Voc Orchestrator

/voc-orchestrator — The Weekly Run

One command, six phases. The intelligence lives in the individual skills; this is a coordinator that runs them in order and fails loudly instead of quietly.

Arguments

Use --dry-run for your first three runs. Read the synthesis by hand and check whether the confirmed themes match what you already believe from talking to customers. If they do not, the problem is almost always upstream configuration, not the rule.

Phase 0 — Preflight

Before pulling anything, verify the setup and print the result as a table. Do not proceed silently past a broken connector.

Preflight
  config/thresholds.json        ok
  context/product-context.md    ok
  CALL_PROVIDER=gong            credentials present
  Apify MCP                     connected
  CRM MCP                       NOT CONNECTED — signals will be tier U
  chat_channels                 3 configured
  notes_source                  not configured — skipping
  VOC_OUTPUT_DIR=./outputs      local, not synced   ok

Check VOC_OUTPUT_DIR for the strings Dropbox, Google Drive, OneDrive, iCloud. If it is inside a synced folder, warn before writing anything. This pipeline writes customer conversations to disk.

If fewer than confirmation.min_independent_sources sources are configured, say plainly that nothing can reach high confidence this run, and suggest either connecting more sources or lowering the threshold deliberately in the config.

Phase 1 — Pull sources

Run in parallel where the connectors are independent:

A source that is not configured is skipped, not failed. A source that is configured and errors stops the run — a partial corpus produces a confident-looking synthesis built on a hole.

Phase 2 — Analyze calls

/voc-analyze-calls. Its report is a structured corpus and feeds synthesis directly rather than being profiled signal by signal.

Phase 3 — Profile

/voc-profiler over every unprofiled signal in outputs/voc/sources/. Parallelize with subagents above 20 signals, respecting cost_guards.max_parallel_subagents.

Deduplicate before profiling, not after. Cross-source duplicates are the one failure the confirmation rule cannot catch, because by the time it runs they look independent.

Phase 4 — Synthesize

/voc-synthesize. Clusters via subagents, then applies the rule via scripts/synthesize.js.

Phase 5 — Route

/voc-route. Skipped under --dry-run.

Phase 6 — Dispatch

/voc-dispatch. Skipped under --dry-run. Produces drafts only. Nothing is sent.

Final summary

VoC Pipeline — week of 2026-04-20 to 2026-04-26

  SOURCES          calls 41 · reviews skipped (monthly) · chat 23 · notes 8 · custom 0
                   duplicates removed 6
  PROFILED         72 signals    tier A 9 · B 41 · C 8 · U 14 (quarantined)
  CALLS ANALYZED   41 in → 12 internal · 8 customer · 21 prospect
  SYNTHESIS        34 clusters → 6 confirmed · 9 watchlist · 19 dropped
  ROUTED           testimonial 2 · case study 1 · objection 3 · product 4 · exec 1
  DISPATCHED       11 artifacts, 1 held

  Not available this run: CRM (not connected) — all weighting is ICP-only
  Runtime 14 min

  Start here: outputs/voc/synthesis/2026-04-26-synthesis.md

Always end with the single file a human should open first.

Cadence

Weekly is right for calls, chat and notes. Monthly for reviews — they move slowly and each run costs money. Quarterly, run a wide window to see what has been steadily true rather than what happened last week.

Daily runs are a mistake. The confirmation rule needs time to accumulate corroboration across channels, and a daily cadence produces a watchlist that never graduates.

When it goes wrong

The orchestrator is deliberately thin. If output looks wrong, run the phases by hand — the failure is almost always in one skill and obvious once isolated.