tdmcp-hype-scout — external trend ideation orchestrator

Coordinate a fan-out of trend scouts + one synthesizer to produce a single prioritized hype tool backlog for tdmcp, grounded in cited community evidence and vetted against the real codebase. This harness finds and ranks trend-driven tool ideas; the tdmcp-pipeline harness builds the chosen ones. Keep the boundary crisp: hype-scout answers "what's the TD community hyped about, and which of that could we easily turn into a tool?", pipeline answers "build this tool".

Boundary vs tdmcp-feature-discovery: that skill does internal gap analysis (what's missing in tdmcp vs roadmap). This skill does external hype analysis (what's trending in the TD community). They are complementary — both can run and feed into each other.

Execution mode: sub-agent fan-out → fan-in

Stage	Mode	Why
Scout (×5)	sub-agent (fan-out, parallel)	scouts are fully isolated — each owns one surface, no inter-comms needed; mirrors the proven tdmcp-feature-discovery shape
Synthesize	sub-agent (×1)	a single reasoning-heavy consolidation pass over result files — no producer↔reviewer loop, so no team needed

No TeamCreate here — scouts pass results via files, so sub-agents are the right tool over team overhead. All Agent calls use model: "opus".

Agent roster

Agent	Type	Skill	Output
td-trend-scout (×up to 5)	custom	td-trend-scout	_workspace/hype-scout/01_scout_<surface>.md
td-hype-synthesizer	custom	td-hype-synthesize	_workspace/hype-scout/HYPE_TOOL_BACKLOG.md

The five surfaces:

• community-showcase — TD forum, Instagram, Vimeo (finished work)
• tutorials — YouTube channels + courses (what's being taught now)
• generative-ai — StreamDiffusion/ComfyUI/realtime-ML bridges into TD
• hardware-interactive — LiDAR / depth / hand-tracking / sensors
• vfx-aesthetics — dominant visual languages of 2025-2026

Workflow

Phase 0 — context check (follow-up support)

1. Check whether _workspace/hype-scout/ exists.
2. Decide the run mode:
   • No _workspace/hype-scout/ → fresh run. Go to Phase 1.
   • Exists + user asks to refresh / rescout one surface / re-rank → partial re-run. Re-invoke only the affected scout(s) and/or the synthesizer, passing prior artifact paths so they refine rather than rewrite.
   • Exists + a materially new ask (e.g. "rescout 6 months later", new aesthetic moment) → new run. Move the old dir to _workspace/hype-scout_<YYYYMMDD_HHMMSS>/, then Phase 1.

Phase 1 — prepare

1. Determine scope: by default scout all five surfaces. If the user scoped it ("just AI integration trends", "only aesthetic trends", "só hardware"), scout only those surface(s).
2. Determine profile for the synthesizer ranking: default is Hype × Build-Ease. The user can pick hype-only, quick-wins, strategic, or conservative (see td-hype-synthesize skill).
3. Create _workspace/hype-scout/ (or reuse, per Phase 0).
4. Briefly tell the user: "Scouting surfaces in parallel, then synthesizing. Expected output: _workspace/hype-scout/HYPE_TOOL_BACKLOG.md."

Phase 2 — fan-out scouts (parallel sub-agents)

Spawn one td-trend-scout per in-scope surface, in parallel (a single message with multiple Agent tool calls). Each call:

subagent_type: td-trend-scout
model: opus
description: "Scout <surface> for TD hype trends"
prompt: |
  You are the td-trend-scout for surface = "<surface>".

  Load your skill (`td-trend-scout`) and follow it. Produce
  `_workspace/hype-scout/01_scout_<surface>.md` with cited candidates
  ranked per the entry format.

  If `_workspace/hype-scout/01_scout_<surface>.md` already exists, read
  it first and apply only the requested change: "<change-request-or-fresh>".

  Cite real URLs from 2025-2026 where possible. ≥2 citations per entry.
  Write incrementally so a partial run is still useful.

While scouts run, do not work on the synthesis — wait for all to return.

Failure policy (per scout): if a scout returns an error, retry it once with the same prompt. If it fails again, mark that surface as SCOUT MISSING and proceed to synthesis without it. Do not block the whole run on one surface.

