mirror of
https://github.com/rennf93/roboco.git
synced 2026-08-03 07:23:24 +02:00
e9d0e0bd48
* feat(video): Phase A — VideoEngine origination spine + held-source gates
New default-off engine skeleton: opens a UX/UI authoring task (source=video, assigned to a ux-dev, LOW complexity to clear the dev-needs-subtasks guard) and materializes a held CEO-approval draft (source=video_post). Excludes video_post from all three held-source skip sites; adds the video_draft marker, six config flags, and the feature-flag entries. Origination + gate behavior unit-tested.
* refactor(orchestrator): fold _dispatch_dev_work skip chain into a helper
The per-source if/continue chain grew past xenon's --max-absolute B when the video_post held source joined it. Extract _is_non_dev_dispatch_source (every held-CEO source plus the two Board exploration sources) so the dev loop's skip is one flat call. Behavior-identical.
* feat(video): Phase B — propose_video do-tool (metadata-only, team-gated)
UX/UI dev records a video's composition ref + per-platform captions onto the authoring task's video_draft marker. Team-gated at runtime via _caller_team (Role.DEVELOPER can't tell a ux-dev from a be-dev). Resolves the caller's ACTIVE task via get_active_task_for_agent, not an oldest-first scan that would clobber a second open video task. Metadata only, no render. Wired through do_server + route + schema; added to _DEV_DO.
* feat(video): Phase D — render loop + RemotionRenderer client
Orchestrator-async _video_render_loop renders a completed authoring task's merged composition to MP4 (vertical + square) via the remotion-renderer sidecar and materializes the held video_post draft. RemotionRenderer tars the read-clone's motion/ source, POSTs it, and saves the returned MP4 bytes to a TASK-scoped local path (no shared volume; a composition is reused across videos so a composition-scoped path would clobber an earlier draft). Render failures bounded-retry (read-clone catch-up window, transient sidecar) up to a cap, then terminal-fail. Client tested vs a mock transport; loop vs a mock renderer + real DB.
* feat(video): Phase C — release / spotlight / on-demand video triggers
Three entry points open a UX/UI video-authoring task via VideoEngine.open_video_task: (1) a published release drafts a companion video — best-effort in ReleaseProposalService.approve, never fails the publish; script from the CHANGELOG via the local model with a template fallback. (2) propose_feature_spotlight gains optional wants_video/video_script — best-effort, gated on video_on_spotlight, default-off leaves the spotlight flow byte-for-byte unchanged. (3) POST /video/request (CEO-only) for an on-demand brief, with clean disabled/not_opened responses. All gated on video_engine_enabled.
* fix(video): savepoint-isolate video-task inserts (F042 poisoned session)
The best-effort try/except around open_video_task (release-publish + spotlight hooks) swallowed the Python exception, but a DBAPI error at the insert flush left the shared session must-rollback — so the caller's next commit (release finalize / request boundary) threw PendingRollbackError: the release stuck 'pending' after actually publishing, or the spotlight draft + HTTP response were lost. Wrap both inserts (open_video_task, _originate_video_post) in a begin_nested savepoint (the repo's established F042 pattern) so a DB error rolls back only the insert. open_video_task returns None (every caller already handles it); _originate_video_post propagates to the render loop's handler. Regression test: an insert FK error returns None with the session left usable. Dormant while the flags were off; armed on the NAS.
* feat(video): Phase G — motion/ package + remotion-renderer sidecar + compose
In-repo Remotion v4 motion/ package (ReleaseAnnouncement composition; calculateMetadata returns 1080x1920 vertical / 1080x1080 square from inputProps.orientation) + a credential-free remotion-renderer sidecar: untar the POSTed motion/ source, bundle (LRU-cached per source sha), selectComposition + renderMedia h264, stream the MP4 bytes back — matching the RemotionRenderer client contract. docker/remotion.Dockerfile on Debian (Chrome apt deps, build-time Chrome pre-warm, ffmpeg bundled in @remotion/renderer). Wired into both compose files (roboco_default only, shm_size 1gb, /health check) + the release publish matrix. Verified via a real local render of both cuts; the Debian docker build is the CEO's to run.
* chore(video): D-hardening — video_post source_task_id + render-loop docstring
Add a source_task_id back-reference to the video_post held-draft marker (traceability from a draft to its authoring task; also makes the render loop's two-key idempotency check wireable later). Fix the render-loop test's stale docstring ('never retried' -> bounded-retry). Both from the Phase D critic's non-blocking follow-ups.
