qaspec 0.1.0

First release.

Added

• qaspec run, check, init, state list|clear CLI (single Rust binary).
• *.qa.ts spec format: TypeScript syntax parsed (never executed) into suites and ordered steps,
  with goal, expect(...) (LLM judge), deterministic expect.console, expect.errors,
  expect.network, expect.url, expect.state, expect.visible, expect.storage.local, and
  capture / capture.state / capture.url. Errors report file:line:col.
• Ordered steps that share browser state, with start, onFail (stop/continue/abort) and needs.
• Per-step signal windows: console errors, uncaught page errors and network requests are checked
  only for what happened during the step.
• One agent-browser session (one Chromium) per run, with one tab per project; --keep-open reuses it
  across runs.
• qaspec.toml: projects, per-environment base_url, health, depends_on, locale, identities with
  secrets (file/env/value), login spec or agent sign-in, valid_if, params with env overrides,
  --set, QASPEC_ENV, QASPEC_PARAM_*.
• Saved sign-in state per identity in .qaspec/state/<env>/ (mode 600, filtered by host), reused
  until valid_if fails.
• Agent loop over any OpenAI-compatible endpoint with tool calls; fill_secret types secrets through
  agent-browser's stdin batch mode and redacts them from logs and reports.
• Verdicts passed / failed / blocked / skipped with exit codes 0/1/2/3, a JSON report
  (report-1) and JUnit XML.
• Several identities per project in one run: the planner groups suites by identity, and a switch saves
  the current state, clears cookies, restores the other projects' identities and loads the new one.
• A suite can move between projects step by step; each project keeps its own tab, so returning to a
  project continues where its tab was.
• Login suites are reported as setup in the JSON report and are excluded from step totals.
• Example examples/shop (two dependent projects) and a runnable fixture app in tests/fixtures/app.
• Planning analysis and roadmap in plan/.

Fixed during validation

• Large agent-browser outputs (hundreds of KB of network requests on real apps) no longer deadlock the
  pipe: stdout and stderr are drained in threads.
• open on apps whose load event never settles (agent-browser times out after about 25s) counts as
  arrived when the tab is already on the target host.
• expect.visible reads the page's visible text. snapshot -c drops static text in agent-browser 0.27.
• The judge gets the visible text plus a larger snapshot, so long chat answers are not truncated away.
• Older tool results in the agent transcript are compacted, so token use does not grow with every action.

Known limitations

• No replay cache yet: every goal calls the model.
• Captures are shared across files in run order, but needs between files is not implemented.
• agent-browser 0.27 errors --clear is a no-op and logs are shared by tabs; qaspec works around it
  with offsets, so the per-step windows assume one active tab at a time.

Install

npm i -g agent-browser && agent-browser install
cargo install --git https://github.com/jaivial/qaspec --tag v0.1.0

Or download a prebuilt binary below.