Voice-Platform pilot 0 of 9 decided

Shared Voice Platform with a Voice-Enabled Task Orchestrator

Plan: /Users/agency/Documents/Agty/Voice-Platform/action-plans/2026-08-09-voice-platform-pilot.md

Review the nine proposed tasks. The first pilot keeps Sonnet and ElevenLabs as the default, uses Grok only as a startup fallback, and leaves every other voice agent unchanged until you accept the result.

How to review this

T-001 — Establish the private Voice-Platform source of truth

Undecided
Task

Create the private Voice-Platform repository as the authoritative home for shared voice runtime code, contracts, operations, transcripts, and reusable Codex and Claude context. Depends on: none.

Problem solved

Voice solutions are scattered across skills, historical tasks, Modal code, and project repositories.

→
Cost of inaction

Each agent spends tokens rediscovering decisions and can regress already-solved reliability failures.

→
Benefit

One factual source shows current ownership, proven patterns, and operating instructions.

Done and 80/20

Private local and GitHub repositories match; architecture, provenance, security, model routing, transcript policy, operations, and anti-slop rules are documented; secrets are absent. The 80/20 is the private repository, README, AGENTS.md, current voice map, verified source pointers, and operations map.

flowchart LR
  A[Voice knowledge scattered] --> B[Create private shared source]
  B --> C[One reliable platform map]
flowchart TD
  S1[Recover verified sources] --> S2[Create private repository]
  S2 --> S3[Add concise docs and tests]
  S3 --> D{Private and verified}
  D -->|No| S1
  D -->|Yes| DONE[Source ready]
Your decision

T-002 — Build the Sonnet-first shared voice gateway

Undecided
Task

Build the shared Modal speech gateway with Sonnet and ElevenLabs as default and native Grok Orion as a transparent fallback only when Sonnet cannot start the conversation. Depends on: T-001.

Problem solved

Provider configuration and reliability behavior are embedded in one application instead of reusable infrastructure.

→
Cost of inaction

The new skill either duplicates the voice bridge or depends on unrelated Telegram-coach internals.

→
Benefit

New agents use one tested gateway with shared VAD, reconnect, and provider policy.

Done and 80/20

The isolated pilot starts with Sonnet, falls back only before the conversation, pins one engine, handles real-word interruption and noise rejection, resumes safely, and retains explicit rollback. The 80/20 is one route with deterministic provider tests and real audio evidence.

flowchart LR
  A[Provider logic embedded] --> B[Build Sonnet first gateway]
  B --> C[Reusable voice route]
flowchart TD
  S1[Extract verified contracts] --> S2[Implement provider adapters]
  S2 --> S3[Deploy isolated pilot]
  S3 --> D{Audio and fallback verified}
  D -->|No| S1
  D -->|Yes| DONE[Gateway ready]
Your decision

T-003 — Build the authenticated SSH Codex runner

Undecided
Task

Create a secure job bridge through which Modal voice sessions invoke Codex CLI, inspect permitted SSH files, run tools, and create bounded sub-agents without a second voice confirmation. Depends on: T-001.

Problem solved

The Modal voice runtime cannot directly use the repositories, Codex CLI, and durable agent state on the SSH box.

→
Cost of inaction

Voice remains an explanation layer and still requires manual context transfer.

→
Benefit

Voice conversations can launch grounded, durable Codex work against real files.

Done and 80/20

Signed replay-safe jobs reach an authenticated worker, produce durable receipts, redact secrets, and preserve non-delegable gates. The 80/20 is one real read job and one bounded Codex sub-agent job with verified receipts.

flowchart LR
  A[Voice cannot reach SSH work] --> B[Add signed Codex job bridge]
  B --> C[Voice can launch grounded work]
flowchart TD
  S1[Define signed job contract] --> S2[Build authenticated worker]
  S2 --> S3[Test Codex and subagents]
  S3 --> D{Receipts complete}
  D -->|No| S1
  D -->|Yes| DONE[Runner ready]
Your decision

T-004 — Create the Git-backed transcript and session ledger

Undecided
Task

Implement one live transcript ledger in the private repository with partial checkpoints, terminal records, conversation resumption, privacy metadata, and durable Git history. Depends on: T-001 and T-002.

Problem solved

Transcripts and session state currently use different queues, volumes, and import paths.

→
Cost of inaction

Disconnects can lose feedback and every new integration invents another transcript format.

→
Benefit

Every session has one searchable record that restores context and feeds future Codex work.

Done and 80/20

Partial turns persist safely, final records are terminal and deduplicated, reconnect restores full context, and Git writes are serialized and secret-scanned. The 80/20 is one forced-drop pilot that resumes without duplicates and produces one terminal private Git record.

flowchart LR
  A[Transcripts scattered] --> B[Write live private Git ledger]
  B --> C[Recoverable shared history]
flowchart TD
  S1[Define versioned schema] --> S2[Write checkpoints and commits]
  S2 --> S3[Force drop and reconnect]
  S3 --> D{One terminal record}
  D -->|No| S1
  D -->|Yes| DONE[Ledger ready]
Your decision

T-005 — Create the voice-enabled task orchestrator skill

Undecided
Task

Create extract-tasks-orchestrator-w-voiceagent as a thin wrapper around the existing task orchestration contract and put a working voice block at the top of its review HTML. Depends on: T-001 through T-004.

Problem solved

The current visual review cannot discuss tasks, answer grounded questions, or accept a brain dump.

