new input,” the #revealed input must become visible within 500ms; more than 500ms reproduces the SLA violation. The official reference source is https://github.com/SeleniumHQ/selenium/blob/trunk/common/src/web/dynamic.html .

Flow:
1. Create artifacts/dsh-dynamic-ui-triage in the working directory. If it exists, append a timestamp instead of overwriting prior results.
2. Call lexmount_session_create to create exactly one temporary normal-mode Session with Recording enabled, and confirm that Recording is active. Open the target URL and use text/DOM results from lexmount_browser_snapshot to verify that #reveal exists and #revealed is initially hidden. Write the URL, accessed_at_utc, selector presence, visibility, and relevant DOM text to evidence/01-before.json.
3. Do not use a fixed sleep. First use lexmount_browser_evaluate to install a read-only timing probe: capture performance.now() during the #reveal click event and use a MutationObserver to record when #revealed becomes visible. Then click #reveal with lexmount_browser_click and use a bounded evaluate to read elapsed_ms. Write the timing start, end, elapsed_ms, sla_ms, and violated fields to evidence/02-timing.json. Call lexmount_browser_snapshot again to verify that the input is visible, then write its visibility and relevant DOM text to evidence/03-after.json.
4. Use lexmount_browser_evaluate to extract the smallest script text containing reveal() and setTimeout from the live DOM, and write it to evidence/04-deployed-source.txt. Then open the official GitHub reference source above, locate the corresponding function in text returned by lexmount_browser_snapshot, and write it to evidence/05-reference-source.txt. Record the former as a deployed_source_fact and the latter as a reference_source_fact. Do not assume the repository trunk is the exact deployed commit.
5. Inspect the current workspace Git remote read-only. Record workspace_code_facts with file paths and line numbers only when it clearly points to SeleniumHQ/selenium and common/src/web/dynamic.html exists. Otherwise set workspace_relation to unrelated or unknown and do not scan unrelated local code.
6. Write report.md, report.json, evidence/*.json, and evidence/*.txt. report.json must separate browser_facts, deployed_source_facts, reference_source_facts, workspace_code_facts, hypotheses, recommended_next_steps, recording, and cleanup. Return REPRODUCED when elapsed_ms is greater than 500 and NOT_REPRODUCED otherwise. In either case, do not describe the test page's intentional delay as an unknown production incident.

Acceptance: doctor is ready; Recording is confirmed; no tool returning image content was called; text evidence exists for the click and timing; elapsed_ms is finite; 01-before.json, 02-timing.json, and 03-after.json are present, non-empty, and parseable; 04-deployed-source.txt and 05-reference-source.txt are present and non-empty; both reports exist and are non-empty; report.json parses; every fact has an evidence category; the Session is closed; and the Recording is retained. Return PARTIAL_SUCCESS if cleanup fails.

Cleanup: on success, failure, timeout, or cancellation, close only the Session created by this task. Do not close pre-existing Sessions, delete the Recording, modify code, or perform purchases, publishing, deletion, or any other irreversible action.

Return only status, REPRODUCED / NOT_REPRODUCED, elapsed_ms, the SLA threshold, first failed step, report directory, text evidence file count, Recording retention, cleanup result, and a sanitized error summary. Do not expose API keys, Project IDs, Session IDs, Recording IDs, connection URLs, or OAuth details.