* feat(video): Phase E1 — VideoPostService + heartbeat mutex (approve->post)
CEO-approve->post service: heartbeat-renewed Redis mutex (fail-closed, grace=ttl-2*heartbeat), re-read-in-lock double-post guard, per-platform durable commits (asyncio.shield-ed, settle-before-rollback on lock-loss), all writes inside the lock (captions validated pre-lock, applied in-lock — no stale whole-column clobber), idempotent, per-platform retry-skip. Poster interfaces (X/TikTok, mocked here). Reject + list-held-drafts. Survived 3 adversarial rounds; residual = a crash in the poster->commit window (CEO-gated low-freq, documented).
* fix(video): G-hardening — renderer leaks + Share Tech Mono brand font
Sidecar: give bundle() an explicit outDir tracked + deleted on LRU eviction (was leaking ~19MB remotion-webpack-bundle-* per source); res.on('close') cleanup so an aborted/retried download no longer leaks its remotion-out-* MP4 dir. Fonts: vendor Share Tech Mono (roboco-website brand font) as the display face (self-hosted woff2, 400-weight, headline fontWeight 700->400 to avoid faux-bold) + self-hosted Inter body — no gstatic fetch at render time (lsof-verified). Extras: composition_id whitelist (400) + Multer error middleware (400/413).
* feat(video): Phase E2 — X v2 + TikTok posters, tiktok_credentials, routes
LiveXVideoPoster (X v2 chunked media upload: init/append/finalize/STATUS-poll -> tweet w/ media_ids, OAuth1 signer reused). LiveTikTokPoster (OAuth2 inbox: init -> chunked PUT with asymmetric final chunk -> status-fetch; 401 -> refresh_token grant, rotated token persisted). tiktok_credentials Fernet singleton + migration 062 (single head). Routes: CEO approve/reject + list held drafts + write-only tiktok creds, wiring real posters into VideoPostService. Residual: a lock-loss right after a token-refresh flush can discard the rotated token (same rare CEO-gated class as the documented post->commit window).
* feat(video): Phase F — panel video-post queue + TikTok creds card + flags
video-post-queue.tsx: <video> MP4 preview with 9:16/1:1 cut switch, per-platform editable captions (280/2200 counters, over-limit disables approve), approve/reject, Request-a-video dialog. tiktok-credentials-card.tsx (4 write-only OAuth2 fields). feature-flags-card inlines TikTokCredentialsForm under video_engine_enabled. Mounted in command-center. tsc/eslint clean, 273 panel tests green. NOTE: needs the GET /video/posts/{id}/media route + mp4_paths on VideoPostResponse (folded into H) for the preview source.
* feat(video): Phase H — media route + e2e smoke + NAS arming + docs
GET /video/posts/{id}/media?cut= (CEO-gated FileResponse of the rendered MP4; closes the panel preview gap) + mp4_paths on VideoPostResponse. e2e smoke tests/e2e_smoke/test_video_pipeline.py (full flow, sidecar+X/TikTok mocked; asserts dispatcher skips, render-loop materialize, propose_video team-gate, approve idempotency). NAS arming: docker-compose.yml/.yaml ROBOCO_VIDEO_ENGINE_ENABLED/ON_RELEASE/ON_SPOTLIGHT default-on (.yaml resynced to .yml); registry stays off. CLAUDE.md video-engine section + CHANGELOG. Fixed 2 pre-existing route-test pollution leaks. Full suite 11763 passed.
* fix(video): auth-carrying preview, media route confinement, VideoPost type drift
Three fixes along the video preview path:
1. panel video preview auth: the <video> element was pointed straight at
GET /video/posts/{id}/media, but a native <video src> GET carries none
of axios's X-Agent-ID/X-Agent-Role headers — so in the default
header-trust deployment the request 401s. Fetch the cut via
videoApi.getMediaBlob (axios, responseType: blob) and drive <video>
off a URL.createObjectURL result instead. The object URL is revoked
on cut-change (the previous cut's URL) and on unmount, so neither
cut switches nor row teardown leak blob URLs.
2. backend media route confinement: GET /video/posts/{id}/media now
resolves mp4_path and refuses it with 404 when it falls outside
settings.video_output_dir. Defense-in-depth against any future
writer of mp4_paths serving files from arbitrary disk locations.
3. panel VideoPost type/comment drift: added mp4_paths to the
VideoPost interface (the committed VideoPostResponse already
carries it), and corrected the stale comment on videoMediaUrl
that claimed no route served the rendered bytes — the route has
existed since the media endpoint landed; the comment now describes
why getMediaBlob exists instead of a direct <video src>.
* Persist rendered videos to data in physical storage.
