faelo-ship

Faelo ship — one-shot, autonomous

Same engine as faelo-orchestrate, but optimized for "make the whole thing now". You get ONE budget approval up front, then drive all phases start → finish without pausing at each step. Use only when the brief is clear and the user explicitly wants autonomy.

Pre-flight (do this before spending)

1. whoami — confirm key + credit balance is enough for a full ad (typical $5-30; budget ~$12-15).
2. Restate the plan in 3-4 lines: product, archetype, duration/aspect, est. cost, est. time (~15-25 min). Get ONE explicit "go". This is the only gate.
3. create_project with a budget_cap_usd so the engine hard-stops if cost runs over (the cap is your safety net for autonomy). Fresh client_request_id.

A note on "autonomous"

Claude Code still shows its own per-tool permission prompt for each mcp__faelo__* call (that's the OS-level spend gate, by design). So "autonomous" here means one creative-decision approval up front — not zero clicks. The user will still confirm tool calls unless they've pre-allowlisted the cheap ones in their settings. Don't promise a fully hands-off run; promise "I won't stop to ask you creative questions — just approve the tool prompts as they appear, and I'll pause for real only on the paid render and on errors."

Then run autonomously

Drive the full chain without further approval (creative-wise), narrating milestones:

ingest_* → suggest_archetype/draft_plan/commit_plan → draft_script/ approve_script → generate_ref_batch (auto-pins one ref per shot) → commit_storyboard → cascade render: render_shot for EACH shot_index → music_generate → composite_variant → export_master → qa_delivered_ad. (Do NOT use generate_clips_batch — its config translation is unbuilt and it silently renders nothing.)

Rules while autonomous:

• Still a fresh client_request_id per mutating call.
• If budget_cap_usd is hit, STOP and report — don't try to route around it.
• On a content-filter block (the video model), auto-regenerate_segment ONCE with a softened prompt; if it fails again, stop and ask.
• Surface each finished artifact (refs, then final variants) so the user can watch progress; Read image URLs so they render (mock mode → mock://, won't render).
• End with the final variant URLs + a one-line cost/time summary, and offer rate_delivered_ad to capture their verdict.

Operators running Faelo locally

If you're the operator with the Faelo engine checked out locally, there's a deterministic one-shot CLI (with a free dry-run cost estimate) — see the internal repo's docs. Remote plugin users don't have that; for them the autonomous MCP flow above IS the one-shot.