Phase 3 — synthesize (single sub-agent)

After all scouts return (or are quarantined), spawn one td-hype-synthesizer:

subagent_type: td-hype-synthesizer
model: opus
description: "Synthesize hype scouts into HYPE_TOOL_BACKLOG.md"
prompt: |
  You are the td-hype-synthesizer.

  Load your skill (`td-hype-synthesize`) and follow it. Read every
  `_workspace/hype-scout/01_scout_*.md`, dedupe across surfaces, vet
  feasibility against the real codebase, rank under profile = "<profile>",
  and produce `_workspace/hype-scout/HYPE_TOOL_BACKLOG.md`.

  If a scout file is missing, note `SCOUT MISSING: <surface>` at the top
  of the output and proceed.

  If `HYPE_TOOL_BACKLOG.md` already exists, read it first and apply only
  the requested change: "<change-request-or-fresh>".

Phase 4 — user-facing summary

Once the synthesizer returns, present a concise summary to the user:

1. One-line headline: "Top trend = X, top buildable = Y."
2. The Top 5 "Ready for tdmcp-pipeline" list (just names + 1-line value each).
3. Any Force multipliers identified.
4. Pointer to the full file: _workspace/hype-scout/HYPE_TOOL_BACKLOG.md.
5. Next-step prompt: "To build one of these, run tdmcp-pipeline with the chosen tool. To re-rank, ask for a different profile (hype-only / quick-wins / strategic / conservative). To rescout one surface, name it."

Keep it short — the file holds the depth.

Data flow

• File-based between scouts and synthesizer (each scout writes its own file; synthesizer reads them all).
• Return-value-based from synthesizer to orchestrator (file path + 1-line headline).
• All artifacts live under _workspace/hype-scout/. Never under the project root.

Error handling

Failure	Strategy
Scout sub-agent errors out	retry once with same prompt; on second failure, mark SCOUT MISSING: <surface> and proceed
Scout returns but 01_scout_<surface>.md is missing	retry once; on second failure, treat as SCOUT MISSING
Scout tags file PARTIAL-DUE-TO-NETWORK	accept; synthesizer will propagate the flag
Synthesizer errors out	retry once; on second failure, ask the user (rare — synthesis is offline file work)
Conflict between scouts	synthesizer averages + annotates per skill — do not intervene

Follow-up handling (re-invocation)

Common follow-ups and how to route them:

• "refresh the backlog" → re-run all scouts (Phase 2) + synthesizer (Phase 3) with <change-request-or-fresh> = "refresh citations and add any new 2026 trends; preserve prior numbering where possible".
• "rescout only " → re-run that one scout + the synthesizer (skip the other scouts).
• "re-rank by quick-wins" → skip scouts, re-run synthesizer with <profile> = quick-wins.
• "deepen trend X" → re-run the scout for trend X's surface with <change-request-or-fresh> = "deepen entry X with more citations and a sharper tool sketch"; then re-run synthesizer.
• "build #N" → do not build here. Hand off to tdmcp-pipeline with the candidate id and detail block as input.

What this skill does NOT do

• It does not build tools — tdmcp-pipeline does.
• It does not do internal gap analysis — tdmcp-feature-discovery does.
• It does not drive a multi-release campaign — tdmcp-backlog-campaign does.
• It does not spawn the build agents (td-architect, td-builder, etc.). Only spawns td-trend-scout and td-hype-synthesizer.

Test scenario (normal flow)

User: "ideias hype — o que está bombando no TouchDesigner agora? quero ferramentas fáceis de criar com o tdmcp"

1. Phase 0: no _workspace/hype-scout/ yet — fresh run.
2. Phase 1: scope = all 5 surfaces, profile = default.
3. Phase 2: spawn 5 td-trend-scout sub-agents in parallel (one Agent call per surface, single message).
4. Phase 3: spawn 1 td-hype-synthesizer with the 5 scout outputs.
5. Phase 4: present Top-5 + Force multipliers + pointer to backlog file.

Test scenario (error flow — one scout fails)

1. 4 scouts return cleanly; generative-ai scout errors out.
2. Retry generative-ai scout once → still fails (e.g. WebFetch rate-limited).
3. Note SCOUT MISSING: generative-ai and proceed to Phase 3.
4. Synthesizer notes the gap at the top of HYPE_TOOL_BACKLOG.md.
5. Phase 4 summary mentions the missing surface and suggests "rescout generative-ai later".