dispatch-hermes-content-run

name	dispatch-hermes-content-run
description	Trigger Hermes's deterministic 11-agent video-production pipeline for an editorially-approved run, and poll it to completion. Use when submit-editorial-review records a human PASS and the approved asset bundle is ready to actually produce.
metadata	{"agent":"chief-editor","source":"Project Atlas Requirements §3.5/10.2","layer":"executive","host":"railway"}

dispatch-hermes-content-run

Hand off an editorially-approved run to Hermes's production pipeline (Soul Reference -> Visual/Presenter/Narration/Music/Asset Sourcing/Thumbnail -> Captioning -> Video Assembly -> Render QC -> Video Publish), and wait for it to finish.

When to use

• submit-editorial-review recorded a human PASS (content_assets.status = APPROVED).
• The upstream judgment outputs (select-vibe, direct-voice, direct-music, direct-visuals, plan-assets, generate-thumbnails, generate-seo-metadata, all owned by content-factory) and the approved script are available.

Workflow

1. Assemble the ContentRunRequest payload:
   • script_text and graphics_cues — both from the approved script (the Script agent's output carries graphics_cues for lower-third identification graphics; pass them through).
   • target_format — "shorts" / "reels" / "long_form" per the topic's format decision. Do NOT omit this: it defaults to long_form, which renders the 16:9 20-minute composition.
   • asset_planner_output (from plan-assets), voice_assignment (from direct-voice).
   • music_director_output (from direct-music) — without it the video has NO music, no beat-synced cuts, and no sound effects.
   • visual_director_output (from direct-visuals) — without it every shot falls back to a default crossfade with no color grade.
   • thumbnail_concepts (from generate-thumbnails), personality_id/personality_name/reference_image_urls.
   • publish_metadata -- title/description/tags from generate-seo-metadata's output, privacy_status set directly (unrelated to SEO, defaults to "private" if omitted).
2. Refuse to dispatch if any upstream judgment output carries "simulated": true — that marks a mock produced by a misconfigured Hermes deployment, and dispatching it spends real generation budget filming fake data. Surface it for human follow-up instead.
3. Call the hermes-bridge MCP tool start_content_run(payload) — returns {run_id, status: "started"} immediately. A 422 response means the run's projected cost exceeded the RUN_BUDGET_USD cap — surface it for human review rather than trimming the payload yourself.
4. Poll get_content_run_status(run_id) (a few minutes apart; this is a multi-minute pipeline, do not poll tightly) until status is completed or failed.
5. On completed, record the result's publish_result.live_url against the run — and surface plan_warnings, remediation.notes, qc_result.vision_warnings, and cost_summary.total_usd to the editorial record; on failed, surface the failed stage for human follow-up.

Function signature (manifest contract)

def dispatch_hermes_content_run(run_id: str, content_run_payload: dict) -> dict:
    """Starts and awaits Hermes's content-run pipeline. Returns the final
    {"status": "completed"|"failed", ...} result."""

Backend dependency

None beyond the hermes-bridge MCP tool (openclaw/mcp/hermes-bridge/) and Hermes's own Temporal worker (hermes/temporal_worker.py) — already wired, no stub.

Model

Chief Editor model (gemini-direct/gemini-2.5-flash).