deAPI skills 2.0.0

Added

• Second skill, deapi-api, for building deAPI into applications (native API v2, webhooks,
  OpenAI SDKs on https://oai.deapi.ai/v1).
• Free price quote before every paid job, with a cost limit (--max-cost, DEAPI_MAX_COST,
  default 0.25 USD) and --quote to get the price without submitting anything.
• Stable exit codes, mapped from the API's code field (key, role, IP allow-list, balance,
  key spend limit, rate limits, suspended account, removed endpoint).
• status (with --wait and --download), account (key role and spend limit, balance,
  usage) and doctor commands; job:<request_id> as input to chain jobs.
• Model choice from the live catalog: standard models by default, partner models only on
  request; WhisperLargeV3Ct2 by default for transcription; parameters fitted to each model's
  limits; text-to-speech language and voice checked against the model.
• Commands audio-video, video-upscale, describe, music, boost and models; voice
  cloning and voice design in tts; speakers and --subtitles srt|vtt in transcribe;
  reference images, videos and audio in animate; --count for images.
• Input checks: file type from the first bytes, size limits, refusal of files in hidden or
  sensitive folders (also through links and for --text-file, --lyrics-file,
  --input-file and output paths); image quotes by size without uploading.
• Quotes that flags cannot lower: a file's measured length wins over --duration, links other
  than TikTok are priced from the link, and the prompt booster fee of --enhance is part of
  the cost check for images and videos. A limit above the default shows as max_cost_raised.
• The API key is never sent along an HTTP redirect (the script and the code examples).
• Downloads keep a media extension, stop above 1 GB and leave no partial file; with --count,
  results that finished before a failure are reported in completed_results.
• Reference docs for the script (commands.md, recipes.md, models.md, network.md) and
  for MCP clients (mcp.md); tested examples for the native API, webhooks and the OpenAI SDKs
  in deapi-api/assets.
• Repository checks: frontmatter rules, the Agent Skills reference validator, OpenAPI
  conformance, MCP tool names, doc examples parsed by the script's own parser, leak scan
  (files, commit messages, PR text, gitleaks); offline unit tests on Python 3.9 and 3.12,
  Linux and Windows; the deapi-api examples with Node.js, curl and pinned OpenAI SDKs; CI.
• transcribe --duration SEC --quote without a link or file: the price for a length alone,
  with --platform (file, youtube, x, twitch, kick, tiktok) priced the way
  deAPI prices that source. The skill tells the agent never to invent a link for a quote.
• Every paid command (not only --quote) prints max_cost and max_cost_source, the cost
  limit in force; the agent reports that value, not the documented default.
• doctor lists blocking issues in problems; ok is true when the script can run jobs,
  even with warnings such as an ADMIN key without a spend limit.
• tools/check_skill_safety.py: keeps out patterns that agent skill scanners block
  (a literal path to an agent's secrets file, os.environ[...], process.env[...],
  chat-role tags as placeholders).
• README: setting the key in Hermes Agent, and a deAPI chat model as the Hermes model.

Changed

• stt is now transcribe (old name kept as an alias), balance is now account (alias kept).
• Results are saved in --out-dir and named after the request ID; text results are saved to a
  file and previewed in the JSON output. Files are never overwritten.
• Configuration by flags only; DEAPI_BASE_URL and DEAPI_POLL_* are gone. --base-url
  accepts only https://*.deapi.ai.
• Paid requests are retried only on HTTP 429; a server error or dropped connection while
  submitting exits with code 9 instead of risking a double charge.
• An HTTP 403 without a JSON body (a proxy, firewall or CDN block) is exit code 1, not
  "account suspended".
• Long text for tts is refused with the model's limit (no automatic splitting); the docs say
  to quote every part and ask the user when the total is above the limit.
• metadata in the SKILL.md frontmatter uses block YAML (accepted by the reference validator).
• The deapi description: its first sentence names deAPI within the 57 characters that the
  Hermes skill index shows, everyday wording ("read text aloud, also in the user's own
  voice", "improve prompts"), "Use when ... unless they name another provider" and "Not for
  chat, summaries or other LLM text generation".
• OpenClaw: requires.anyBins: [python3, python, py] instead of bins: [python3], so the
  skill is available on Windows, where Python is usually python or py.
• The script reads only DEAPI_API_KEY and DEAPI_MAX_COST from the environment, one by
  one; the docs, the script and the deapi-api examples no longer spell out the Hermes
  secrets file path or use os.environ[...], which the Hermes skills hub scanner refused.
• mcp__deapi__{tool} instead of an angle-bracket placeholder, which some gateways reject
  as a chat-role tag.
• SKILL.md: one script call per command line (no ; echo $?, chains or pipes; the exit code
  is in the JSON) with an example Claude Code permission rule; --max-cost after approval
  equals the approved quote exactly; doctor once per session before the first paid job and
  after key, role, balance or spend-limit errors.

Removed

• deapi-audio skill (API v1). Use deapi tts and deapi transcribe.
• The --scale option of upscale (the API has no such field for images).

This repository is the former deapi-ai/claude-code-skills; the old Claude Code commands on the deprecated API v1 remain at the v1-final tag.