Changelog

All notable changes to the transcribe-so CLI (0.x) and to the plugin manifests (1.x, tracking the MCP server version).

Bump the version in every plugin manifest with each release; CI checks it.

Plugin 1.8.2

Manifests only; no skill or tool changes.

Added

• The root plugin.json now declares the Agent Plugins 1.0.0 schema ($schema, author, homepage, repository, license, keywords), and a root mcp.json lists the remote server as streamable-http. Kiro Powers and the Cursor Marketplace read this format; Antigravity still validates it and loads it as a native plugin unless its Agent Plugins flag is on.

Changed

• .cursor-plugin/plugin.json drops the non-standard note key from its MCP server entry.

Plugin 1.8.1

README only: a How-to videos section (Claude Code, Claude, the Claude desktop app, ChatGPT, WhatsApp) and the playlist link. No manifest, skill or tool changes.

Plugin 1.8.0

Plugin manifests and skills only; the CLI stays at 0.2.0. Tracks MCP server 1.8.0.

Changed

• All manifests move to 1.8.0 with the MCP server. Server 1.8.0 adds the clip options from transcribe.so#1583 where the server enables them (output: audiogram, video or audio; caption_position, words_per_line, framing, highlight_color; rights_affirmed for video from a link, when the source video can be used; up to 180 s) and accepts public TikTok, X and Facebook video links as platform_url.
• skills/transcribe-audio: the platform list names TikTok, X and Facebook.
• Antigravity CLI plugin: root plugin.json (name, version, description) and mcp_config.json (serverUrl to the hosted MCP server); install with agy plugin install <repo URL>. The CLI cannot run the OAuth sign-in, so the README adds the server with an API key header under its own name. The build's version and listing checks cover the new manifest.
• Codex longDescription (the OpenAI listing text): drops the claim that recordings "stay private"; the listing only says to sign in with an existing account.
• skills/get-transcript, Clips: output and the caption options, the 180 s qualifier, and getClip's file_url (mp4_url is null for audio).
• README: a "What the Claude plugin runs and sends" section for directory review; the ChatGPT section points at the developer-mode connector only.
• LICENSE: copyright holder is SUNMOON.CO PTE. LTD. (transcribe.so).

Plugin 1.7.0

Plugin manifests and skills only; the CLI stays at 0.2.0. Tracks MCP server 1.7.0.

Changed

• skills/get-transcript, Clips: call the new read tool getClipQuote for the range first, tell the user its charge_usd and wait for a go-ahead (unless it is within a limit the user already gave), then pass the agreed figure to renderClip as max_charge_usd. On max_charge_exceeded nothing is held or rendered: report the figure and ask, never raise the ceiling.
• Codex commerce_description now covers clips (getClipQuote preview and the renderClip ceiling).
• README: the MCP tool count is corrected to 24 (it said 21).
• idempotency_key retry rule for transcribe and renderClip in both skills, aligned with the server: keep the same key by default (a retry with the same arguments replays the first result; queue_full and server errors after retry_after; not_ready while an earlier call may still be running). Refusals are replayed under the key for 24 hours, so a new key is needed only after a refusal actually received that says nothing was started or drawn, once its cause is fixed; the listed codes are examples, not a closed list. If the same key is rejected because the arguments differ and the first call was never answered, it may have gone through: re-send the same key with the original arguments or check the job or clip first.
• Retry bounds now match the server (transcribe.so #1541): the preview phase ends within 45 seconds (retry once, then tell the user), and a same-key not_ready carries stale: true after 120 seconds with no result recorded. transcribe: on stale, a job with your client_reference in any status counts as found; with no client_reference, tell the user instead of switching keys. renderClip: on stale, tell the user. The earlier "keep retrying for up to 5 minutes" rule is gone.
• Every plugin manifest is 1.7.0; the Codex release notes name MCP server 1.7.0 and the clip preview.
• scripts/check-api-coverage.ts: getClipQuote added to the skip list with the other clip operations (no CLI clip commands yet).

Plugin 1.6.1

Plugin manifests only (.codex-plugin, .claude-plugin, .cursor-plugin, .grok-plugin, gemini-extension.json); the CLI stays at 0.2.0.

