tdmcp-feature-discovery — new-feature ideation orchestrator

Coordinate a fan-out of surveyors + one synthesizer to produce a single prioritized feature backlog for tdmcp, deduped and reconciled against the roadmap. This harness finds and ranks ideas; the tdmcp-pipeline harness builds the chosen ones. Keep the boundary crisp: discovery answers "what should we consider next?", pipeline answers "build this."

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

Stage	Mode	Why
Survey	sub-agent (fan-out, parallel)	the five surveyors are fully isolated — each owns one surface, no inter-comms needed; the textbook fan-out case (mirrors the pipeline's design/build stages)
Synthesize	sub-agent (×1)	a single reasoning-heavy consolidation pass over result files — no producer↔reviewer loop, so no team needed

No TeamCreate here — surveys are pure result-passing 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-surveyor (×up to 5)	custom	td-feature-survey	_workspace/discovery/01_survey_<surface>.md
td-synthesizer	custom	td-feature-synthesize	_workspace/discovery/FEATURE_BACKLOG.md

The five surfaces: controls (Layer 1/2 creation & performance), library (vault + recipes + .tox/component packaging + distribution), cli (CLI/DX), ai (prompts + local-LLM copilot), td-depth (Layer 3 + bridge + operator KB).

Workflow

Phase 0 — context check (follow-up support)

1. Check whether _workspace/discovery/ exists.
2. Decide the run mode:
   • No _workspace/discovery/ → fresh run. Go to Phase 1.
   • Exists + user asks to refresh/deepen/re-prioritize part → partial re-run. Re-invoke only the affected surveyor(s) and/or the synthesizer, passing prior artifact paths so they refine rather than rewrite.
   • Exists + a materially new ask (e.g. post-release, new competitor) → new run. Move the old dir to _workspace/discovery_<YYYYMMDD_HHMMSS>/, then Phase 1.

Phase 1 — prepare

1. Determine scope: by default survey all five surfaces. If the user scoped it ("just CLI ideas", "AI features only"), survey only those surface(s).
2. Determine the weighting profile for synthesis: default live-show; if the user signals otherwise ("favour quick wins", "what closes competitor gaps", "make the agent cheaper"), pass quick-win / parity / agent-dx to the synthesizer in Phase 3.
3. Create _workspace/discovery/.

Phase 2 — survey (sub-agent fan-out, parallel)

Spawn one td-surveyor per in-scope surface in a single message (subagent_type: "td-surveyor", model: "opus"). Each prompt states its surface assignment (controls / library / cli / ai / td-depth), the hard rule stay inside your surface; write _workspace/discovery/01_survey_<surface>.md incrementally, and asks for a short summary return (counts + headline ideas) so the leader's context stays lean.

Resilience (transient errors are common at this fan-out width). When the batch returns, verify by file, not by return value: check that each in-scope 01_survey_<surface>.md exists and looks complete (has a tally). For any surface whose agent returned a socket/API error or left a missing/truncated file, re-spawn just that one surveyor once before synthesis — the incremental writes mean a retry resumes cheaply. If a surface still fails after one retry, proceed without it and record the gap (the synthesizer notes coverage). Do not block the whole run on one surface.

Phase 3 — synthesize (sub-agent ×1)

Spawn one td-synthesizer (model: "opus"), telling it the weighting profile from Phase 1 (default live-show). It reads every 01_survey_*.md, plus docs/ROADMAP.md / CLAUDE.md / CHANGELOG.md / competitive memory, dedupes, reconciles against the roadmap, prioritizes under that profile (high-confidence first within each tier), and writes _workspace/discovery/FEATURE_BACKLOG.md. It returns a self-contained prose summary.

Single-surface shortcut: if only one surface was in scope, you may skip the synthesizer and relay that survey directly — but still run it through synthesis if the user wants priority/roadmap reconciliation.

Phase 4 — report + handoff

1. Relay the synthesizer's summary to the user (in their language): headline NEW ideas, the recommended shortlist, and any coverage gap. Point to _workspace/discovery/FEATURE_BACKLOG.md.
2. Offer the handoff: any chosen items can go straight to the tdmcp-pipeline build harness.
3. Offer the Phase-7 evolution loop (feedback → tune the surveyor/synthesizer skills). Preserve _workspace/discovery/ for audit.

Data flow

[leader] ──scope──▶ td-surveyor ×5 (parallel, one message)
   controls │ library │ cli │ ai │ td-depth
        └── each writes _workspace/discovery/01_survey_<surface>.md ──┐
            (leader retries any that errored, once)                   ▼
                                          td-synthesizer (reads 5 + roadmap, weighting profile)
                                                      │
                                       _workspace/discovery/FEATURE_BACKLOG.md
                                                      │
                                   summary relayed ──▶ user ──▶ (optional) tdmcp-pipeline

Error handling

Situation	Strategy
One surveyor returns thin/empty	Keep its (small) report; synthesizer notes the lean surface. Don't block the batch.
One surveyor returns a socket/API error or a missing/truncated file	Re-spawn just that one surveyor once (incremental writes make the retry cheap). Only if it fails again, synthesize from the surveys that landed and name the uncovered surface in the coverage line.
Surveyor cites an unconfirmed operator	Idea is kept with UNVERIFIED — probe live; the flag rides through into the backlog so the build pipeline validates first.
Two surveys overlap on the same idea	Synthesizer merges into one entry under the best-fit surface with a cross-ref; not an error.
User scoped to one surface	Spawn only that surveyor; skip or keep synthesis per whether they want prioritization.
_workspace/discovery/ already exists	Phase 0 decides partial-refresh vs. new-run; never silently overwrite a prior backlog.

Test scenarios

Normal: user asks "what new features could we add?" → Phase 1 sets scope = all five, profile = live-show → 5 td-surveyor fan out in one message, each writing its 01_survey_*.md incrementally → leader verifies all five files landed (re-spawns any that errored, once) → td-synthesizer merges, dedupes against ROADMAP Phase 13 / deferred-v0.6.0+, ranks into P0/P1/P2 (high-confidence first), writes FEATURE_BACKLOG.md → leader relays the executive summary + Top-N shortlist and offers to feed picks into tdmcp-pipeline.

Error / scoped: user asks "just give me CLI feature ideas" → scope = cli only → one td-surveyor runs → synthesizer (or direct relay) reconciles the CLI candidates against the roadmap and reports; the other three surfaces are explicitly out of scope, noted in the coverage line.