Panda video scaffold - #470
Closed
Philipcyrus wants to merge 11 commits into
Closed
Conversation
Panda Mobile agentic video pipeline on top of pristine upstream: - pipeline_defs/panda-video.yaml: stages + 3 human-approval gates (script, storyboard stills, final video) - skills/pipelines/panda-video/storyboard-director.md: one-still-per-scene then gate - tools/video/montage_svc_compose.py: THE HINGE — compose delegates to montage-svc - config/panda-elements.json: panda brand/element wiring (placeholders) - styles/panda.yaml: look of generated imagery (overlays stay in montage-svc) Ported from OpenMontage-main (NOT in upstream — verified): - tools/video/higgsfield_mcp_video.py: Higgsfield MCP-bridge provider - skills/meta/higgsfield-mcp-bridge.md: agent-side MCP protocol SCAFFOLD: stubs carry TODOs; not yet runnable end to end. Prove the hinge first. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Consolidate to ONE deployable: vendor montage-svc's proven render pipeline (vendor/montage_svc + vendor/brand) and call it in-process instead of over HTTP. - tools/video/panda_render.py: CLEAN compose tool (ugc/no-brand profile), stages assets into a montage run and calls run_compose verbatim. Round-trip tested: 1080x1920 h264, clips + xfade + audio mix, branded=False. - pipeline_defs/panda-video.yaml: compose stage now uses panda_render; produces the clean master that hits GATE 3. Branding is explicitly NOT applied here. - Removed tools/video/montage_svc_compose.py (the HTTP hinge) + MONTAGE_SVC_URL. - vendor/data gitignored (render scratch). Branding stays a separate, on-demand, post-approval step (panda_brand — next). Per architecture decision: branding only when the user asks, on the approved master. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds vendor/brand/fonts/msyhbd.ttc = Noto Sans SC (SIL OFL), instanced to a static Bold (wght=700), aliased to the filename the brand profiles expect. Replaces the system dependency on proprietary Microsoft YaHei — the repo now renders Chinese captions on Linux/EC2 with no font install step. OFL.txt + README included. Verified: full SC glyph coverage (避坑套餐…), usWeightClass=700, CJK caption render through panda_render succeeds. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- brand/panda-mobile/mascot/Panda-Mobile-Smiling-Hero-Turnaround-v1.png (6336x2688) - brand/panda-mobile/customer/Customer-V3-Hero-Turnaround-v1.png (6336x2688, new customer/) - config/panda-elements.json: reference both as character-consistency master sheets (panda + customer), for seeding Higgsfield element ids. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The engine is not a service; this small FastAPI app is the networked surface Dify
talks to. It starts/resumes runs and surfaces the 3 approval gates, with local
file-based job storage (no S3/Postgres — Phase 5 deferred).
- app.py: POST /jobs, GET /jobs/{id}, POST /jobs/{id}/respond, GET artifacts, /health
- store.py: per-job local folder (state.json + artifacts/)
- runner.py: MockRunner (no LLM; fakes script/storyboard, REALLY renders a clean
master via folded panda_render) + ClaudeCodeRunner skeleton (EC2 path)
- test_dify_flow.py: in-process mock-Dify test — PASSES the full gate handshake
start -> script -> storyboard(+revision) -> real final.mp4 (branded=False) -> done
- Storyboard gate accepts user-supplied stills from Dify.
data/jobs/ gitignored. Swap MockRunner->claude + local storage->S3 on EC2.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per request, generated motion clips now get their own human-approval gate before
assembly — not just stills and the final cut.
- panda-video.yaml: assets stage human_approval_default=true (GATE 3); final is GATE 4
- runner.py: new _do_clips (one clip per approved still); revise supports per-shot
regeneration via {"decision":"revise","shots":[i,...]} — only those shots re-run;
refactored resume() so revise regenerates the current gate's artifact
- app.py: Respond gains `shots`; test covers the clips gate + a per-shot revision
Flow verified end to end: script -> storyboard -> clips (+revise shot) -> final -> done.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Exact Dify workflow wiring: auth header, the gate loop, per-gate request/response
bodies, artifact fields to display, the polling requirement for the real agent,
a curl smoke test, and the deploy snippet to obtain {BASE_URL}.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Makes the repo clone-and-run ready on the box behind dev.om.mvnoc.ai: - deploy/install.sh: system deps + venv + launcher deps + import/render smoke test - deploy/panda-launcher.service: systemd (uvicorn 127.0.0.1:8600, --root-path /panda) - deploy/nginx-panda.conf: reverse-proxy block (subpath or subdomain options) - deploy/requirements-launcher.txt: minimal launcher+render deps - deploy/README.md: end-to-end deploy steps No secrets; .env is created from .env.example on the box. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ange dev.om.mvnoc.ai already forwards to 8501; the launcher now listens there (montage-svc retired, since its render is folded into the engine). So: - systemd: uvicorn on 127.0.0.1:8501, no --root-path - README: stop montage-svc, run launcher on 8501, proxy unchanged - DIFY_INTEGRATION + curl: BASE_URL = https://dev.om.mvnoc.ai (root) - nginx: documents "no change needed"; subdomain kept as optional Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ClaudeCodeRunner now drives Claude Code headless against the engine and adapts OpenMontage's checkpoint system to the launcher gates: - start/resume run `claude -p`; checkpoint-based resume (each leg reads latest checkpoint) - _sync maps stage+status -> gate (script/scene_plan/assets/compose) and done/failed - _mirror_artifacts copies project files into the job store, grouped stills/clips/final - _approve_stage writes completed + human_approved to clear a gate - config via CLAUDE_BIN/CLAUDE_EXTRA_ARGS/CLAUDE_TIMEOUT_S/PANDA_PIPELINE_TYPE FIX: panda-video.yaml had stability: experimental (schema allows only production|beta) and a non-schema `notes:` field — the whole manifest silently failed to load, so the engine used the DEFAULT 9-stage pipeline and our gates were NOT enforced. Now loads as idea/script/scene_plan/assets/edit/compose with gates on script/scene_plan/assets/compose. test_claude_adapter.py verifies the adapter (mapping/mirroring/sync/approval) against the real lib/checkpoint. Mock flow still passes. .env.example + README document the runner. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.