12 TDD tasks: request counter, progress rendering, counter wiring,
prompter abstraction, run_init extraction, seven wizard sections,
orchestrator, aai onboard command, help reordering + bare-aai offer,
hint updates, and snapshot/gate.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Template requirements.txt files listed bare package names (fastapi,
uvicorn, httpx2, python-dotenv, ...). SCA scanners read a starter app's
requirements.txt as a lockfile and report unpinned lines as "missing
versions", blocking vulnerability analysis.

Pin every template dependency to a >= floor matching the project's own
(pyproject deps), and add a _requirements_pin_versions check to both the
template contract gate (the check.sh stage) and the parametrized pytest
contract test so a future unpinned dep fails loudly.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ctions

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Adds the per-command help snapshot for the new `aai onboard` command. The
root --help ordering and login-success hint are pinned by assertion-based
tests, not syrupy, so only this one snapshot changed.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Two gate fixes for the onboard work:

- mypy strict (no_implicit_reexport): switch monkeypatch targets to the
  string-target form (e.g. "aai_cli.main.typer.confirm") and import the
  prompter classes from their source module instead of re-importing them
  through the command module.

- mutation gate (diff-scoped): add assertions that kill the surviving
  mutants on changed lines — pin the onboard exit-code guard (code != 0),
  auto_login=False, the request-counter `or 0` fallback,
  _interactive_session's `and`, the bare-aai offer's confirm default and
  json_mode, the prompter confirm defaults, cursor restore (show=True), the
  exact run_init/claude-setup kwargs, the doctor render `ok=True`, and
  init's `api_key is None` launch guard (via a --json step assertion).
  One genuinely-equivalent mutant — the Prompter Protocol stub default — is
  marked `# pragma: no mutate` (not a forbidden escape hatch).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… source

Drop the 'N of 100 requests' counter entirely (no accurate way to track
request counts) and make the wizard's first transcription prompt for a
file path or YouTube URL, defaulting to the hosted sample on Enter.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The AMS usage endpoint returns total:0.0 on every window; derive each
window's spend from line_items[].price (cents) instead and render it as
dollars. Show a clear message when an account has no custom rate limits.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Update the onboard --help snapshot (drops the --status option and its
"progress toward 100 requests" example) and assert the first-transcription
status label so the diff-scoped mutation gate kills the source-vs-sample
Or->And mutant on sections.py:76.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- Remove the `samples` command and all related code (module, tests, help
  snapshots, import-linter contract, README/AGENTS/skill docs).
- Remove the `version` command; keep `--version`/`-V`. Repoint smoke,
  __main__, and install-script tests to the flag.
- Shorten the `--show-completion` and `init` help text so they stop wrapping.
- Onboarding: browser-only sign-in (drop the API-key paste path so a secret
  never lands in scrollback/history), fix the duplicated "Environment check"
  heading, space the setup prompt off the banner, and reword the source and
  build-path prompts.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Rename the "Transcription & AI" panel to "Run AssemblyAI" and give `init`
its own "Build an app" panel, so the help screen distinguishes commands that
use AssemblyAI directly from the one that scaffolds a new project. Reword
init's help to "Scaffold a new project from a template, then launch it."

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Match the casing of the other multi-word panels (Quick Start, Setup & Tools).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Implements the prioritized recommendations from the persona journey review:

- transcribe: add --out FILE to save a clean text artifact; annotate -o
  values; document the human-by-default --json policy; lead the docstring
  with the simplest invocation; --show-code now works with no source
  (emits a placeholder path).
- usage: show the per-product breakdown in dollars (from line_items[].price),
  aggregated by product and reconciling to the window total — no more raw
  quantities mixed with dollars.
- env: the mismatch warning now also fires for AAI_ENV (not just --env) and
  names the source, catching silent sandbox/prod key swaps.
- errors: root-callback failures (e.g. bad --env) honor a --json/-o json
  request and emit the uniform {"error": ...} shape; soften the not-signed-in
  message and point it at `aai onboard`.
- stream/agent: document stdin raw-PCM streaming (`aai stream -` + ffmpeg),
  cross-reference the in-process vs piped LLM paths, surface the headphones
  caveat in `agent --help`.
- setup: enumerate the three installed artifacts (docs MCP, AssemblyAI skill,
  bundled aai-cli skill) in help.
- limits: clarify the empty state (standard limits apply).
- root help: document that global options go before the subcommand.
- templates: comment the intentional httpx2 import so integrators don't "fix" it.
- AGENTS.md: fix the stale `aai claude` reference (now `aai setup`).

Snapshots regenerated; new tests added; full gate (ruff/mypy/pyright/vulture/
deptry/import-linter/diff-cover 100%/diff-scoped mutation) green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Audited every command's `--help` Examples panel and reworked them to
surface high-value capabilities that were previously undiscoverable:

- root: lead with breadth (transcribe/stream/agent) + a `--llm` one-liner
- transcribe: add YouTube input, speaker labels, PII redaction,
  summarization, and `--llm` Q&A; drop the redundant `-o text` example
- stream: add file/URL source, speaker labels, and keyterm prompts
- agent: show `--system-prompt` persona
- llm: add model + `--system` example, reword transcript-id example
- transcripts/sessions: add `--json | jq` pipelines and id-chaining
- account/keys/audit: add `--json | jq` scripting recipes; fix the
  `audit --limit 20` example that showed the default as if required
- init: show the voice-agent template and `--no-install`
- doctor/setup: add `--json` and `--force`/`--scope` examples

No piping into `aai llm` and no `--sandbox` examples per review.
Snapshots regenerated.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-authored-by: aikido-pr-checks[bot] <169896070+aikido-pr-checks[bot]@users.noreply.github.com>

…-flow

The starter-app requirements.txt pins used floors low enough to admit
known-vulnerable releases (python-multipart>=0.0.9, python-dotenv>=1.0.0),
and fastapi>=0.115.0 transitively pins starlette<0.39.0 — pulling a
starlette with CVE-2024-47874 / CVE-2025-54121 / CVE-2026-48710.

Raise fastapi/dotenv/multipart/httpx2 floors to current patched releases
and add an explicit starlette>=1.2.1 pin (FastAPI's own starlette>=0.46.0
floor still admits CVE'd versions).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…--out

Move the transcription-execution + output-shaping helpers out of the
transcribe command into a new core module aai_cli/transcribe_exec.py
(run_transcription, out_payload, render_transform_steps). This keeps the
command a thin option surface, drops it back under the 500-line gate, and
lets onboard import run_transcription from a core module instead of
reaching into the command's private _transcribe_audio (removing a
reportPrivateUsage suppression).

Also harden the merged-in --out path-traversal guard: replace the bare
`raise Exception` with a clean UsageError (exit 2, no traceback) and check
`".." in out.parts` (path components) instead of a substring match, so a
filename merely containing ".." isn't wrongly rejected.

Tests: split the --out cluster into tests/test_transcribe_out.py (keeps
test_transcribe.py under 500) and add a path-traversal rejection test.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

# Conflicts:
#	aai_cli/theme.py

…nternals

The onboard wizard reused command internals by reaching into private
functions with reportPrivateUsage suppressions. Promote exactly the
reused functions to public API and call those instead:

- doctor: check_python / check_ffmpeg / check_audio / render
- setup:  install_mcp / install_skill / install_cli_skill / render

(_check_api_key, _check_coding_agent and the install helpers onboard
doesn't use stay private — minimal public surface.) Removes all eight
reportPrivateUsage suppressions in onboard/sections.py; the wizard now
depends only on each command module's intended public surface. Test
patch targets updated to the public names.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>