> For the complete documentation index, see [llms.txt](https://getsquish.gitbook.io/squish/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://getsquish.gitbook.io/squish/resources/changelog.md). # Changelog User-facing changes per published version of [`@getsquish/squish`](https://www.npmjs.com/package/@getsquish/squish), newest first. Dates are the npm publish dates. Release notes also live on [GitHub Releases](https://github.com/getsquish/squish/releases). ## 0.3.1 — 2026-07-18 Audio activity now preserves high-frequency energy when building its low-rate timeline. * Fixed a resampling error in 0.3.0 that could make continuous tones above 500 Hz appear nearly silent. Squish now builds a full-band activity envelope before reducing it to the compact 1 kHz representation used by sheets and JSON. * No contract or command changes: `audio.samples[]`, `clip_peak`, CLI, and local MCP retain the 0.3.0 shapes. Videos without audio and videos with silent tracks retain their distinct results. {% hint style="success" %} **Try 0.3.1 on a video you know:** ```bash npx -y @getsquish/squish@0.3.1 clip.mov --json ``` The JPEG contact sheet includes frames, absolute timecodes, and an audio-activity band. Ask an AI to find one specific moment, then run Squish again with that range—for example, `--start 12 --end 18`—to inspect it more closely. The CLI requires Node.js 20+ and ffmpeg; video processing stays on your machine. {% endhint %} ## 0.3.0 — 2026-07-18 Audio activity becomes a navigation channel on the local mouths. * Every CLI/local-MCP sheet now carries a thin **audio-activity band** above the visual grid, aligned to the same absolute source timeline. High-energy peaks are marked so an agent can choose a range to inspect even when adjacent frames look alike. * JSON adds an `audio` object without removing or renaming any v0 field: `present`, `normalization`, absolute `window`, and `samples[]` (`time`, `level`). Local MCP returns the same object alongside `timecodes[][]`. * Levels use `normalization: clip_peak` across the full source clip, including windowed calls. A quiet zoom therefore stays quiet relative to the overview instead of being independently normalized until every interval looks loud. * The activity envelope keeps short transients during overview downsampling. Videos with a silent track return zero levels; videos with no audio track still produce sheets and report `present: false`. * Scope boundary: activity answers *when audio energy changed*, not what was said or what made the sound. No transcription, diarization, audio-event classification, or emotion inference. The web app, hosted API, and remote MCP remain visual-only until separately released. * README and `video-navigation` skill now include a five-minute test protocol: bring a video whose answer you know, ask for one specific moment, navigate overview → zoom, and verify the cited time against the source. ## 0.2.3 — 2026-07-14 The discoverable release: agents can now find Squish on their own. * Listed in the official **MCP Registry** as [`io.github.getsquish/squish`](https://registry.modelcontextprotocol.io/v0/servers?search=io.github.getsquish/squish): `package.json` carries `mcpName` (the registry's npm ownership marker), `server.json` describes both ways in (this npm package over stdio + the hosted endpoint), and a tag-triggered OIDC workflow publishes the listing on every release. * **`video-navigation` agent skill** in the repo's `skills/` dir — `npx skills add getsquish/squish` teaches an agent when to reach for a contact sheet, the zoom loop, and the citation discipline. Listed on skills.sh. * stdio MCP truthfulness: `serverInfo` now reports the real package version (it had sat at a hardcoded `0.2.0` through two releases), and `squish_video` carries behavior annotations — writes sheets to your disk, nothing leaves the machine (closed world), reruns overwrite the same outputs (idempotent). * New docs: [For AI agents](/squish/getting-started/for-ai-agents.md) (pick-your-surface decision table + behavior rules) and [Support](/squish/resources/support.md). * Hosted endpoint (same engine, **not in this package**): Apps SDK scan metadata (`outputSchema` + `structuredContent`, attachment `fileParams`), a subject-keyed free lane, and a commerce-free quota voice for ChatGPT traffic — the ChatGPT app was submitted for review 2026-07-12. * Engine unchanged: the `squish-cli-v0` and `squish-mcp-v0` contracts are identical to 0.2.0. ## 0.2.2 — 2026-07-10 The lens README, and the remote mouth grew auth. * README rebuilt around the lens canon — *AI sees through lenses, not answers: Squish adjusts the lens; the model interprets* — with a new cover ("Random access to video") and the official-app path (add `https://api.getsquish.app/mcp` as a connector, zero install). * `squish_video` tool description tuned from field evidence: the one-call economy line ("one call replaces a whole download → ffmpeg → extract → montage pipeline"), density-vs-window guidance, and the `timecodes[][]` mirror note — agents pick the tool and drive the loop better with no prompting changes. * Hosted endpoint (same engine, **not in this package**): `api.getsquish.app/mcp` gained optional `Authorization: Bearer` API keys (same credits as the hosted API, deduct-before-extract with auto-refund) alongside a small anonymous free lane; quota exhaustion answers with a structured error an agent can relay. **The local CLI and MCP server in this package stay ungated forever.** * Engine unchanged: the `squish-cli-v0` and `squish-mcp-v0` contracts are identical to 0.2.0. ## 0.2.1 — 2026-07-08 First open-source release: the engine behind `@getsquish/squish` went public at [github.com/getsquish/squish](https://github.com/getsquish/squish) (Apache-2.0, curated mirror of the founding monorepo — history starts there). * README rebuilt around the primitive: video as an addressable representation, the navigation loop (overview → zoom → cite), and the meta-demo — a 76-second explainer video shown next to the same video as one contact sheet. * `package.json` `repository` now points at the public repo, so npm's "Repository" link works. * Added `SECURITY.md` (private vulnerability reports via GitHub security advisories) and npm / CI / license badges. * Engine unchanged: the `squish-cli-v0` and `squish-mcp-v0` contracts are identical to 0.2.0. ## 0.2.0 — 2026-07-07 Windowed sheets — the release that turns a sheet into a navigable address space. * `--start` / `--end` on the CLI and `start` / `end` on the `squish_video` MCP tool window a run to a range. Both accept seconds (`90`) or a timecode exactly as stamped on a sheet (`1:30`, `1:07.3`). * Timecodes stay **absolute to the source video at every zoom depth** — a windowed run never re-bases them, so you can drill repeatedly: overview → range → moment. * Adaptive timecode precision: short windows stamp sub-second timecodes (`1:07.3`) so adjacent cells stay distinguishable. * Addressable floor: a window narrower than 2 ms per cell is rejected with a teaching error (zoom out or lower the density) instead of producing indistinct cells. * A windowed `--json` run echoes the resolved bounds as `"window": { "start": …, "end": … }`. ## 0.1.1 — 2026-07-06 README rewrite for the npm package page. No engine or contract changes. ## 0.1.0 — 2026-07-06 First publish of `@getsquish/squish`: the CLI and the MCP server in one package. * CLI: `squish