Reduce per-session context footprint of flow + design plugins #156

Problem

Every Claude Code session that has flow + design enabled pays for a large plugin footprint in the
skill / agent listing before any work starts. Measured in AI-StudyMate (2026-10-14):

Proposed changes

1. Hide Codex-only source-command-* skills from Claude. They duplicate every slash command
   (user-invocable: false) and exist only for Codex, but Claude Code still lists all 55 to the
   model. Move them to plugins/*/codex-skills/ and point only .codex-plugin/plugin.json
   ("skills") there. Generator: cli/lib/harness/codex-runtime.mjs.
2. Cap agent descriptions (~200 chars) + a test that enforces it. Worst offenders:
   ux-research-agent 1274, design-system-keeper 1112, footage-analyst 715,
   media-generation-director 708, motion-critic 687, footage-director 620, draw-critic 604,
   signature-moment-critic 590. The description only decides when to spawn; detail belongs in
   the body.
3. Internal agents → one-line descriptions ("Internal — spawned only by X"): flow:builder,
   breaker, shipper, investigator, user-advocate (debate-protocol only),
   design:reconstruct-*, footage-*, _draw-*, _security-regex-catalog. Hide from the model's
   picker if the harness allows it.
4. Split video out of design into an opt-in design-video plugin: reel, video-analyze,
   footage-*, video-comp, to-lottie, footage-keyframes, media generation for video.
   Most repos never cut video.
5. Trim the flow plugin.json description (it lists every command/skill and is stale —
   says "29 slash commands", there are 31).

Expected effect

Roughly 5–8k tokens less per session start for anyone with both plugins enabled, no behaviour change
for Codex.