{"full_name":"benjiyaya/Calliope","owner":"benjiyaya","name":"Calliope","shard":"1","url":"https://github.com/benjiyaya/Calliope","description":"Local-first Agentic Idea-to-video Studio Using ComfyUI + FastAPI + SvelteKit","homepage":"","topics":["ai-video","ai-video-editing","ai-video-editor","aivideogenerator","comfyui","comfyui-extension","comfyui-workflow","fastapi","sveltekit","sveltekit-example"],"tag_pages":["ai-video","comfyui","fastapi","sveltekit"],"language":"Python","license":"MIT","avatar":"https://avatars.githubusercontent.com/u/162032034?u=8ba4037c9a02c6e7d0c4e5c1158aad88d5760dc3&v=4","created_at":"2026-07-27T20:50:33Z","pushed_at":"2026-11-15T09:41:53Z","metrics":{"stars":302,"forks":72,"watchers":8,"issues":55,"prs":34,"releases":20,"issues_open":25,"issues_closed":30,"prs_open":0,"prs_closed":8,"prs_merged":26,"discussions":17},"growth":{"day":{"stars":1,"forks":1,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[299,299,300,301,302],"forks":[71,71,71,71,72],"watchers":[8,8,8,8,8],"issues_open":[26,26,26,25,25],"prs_open":[0,0,2,2,0]},"boards":{},"same_owner":[],"issues":[{"number":51,"title":"[Coming update] V.1.5 - Build Scene With 3D Blocks - Story-Script-Clips Enhancement","comments":5,"created_at":"2026-09-08T15:17:56Z"},{"number":9,"title":"[Feature Request] UX Improvements & Much Needed Features for Video/Asset Workflow","comments":3,"created_at":"2026-08-22T03:12:19Z"},{"number":3,"title":"Ideas to improve","comments":3,"created_at":"2026-08-19T19:38:03Z"},{"number":22,"title":"extract_json returns the first inner object when the envelope is truncated or malformed — retry never fires, surfaces as \"0 beats\"","comments":2,"created_at":"2026-08-24T05:23:46Z"},{"number":19,"title":"[Bug / Feedback] LLM Agent context desynchronization and permission errors","comments":2,"created_at":"2026-08-23T18:39:11Z"},{"number":11,"title":"Expand Playground into Asset Creation with Character Looks, Objects, and Video Guides","comments":2,"created_at":"2026-08-22T16:39:02Z"},{"number":8,"title":"Hardcoded 10-minute timeout in _poll_history causes long ComfyUI video generations to fail","comments":2,"created_at":"2026-08-20T09:59:20Z"},{"number":98,"title":"[Doc] Documentation points to an incompatible H3 Motion Context node","comments":1,"created_at":"2026-10-16T12:24:44Z"}],"discussions":[{"number":103,"title":"Update Calliope","comments":0,"category":"Ideas","created_at":"2026-10-28T16:27:35Z"},{"number":101,"title":"Custom node Calliope","comments":1,"category":"General","created_at":"2026-10-21T15:10:19Z"},{"number":82,"title":"Entrée Boolean","comments":0,"category":"Ideas","created_at":"2026-10-04T06:20:36Z"},{"number":77,"title":"Ajout d'autres langues !","comments":0,"category":"General","created_at":"2026-09-29T17:14:09Z"},{"number":70,"title":"Storyboard needed please!","comments":1,"category":"Ideas","created_at":"2026-09-25T21:01:04Z"},{"number":65,"title":"Telechargement de la dernière Version 1.5.6","comments":2,"category":"Q&A","created_at":"2026-09-22T16:45:47Z"},{"number":61,"title":"Amélioration \"Construction de Scene\"","comments":0,"category":"Ideas","created_at":"2026-09-14T16:46:24Z"},{"number":56,"title":"My journey with Calliope (Raw thoughts & experiments)","comments":2,"category":"Show and tell","created_at":"2026-09-11T19:16:43Z"}],"latest_release":{"name":"Calliope 1.7.0 — agent engine rewrite and gap fixes","tag":"v1.7.0","published_at":"2026-11-14T16:47:49Z","body":"Release **v1.7.0**, 2026-11-14. The plugin harness is gone; the agent loop is a grant-scoped studio runtime, the capabilities it dropped are restored, and plain chat converses again.