→
Cost of inaction

Refinement remains a manual transfer loop and the shared platform has no proving workflow.

→
Benefit

One review page lets Matt understand the plan, talk through it, and preserve feedback.

Done and 80/20

The new skill retains the existing Markdown, decision, comment, and boss contracts while adding an authenticated voice block. Existing skills remain unchanged and Codex and Claude mirrors match. The 80/20 is one published page with grounded voice and live transcript saving.

flowchart LR
  A[Visual review only] --> B[Add voice review skill]
  B --> C[Discuss plans on page]
flowchart TD
  S1[Wrap existing skill] --> S2[Add first page voice block]
  S2 --> S3[Test parity and interface]
  S3 --> D{Review works}
  D -->|No| S1
  D -->|Yes| DONE[Skill ready]
Your decision

T-006 — Ground the review agent in mission context and tools

Undecided
Task

Build the capsule and tool policy for full mission-critical context up to 18K tokens, complete review HTML, selected Notion pages, changelog evidence, repository graph queries, and SSH Codex jobs. Depends on: T-003 and T-005.

Problem solved

A generic voice assistant cannot explain why tasks exist or answer repository questions accurately.

→
Cost of inaction

The agent gives shallow answers and sends under-grounded requests to Codex.

→
Benefit

Matt can discuss the plan with an agent grounded in the meeting, project, codebase, and review state.

Done and 80/20

Context selection is deterministic and provenance-bound, tools are scoped, receipts are durable, and stale evidence is stated. The 80/20 is a fixture that answers one plan, one Notion, and one graph question with source evidence.

flowchart LR
  A[Generic agent context] --> B[Bind mission sources and tools]
  B --> C[Grounded review answers]
flowchart TD
  S1[Prioritize source context] --> S2[Bind sources and tools]
  S2 --> S3[Run evidence questions]
  S3 --> D{Answers grounded}
  D -->|No| S1
  D -->|Yes| DONE[Capsule ready]
Your decision

T-007 — Implement the brain-dump refinement round trip

Undecided
Task

Turn a completed voice brain dump into a refinement packet that the originating Codex session uses to revise the Markdown plan, regenerate the HTML, and republish the same review. Depends on: T-004 through T-006.

Problem solved

A saved transcript alone does not update the plan or reconnect to the originating agent.

→
Cost of inaction

Matt must restate feedback in chat and the voice conversation remains passive documentation.

→
Benefit

One conversation produces visible, transcript-backed improvements while authorization stays explicit.

Done and 80/20

The origin task imports once, separates decisions from suggestions, produces a plan diff, regenerates both charts, republishes the same slug, and preserves receipts. The 80/20 is one brain dump that visibly changes at least one proposed task.

flowchart LR
  A[Brain dump disconnected] --> B[Import and refine plan]
  B --> C[Updated review page]
flowchart TD
  S1[Import terminal transcript] --> S2[Classify voice feedback]
  S2 --> S3[Regenerate plan and page]
  S3 --> D{Receipts preserved}
  D -->|No| S1
  D -->|Yes| DONE[Round trip ready]
Your decision

T-008 — Prove the complete pilot on a real action plan

Undecided
Task

Run the acceptance session where Matt reviews a real plan by voice, gives a brain dump, resumes after a disconnect, launches a Codex sub-agent, and receives a correctly regenerated page. Depends on: T-002 through T-007.

Problem solved

Component tests cannot prove that the entire workflow is usable in real work.

→
Cost of inaction

Microphone, permissions, reconnect, transcript, Codex job, or republish failures may remain hidden.

→
Benefit

Matt gets a working workflow and a trustworthy migration decision point.

Done and 80/20

A real Sonnet session handles interruption and reconnect, a deterministic startup test proves Grok fallback, a Codex sub-agent returns a receipt, private Git holds one complete transcript, and the plan reflects the brain dump. The 80/20 is one live review, reconnect, plan change, and sub-agent job.

flowchart LR
  A[Components only] --> B[Run real integrated pilot]
  B --> C[Matt can accept workflow]
flowchart TD
  S1[Run voice review] --> S2[Test reconnect and subagent]
  S2 --> S3[Regenerate review page]
  S3 --> D{Matt confirms pilot}
  D -->|No| S1
  D -->|Yes| DONE[Pilot accepted]
Your decision

T-009 — Prepare the post-acceptance migration plan

Undecided
Task

After Matt accepts the pilot, inventory every remaining voice surface and produce the reviewed plan that moves them into Voice-Platform one at a time. Depends on: T-008 and explicit acceptance.

Problem solved

The final target includes several voice agents whose behavior and privacy boundaries must not be guessed.

→
Cost of inaction

The pilot remains isolated and duplicated voice infrastructure continues elsewhere.

→
Benefit

Matt gets an evidence-based consolidation path without a big-bang outage.

Done and 80/20

Only after acceptance, an inventory maps routes, code, deployments, providers, prompts, tools, transcripts, privacy, and rollback, then orders staged migrations. The 80/20 is inventorying active surfaces and ranking the next two migrations by value and risk. This task is planning only.

flowchart LR
  A[Other agents still scattered] --> B[Inventory after pilot]
  B --> C[Safe migration sequence]
flowchart TD
  S1[Confirm pilot accepted] --> S2[Inventory voice surfaces]
  S2 --> S3[Create staged migration plan]
  S3 --> D{No production moves}
  D -->|No| S1
  D -->|Yes| DONE[Migration review ready]
Your decision