Changed

• Directory-safe copy in skills/transcribe-audio, skills/get-transcript and commands/transcribe.md: the model-readable text describes what getQuote reports and that the agent waits for the user's go-ahead; it carries no money, promotion or account-tier wording and no em-dashes. The skills are based on the copy OpenAI's directory scan accepted, with factual corrections: getTranscript inline truncation (truncated, download_url), idempotency_key on transcribe and renderClip, artifact_missing as a not_ready reason, and the approval rule (go ahead when retail_usd is 0 or within a limit the user already gave, otherwise ask first).
• Plugin versions now track the MCP server version (1.6.1) in every manifest instead of a separate 1.x line.

Added

• .codex-plugin/plugin.json: Codex listing block (interface) and review metadata (extensions.com.openai.review test cases, commerce declaration, extensions.com.openai.publication.release_notes). The demo recording URL is injected at build time, never committed.
• scripts/build-openai-plugin.mjs (pnpm run build:openai-plugin): builds the OpenAI directory submission ZIP into out/ and fails on listing-tier copy violations. See README, "Directories".

0.2.0

Targets transcribe.so API 1.1.0.

Added

• transcript <id>: the COMPLETE transcript as a raw body, never capped (--format txt|md, --no-speaker-labels, --no-timestamps, --out FILE). A 409 not_ready is acted on by its reason: only transcription_processing is retried (honouring Retry-After, bounded by --wait-seconds, default 0 = no waiting). Every other reason exits non-zero immediately with the recovery on stderr, because retrying it can never succeed.
• captions <id> --for <destination>: paste-ready captions and chapter timestamps. Destinations: instagram, x, threads, linkedin, youtube, spotify, apple_podcasts, markdown, plain. Variants: standard, highlights, clips, quoted_sections, show_notes, original. --cta appends the transcribe.so footer (off by default: the caption belongs to the user). Prints the text raw; --json prints the full envelope (warnings, ok_to_paste, thread[], constraints). On artifact_missing it prints the exact recovery, and --regenerate runs that one POST /timestamps/regenerate for you and retries. --regenerate is not side-effect free: it regenerates and REPLACES all four cached caption/chapter variants and burns 1 of the 10 regenerations allowed per transcription. Once those are gone the API answers 429 rate_limited with no Retry-After, which the CLI maps to exit 1 (not the transient exit 5), because only a re-transcribe resets the count.
• search <q> [--id <id>] [--limit N] [--offset N]: "who said X, and when" across the library, or scoped to one transcription. Hits carry speaker, start_ms, end_ms and a deep-link url.
• capabilities: one JSON object: live /me and /pipelines, the API version from the spec, every format enum this CLI accepts, and the exit-code table. No guessing flag values.
• --max-charge-usd <n> on create, run and retry: a server-side charge ceiling (max_charge_usd) enforced at the wallet hold. run sends its --max-usd as the ceiling by default; --no-server-ceiling opts out. A retry is a new paid attempt with its own hold, so the ceiling has to be sent again there.
• Exit code 7 for 402 max_charge_exceeded, distinct from 4 (the account cannot pay) and from 6 (the local --max-usd refusal). Nothing was charged and no job was started; the remedy is to report the real price, not to top up. Exit codes 0-6 are unchanged.
• --segments-offset / --segments-limit on result and wait, with segments_meta surfaced on stderr. A windowed transcript now says so instead of looking complete.
• Examples: examples/ig-caption.sh, examples/who-said-what.sh.
• pnpm run test: unit tests for the mapping/parsing helpers.
• Multi-runtime manifests so the same repo installs outside Claude Code: .cursor-plugin/plugin.json (Cursor plugin: both skills plus the hosted MCP server) and gemini-extension.json (Gemini CLI extension bundling https://transcribe.so/mcp). The README Install section now leads with per-runtime instructions for Claude Code, Codex, Cursor, Gemini CLI, ChatGPT, and any skills.sh agent.

Changed

• stdout is JSON everywhere except the three raw-output commands: subtitles, transcript, and captions (without --json).
• not_ready hints now name the specific reason and its recovery instead of always suggesting wait.
• API coverage gate: transcript, search (both routes), timestamps and timestamps/regenerate moved from the skip list to covered.

0.1.0

Initial release: auth:status, me, pipelines, quote, create, wait, result, run, list, get, delete, retry, upload, subtitles, ask.