* ++
* docs(video): 0.18.0 CHANGELOG entry + RAG + map reference for video engine
- Move the video engine bullet from [Unreleased] into [0.18.0] and note
the ROBOCO_VIDEO_OUTPUT_DIR bind-mount persistence.
- Add docs/rag/architecture/video-engine.md (mirrors x-engine.md shape:
enable/disable, three triggers, render loop + sidecar, CEO gate, media
route confinement, credentials).
- Reference the video render loop in docs/map/orchestrator.md's engine list.
* chore(video): re-bump to 0.19.0 + sync registry compose defaults
Version was wrongly bumped to 0.18.0; 0.18.0 is an already-released
section. Restore its 2026-07-04 date and move the video-engine CHANGELOG
bullet into a new [0.19.0] - 2026-07-05 section above it. Bump
pyproject.toml, roboco/__init__.py, roboco/config.py (app_version),
panel/package.json, and the motion/README inputProps example to 0.19.0.
docker-compose.registry.yml: add ROBOCO_VIDEO_ENGINE_ENABLED /
_VIDEO_ON_RELEASE / _VIDEO_ON_SPOTLIGHT defaulted false (NAS arms them
true), and comment out the video-renders bind mount with a short note
so the public registry image ships video off by default. Structural
sync with docker-compose.yml maintained.
* fix(video): rate-limit /render + reflow motion/README
CodeQL flagged js/missing-rate-limiting on the renderer /render route.
The sidecar is container-network-only with one trusted caller (the
orchestrator, which renders cuts serially), so this limiter is a
retry-storm ceiling (30/min, well above legit render rate), not the
primary control. Also reflows motion/README.md hard-wrapped prose that
failed the markdown quality gate.
* fix(build): finish pnpm 11 migration + regen verb tables
The panel Docker image build failed on `pnpm install --frozen-lockfile`:
node:22-alpine's corepack resolved to its bundled pnpm 11, but
panel/package.json pinned packageManager to pnpm@10.25.0, and pnpm 11
refuses to run against that pin. The Dockerfiles were already written for
pnpm 11 (comments, CI=true, strictDepBuilds); the package.json pin was the
stale outlier. Finish the migration instead of working around it:
- panel/package.json: packageManager pnpm@10.25.0 -> pnpm@11.10.0; drop the
`pnpm` field (pnpm 11 ignores it — build approval lives in
panel/pnpm-workspace.yaml's allowBuilds). Lockfile unchanged (pnpm 11
accepts it as-is); frozen-lockfile verified.
- remotion-renderer/package.json: pin packageManager pnpm@11.10.0 for
determinism (was relying on corepack's implicit default); engines.node
>=22.13 (pnpm 11 requirement).
- docker/panel.Dockerfile + docker/remotion.Dockerfile: `corepack prepare
pnpm@11.10.0 --activate` so the build uses the pinned version explicitly
instead of trusting corepack's bundled default (which a future
node:22-alpine could change).
- .github/workflows/panel-ci.yml: Node 20 -> 22 (pnpm 11 requires
Node >=22.13; Node 20 fails the engines check).
Also regenerate agents/prompts/_generated/{developer,head_marketing,verbs}.md
— the video engine added propose_video and extended propose_feature_spotlight
(wants_video, video_script) but the verb tables weren't refreshed, failing
the foundation-check quality gate.
* chore(build): approve esbuild build script in remotion pnpm-workspace.yaml
pnpm 11 generated this file with a placeholder ('set this to true or false')
during install; resolve it to true so local dev of the renderer doesn't
re-prompt. esbuild's postinstall only verifies the prebuilt platform binary
(@esbuild/<platform> is installed as an optional dep), so approving it is
safe and silences the ERR_PNPM_IGNORED_BUILDS warning.
* fix(build): copy pnpm-workspace.yaml into panel + remotion images
pnpm 11 hard-errors with [ERR_PNPM_IGNORED_BUILDS] (exit 1) when a
dependency ships a postinstall script that isn't approved in
allowBuilds. Both Dockerfiles copied only package.json + pnpm-lock.yaml,
so the build-approval map in pnpm-workspace.yaml never made it into the
image — the remotion image build died on esbuild@0.28.1's postinstall.
Copy pnpm-workspace.yaml alongside the manifests in both images. In
panel, this also drops the --config.strictDepBuilds=false workaround:
with sharp and unrs-resolver now approved, their postinstalls run and
install the platform-specific binaries (previously skipped, leaving
sharp without its @img/sharp-* binary at runtime).
Verified locally: remotion + panel `pnpm install --frozen-lockfile`
exit 0 with the workspace file present; both exit 1 without it.