\r\n\r\n## The agent loop has been rewritten\r\n\r\nThe old plugin harness (`harness/loop|orchestrator|policy|registry|tools`, `harness/plugins/*`) is gone, replaced by a grant-scoped studio runtime (`agent/runtime/` + `agent/capabilities/`):\r\n\r\n- Your message is matched into **scoped studio calls** (story → script → assets → render → canvas). Each call runs ONE sub-agent that only sees its grant's tools, and the run is verified against the database — a write counts only when the grant's tables actually changed, never because the model said \"Done.\"\r\n- A failed call retries up to twice, carrying the previous transcript plus a recovery note that names the tools that ran.\r\n- **Review cards are the only queue door.** The agent prepares prompts and references and stops; you press Generate. No prose approval games, no direct queueing.\r\n- Steering works mid-run: send a message while the agent works and it lands at the next step boundary. Stop cancels the LLM turn **and** the GPU jobs (ComfyUI gets an interrupt).\r\n- Every chat grant now shares one tool surface for scratch files (`list_files`/`read_file`/`write_file`/`delete_file` under `DATA_DIR/workspace/sessions/<id>`) and the opt-in shell (`run_command`, Settings → Agent) with path containment and an approval card.\r\n\r\n## The gaps the rewrite dropped are fixed\r\n\r\nThe first cut of the rewrite silently lost capabilities the old harness had. All restored:\r\n\r\n- **Live streaming** — assistant tokens and per-agent reasoning stream into the chat (`agent.token` / `agent.thinking`), and token usage lands on the final message when the endpoint reports it.\r\n- **Multimodal attachments** — images (gated + downscaled, 4K included), videos (up to 8 evenly-spaced frames with a timestamp map), and documents (.txt/.md/.docx) reach the model, not just the event log.\r\n- **Composer mentions** — `@workflow` / `@skill` tags are consumed as context again; the pickers work on canvas and Build Scene.\r\n- **JSON-action bridge** — local models that can't emit function-call JSON still operate the studio (fenced `{\"tool\": …, \"arguments\": …}` replies execute through the same audited path, capped 8/turn).\r\n- **Hardening + history budget** — the Settings → Agent hardening rules append to every sub-agent prompt; retry transcripts trim by a character budget so a long turn can't overflow a local model's window.\r\n- **Build Scene poses** — the real 16-pose library is back (it was stubbed to two empty poses).\r\n- **Question cards, memory, skills, contracts** — `question/asked`/`answered` events, the memory panel, runtime skill reading, and the contracts staleness gate all work against the new layout (877 → 901 tests green).\r\n\r\n## Plain chat now converses\r\n\r\nA message that names no studio verb used to get a canned \"No studio change was requested.\" without the model ever being called. It now schedules one conversational turn — story-aware when the session is linked, canvas-aware in sandbox — and streams a normal reply. Attachments-only messages work the same way; the model sees the media and answers.\r\n\r\nStatus questions route as reads, not GPU writes: \"is the render done?\" or \"check if the portrait is ready\" no longer schedule an enqueue. \"Wait for the video to finish\" waits, then reports the outcome in the agent's own words instead of \"The project has 1 job.\" A render/assets status turn is not declared \"ready\" while its jobs are still running — the agent is told to call `wait_for_jobs` or state the handoff.\r\n\r\n## The workflow run-path is back on AI Canvas\r\n\r\nWorkflow cards are visible on the board again, with the full manual run loop:\r\n\r\n1. **＋** (bottom-left) places a workflow node.\r\n2. Drag from any media card's asset handle (characters, locations, items, and finished artifacts) onto a workflow card — the role chooser offers a **reference slot** (prom","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}