anabologyco-maker/claudia-bot/mv-director

Director tier - delegate the heavy work, keep the gates

Read runner/README.md §1b and §2 and docs/95-delegation.md §2 (what to delegate) and §4 before the first job. You are the gatekeeper; the owner's agent does the work on their plan. A delegated job never spends money: spending stays with you, after the owner's yes.

Where the jobs run (once per project)

slop --json agents status. An engine signed in on the bot's computer (the box): run there ("On the box"): no copies, no approval cards, the owner can be away. None: skill setup-agents signs one in on the box; or, if the owner prefers their own computer, "On their computer". The engine for jobs: slop --json agents use ENGINE --for director (it prints which briefs follow it; ENGINE = claude, codex or muse). Once per box: slop --json agents test ENGINE. Claude workflow briefs (research-lyrics, storyboard-panel, chapter-motion, safety-scan) also need, as a background command (1-3 min), node /workspace/claudia/runner/deleg.mjs preflight claude --workflow with subagent_tools_ok true.

Which brief for which stage (docs/95-delegation.md §2)

research-lyrics (concept facts, lyric drafts) · storyboard-panel · prompt-book (one context writes the clip prompts; you still run G5 and the spend) · chapter-motion · review · safety-scan. Gates, spend and the final edit decisions never leave you.

On the box

1. node /workspace/claudia/runner/deleg.mjs new auto --project NAME --brief storyboard-panel --set OWNER_QUOTES='<their words, verbatim>'. Read job_dir, engine (not signed in here per slop --json agents status: do not start; do that stage in chat), engine_note and gates_needed (a gate flag needs the owner's yes and what it allows, first).
2. node /workspace/claudia/runner/deleg.mjs start JOB_DIR (JOB_DIR = its job_dir; the gate flags after it). It returns at once. Record: slop --json -p NAME runs add ID --kind deleg --what '<kind> job' (ID = the job folder's name).
3. A workflow job takes 1-3 hours: tell the owner and end with the profile's turn rule line (say "continue"). Then node /workspace/claudia/runner/deleg.mjs wait JOB_DIR 25, once per turn, never in a loop. At most 2 agent jobs run on the box at once; while they run, render frames with --workers 2.
4. Done: the results are in JOB_DIR/outbox/. Check and apply them (below).

On their computer (local exec; every command and copy is an approval card: tell them first; "Always allow" is their decision, never push it)

Placeholders (runner/README.md, the table at the top): HOME = their home folder, TK = HOME/claudia, PROJECTS = HOME/slop/projects, PROJ = PROJECTS/NAME. Absolute paths, forward slashes, in double quotes.

• Once: the engine signed in there (skill setup-agents); macOS needs bash 4.4+ (brew install bash). git clone https://github.com/anabologyco-maker/claudia-bot "TK", then in the background bash "TK/bin/setup.sh" --minimal (Windows: runner/README.md §1). Checks: node "TK/runner/deleg.mjs" doctor --render, node "TK/runner/deleg.mjs" selftest, node "TK/runner/deleg.mjs" preflight ENGINE.
• Sync (runner/README.md §2, part a): the "small" tar line for research, storyboard and prompt-book jobs, "render" for chapters and reviews; CopyFromBox each part to HOME; node "TK/runner/deleg.mjs" unpack "HOME/NAME.tgz" "PROJECTS". Never while a job of that project runs.
• Build it on the box with new as above (add --engine ENGINE for their engine), CopyFromBox its job.json to PROJ/jobs/ID.job.json, then run its user_run_command as ONE background local command and record it with runs add. 1-3 hours: the same "continue" line; then read its terminal output.
• Done: CopyToBox PROJ/jobs/ID/outbox.tgz to /workspace/slop/projects/NAME/jobs/ID/, then tar -xzf /workspace/slop/projects/NAME/jobs/ID/outbox.tgz -C /workspace/slop/projects/NAME/jobs/ID/.

Check, then apply (runner/README.md §2, part f; the worker's own PASS is only a hint)

• Storyboard: slop --json -p NAME board audit --shots /workspace/slop/projects/NAME/jobs/ID/outbox/files/board/shots.json, then copy outbox/files/board/* into the project's board/.
• Lyrics: slop --json -p NAME lyrics lint /workspace/slop/projects/NAME/jobs/ID/outbox/lyrics/lyrics.txt, then the owner.
• Prompt book (the clips exist: skill clips step 1): each entry of outbox/prompts/clips.json to clips/prompts/CLIP.txt, cd /workspace/slop/projects/NAME && slop --json -p NAME clip set CLIP --prompt-file clips/prompts/CLIP.txt, then slop --json -p NAME clip lock CLIP; then G5 with outbox/prompts/book.md.

Report in at most 6 lines: done or failed, what changed, flags for the owner.

Gates and stops

The same six gates and spend rules as skill mv-studio. An owner quality flag: stop the job (node /workspace/claudia/runner/deleg.mjs stop JOB_DIR, or node "TK/runner/deleg.mjs" stop "PROJ/jobs/ID" on their computer) and slop --json -p NAME spend --freeze '<their words>' before you reply.