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
4.8 KiB
4.8 KiB
Video Engine (Remotion)
What It Is
RoboCo can author and post short marketing videos for the company's X (Twitter) and TikTok accounts — a release announcement, a feature spotlight, or an on-demand CEO request — implemented in roboco/services/video_engine.py (VideoEngine), roboco/services/video_post_service.py (VideoPostService), and the motion/ composition package, rendered by a credential-free remotion-renderer sidecar. Every clip is HELD for an explicit, per-clip CEO approval; nothing is ever posted automatically. It mirrors the XEngine "detect → originate a CEO-gated artifact → hold" shape, extended with a render pass between authoring and the held draft.
Enable/Disable
| Variable | Default | Effect |
|---|---|---|
ROBOCO_VIDEO_ENGINE_ENABLED |
false |
Master switch. Off = no video-authoring task is ever opened and no render/post happens. Panel-toggleable (Settings → Feature Flags). |
ROBOCO_VIDEO_ON_RELEASE |
false |
Sub-switch: open an authoring task when a release publishes. Off even with the master switch on. |
ROBOCO_VIDEO_ON_SPOTLIGHT |
false |
Sub-switch: open an authoring task when the CEO approves a feature-spotlight draft that requests one. Off even with the master switch on. |
Even when enabled, distribution requires an explicit per-clip CEO approval — a second independent gate beyond the flag.
Three trigger sources
Release videos (event-driven). A release publish fires the video-on-release trigger — gated by ROBOCO_VIDEO_ON_RELEASE independently of the master switch.
Feature-spotlight videos (event-driven). The CEO's approval of a feature-spotlight draft that requests a video fires the video-on-spotlight trigger — gated by ROBOCO_VIDEO_ON_SPOTLIGHT.
On-demand CEO request (panel-triggered). POST /api/video/request (CEO-role-gated) opens an authoring task directly — the CEO's escape hatch independent of the two automatic triggers.
All three open a normal, assigned UX/UI authoring task (balanced across the two ux-devs) rather than a held draft — the dev builds a Remotion composition under motion/ and proposes its composition id + per-platform captions via the team-gated propose_video do-tool, then ships it through the standard commit/PR/QA/doc/review lifecycle.
Render loop and the sidecar
Once the authoring task completes, an orchestrator render loop (_video_render_loop, interval ROBOCO_VIDEO_RENDER_INTERVAL_SECONDS) tars the merged motion/ source and POSTs it to the remotion-renderer sidecar (ROBOCO_REMOTION_BASE_URL). The sidecar is credential-free and git-free — it reads only what it's POSTed, bundles the composition, renders both the 9:16 and 1:1 MP4 cuts (ROBOCO_VIDEO_RENDER_TIMEOUT_SECONDS per render), and streams the bytes back. The orchestrator writes the MP4s to ROBOCO_VIDEO_OUTPUT_DIR (bind-mounted in all three compose files so renders survive container recreation) — the sidecar never writes to disk directly.
Ownership and the CEO gate
The render pass materializes a held video_post draft — mirroring the X-post/release-proposal shape: team=main_pm, assigned_to=secretary-1, confirmed_by_human=False (HELD — skipped by every dispatcher, never delivered to an agent). The CEO previews (an axios-blob fetch that carries X-Agent-ID/X-Agent-Role), edits, approves, or rejects each draft in a new panel video queue.
| Endpoint | Effect |
|---|---|
GET /api/video/posts |
List every held video draft awaiting decision. |
POST /api/video/posts/{task_id}/approve |
Post the rendered clip to X (native video, v2 media upload) and/or TikTok (inbox upload). Idempotent — approving an already-posted draft is a no-op. |
POST /api/video/posts/{task_id}/reject |
Cancel the draft with a reason. Terminal. |
POST /api/video/request |
Open an on-demand authoring task (CEO-only). |
Approval runs under a Redis heartbeat-renewed lock so a double-click can't double-post; the task is marked COMPLETED under the same lock before it releases.
Credentials
TikTok's OAuth2 secrets are Fernet-encrypted (ROBOCO_ENCRYPTION_KEY) alongside the existing X credentials, entered in the panel only — never in .env or an agent-visible setting. Every unconfigured leg (renderer, X, TikTok) degrades to a graceful no-op rather than a crash — the same graceful-null pattern as the X NullXClient.
Media route confinement
The panel's media fetch for video previews is confined to a single media route that carries the agent-identity headers and streams bytes from ROBOCO_VIDEO_OUTPUT_DIR — it cannot traverse the filesystem.
Related
docs/rag/architecture/x-engine.md— the held-draft marketing engine this mirrors and extendsdocs/rag/architecture/config-reference.md— full env var tabledocs/rag/roles/ceo.md— the CEO approval queues