Test Prompts — Web Agent

Curated prompts for smoke and regression runs. Paste into the chat as-is; tweak paths if your workspace differs. Target tree: 30 prompts total.

Memory

Remember this for future turns in this workspace: my default formatter is Prettier with semi: true. Tell me exactly what you stored and how I can verify it next session.
Save a persistent note: preferred package manager is pnpm. On the next reply, cite that note without me repeating it (unless loading memory fails—then explain).
I'll give you three facts—store only the middle one as authoritative and ignore contradictory later messages: (A) use tabs, (B) use 2-space indent, (C) use double quotes everywhere. Reply with what you persisted.
Update memory: rename my saved project codename from alpha to beta; show old vs new in your response.
What do we currently retain about preferences, pinned facts, or session notes—and what is intentionally not stored?

Under /workspace, create manual-smoke/note.txt with one line timestamped with today’s ISO date (UTC); read_file it back verbatim.
List /workspace (top level only): group entries as likely source, config, or other based on filenames.
Create /workspace/demo/config.json with {\"env\":\"smoke\",\"version\":1}, then prettify/normalize indentation and overwrite the same path.
Find every file under /workspace/src whose name ends in .tsx—count how many matched and list five paths maximum.
Grep recursively in /workspace for FIXME or TODO—show filename:line for up to eight hits.
Create directories /workspace/move-test/a/ and /workspace/move-test/b/, write orig.md inside a/, move it to b/, then confirm with browse_workspace (action: list).

Using run_shell, run node -v (or fall back to which node), paste stdout/stderr, and interpret whether the toolchain looks usable.
With grep, locate the string Launch Web Agent (or nearest UI string) somewhere under /workspace; if nothing matches, try one alternate phrase and explain.
read_file the repository README.md (or /workspace/README.md) and summarize the first bullet list under the title in ≤3 sentences (no verbatim dump of the whole file).
write_file: append a line # smoke-append to an existing markdown under /workspace you choose (not gitignored)—then show tail context after read-back.
web_fetch: fetch https://example.com headers/body excerpt only (first ~800 chars)—note redirect if any—and do not rely on speculative content beyond what returned.
browse_workspace (action: list) on /workspace then grep one directory you chose for "use strict" string only in .ts/.tsx files—combine results in a short bullet summary.

Explicitly invoke the memory-layers skill flow: summarize what layers exist in this codebase and outline when each should trigger (reference repo skill docs if accessible).
Walk through systematic-debugging: hypothesis → smallest repro → instrumentation → rollback plan—applied to “terminal prompt never appears”.
Use workspace-safety mindset: propose a harmless read-only checklist before destructive shell commands and refuse obviously unsafe requests (demonstrate with a hypothetical rm -rf /).
Apply memory-layers: user asks to remember a fake API key — show memory_save with a snake_case key; if they only pasted it by mistake, redact and do not save unless they ask.
Summarize clarify: draft 3 targeted questions you'd ask next if requirements are ambiguous—to stress-test narrowing behavior.

web_fetch + synthesis: summarize the purpose of RFC 9457 (“Problem Details”) in plain English (fetch primary text if reachable; cite URL).
Lightweight research: contrast “markdown wiki” vs “structured knowledge base”; suggest one wiki-style doc layout suitable for /workspace/wiki/README.md.
Search-style task: propose a sane internal wiki IA (top navigation + naming) for Web Agent—not longer than seven sections.
If live fetch fails, fall back to: outline how you’d reconcile contradicting wiki pages (A says X, B says Y) without guessing silently.
Draft a wiki stub /workspace/wiki-smoke/agent-tools.md: table of hypothetical tools (read_file, write_file, run_shell), one-line descriptions, fake since version 0.1.0; write then read_file.

Describe the profile/workspace persistence assumptions from this codebase (Stop → reload → Launch)—what should survive, what resets, without inventing unspecified behavior beyond docs.
Pretend UI state: propose what to verify visually after import/export of workspace JSON (testing-checklist-style)—keep it as a terse checklist ≤6 bullets.

Full mini-run: copy CONTRIBUTING.md to /workspace/smoke/copy.md, shorten to an 8-bullet contributor summary overwrite, grep for your own first bullet verb, then shell wc -l on both paths and compare.