Add phone call voice stack support

.env.example

@@ -19,7 +19,10 @@ INKBOX_SIGNING_KEY=whsec_xxxxxxxxxxxx
 # INKBOX_WEBHOOK_SECRET_GITHUB=...                     # verification secret for a registered source
 # INKBOX_GATEWAY_PORT=8767
 
-# --- Voice (on by default; Realtime auto-enables when a key is set) ---
+# --- Phone call voice stack ---
+# INKBOX_VOICE_STACK=inkbox_voice_ai                   # or openai_realtime / inkbox_tts_stt
+# INKBOX_VOICE_AI_AUTHORITY_MODE=contact_scoped        # or yolo; local mirror checked by doctor
+# INKBOX_VOICEMAIL_DETECTION=enabled                   # set disabled for calls that must ignore detection
 # OPENAI_API_KEY=sk-openai                             # enables OpenAI Realtime calls
 # INKBOX_REALTIME_API_KEY=sk-realtime                  # dedicated key; OPENAI_API_KEY backstops
 # INKBOX_VOICE_ENABLED=false                           # stop answering calls entirely

.github/workflows/canary.yml

@@ -15,7 +15,7 @@ jobs:
       - uses: actions/checkout@v7
         with:
           repository: inkbox-ai/inkbox
-          ref: a4dd76b534e33c0a8352148d8f0e9ab25199f05b
+          ref: 73f18a2b8c0e9dc6887c5663e6e904d54869927e
           path: .ci/inkbox
       - uses: actions/setup-node@v7
         with:
@@ -40,7 +40,7 @@ jobs:
       - uses: actions/checkout@v7
         with:
           repository: inkbox-ai/inkbox
-          ref: a4dd76b534e33c0a8352148d8f0e9ab25199f05b
+          ref: 73f18a2b8c0e9dc6887c5663e6e904d54869927e
           path: .ci/inkbox
       - uses: actions/setup-node@v7
         with:

.github/workflows/live-a2a.yml

@@ -58,13 +58,13 @@ jobs:
           python-version: "3.12"
 
       - name: Install protocol driver
-        run: pip install 'inkbox==0.5.8'
+        run: pip install 'inkbox==0.5.9'
 
       - name: Install plugin and host
         run: |
           npm ci
           npm install --no-save --package-lock=false \
-            @inkbox/sdk@0.5.8 \
+            @inkbox/sdk@0.5.9 \
             @opencode-ai/sdk@1.17.18 @opencode-ai/plugin@1.17.18
           npm install -g opencode-ai@latest
 
@@ -75,6 +75,7 @@ jobs:
           AUT_INKBOX_SIGNING_KEY: ${{ secrets.AUT_INKBOX_SIGNING_KEY }}
           INKBOX_BASE_URL: ${{ vars.INKBOX_BASE_URL || 'https://inkbox.ai' }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          INKBOX_VOICEMAIL_DETECTION: "disabled"
         run: bash scripts/live-aut.sh
 
       - name: Run ${{ matrix.scenario }}

.github/workflows/live-channels.yml

@@ -40,6 +40,7 @@ concurrency:
 jobs:
   live:
     runs-on: ubuntu-latest
+    timeout-minutes: 25
     strategy:
       fail-fast: false
       max-parallel: 1          # legs share the AUT identity → one at a time
@@ -51,7 +52,7 @@ jobs:
       - uses: actions/checkout@v7
         with:
           repository: inkbox-ai/inkbox
-          ref: a4dd76b534e33c0a8352148d8f0e9ab25199f05b
+          ref: 73f18a2b8c0e9dc6887c5663e6e904d54869927e
           path: .ci/inkbox
       - uses: actions/setup-node@v7
         with:
@@ -72,7 +73,8 @@ jobs:
           nohup node tests/live/mock-openai.mjs 8088 > "$RUNNER_TEMP/mock.log" 2>&1 &
           echo $! > "$RUNNER_TEMP/mock.pid"
           for i in $(seq 1 10); do
-            curl -sf http://127.0.0.1:8088/v1/models >/dev/null && { echo "mock model ready"; exit 0; }
+            curl -sf --connect-timeout 1 --max-time 3 \
+              http://127.0.0.1:8088/v1/models >/dev/null && { echo "mock model ready"; exit 0; }
             sleep 1
           done
           echo "::error::mock model did not start"; cat "$RUNNER_TEMP/mock.log"; exit 1
@@ -85,6 +87,7 @@ jobs:
           AUT_INKBOX_SIGNING_KEY: ${{ secrets.AUT_INKBOX_SIGNING_KEY }}
           INKBOX_BASE_URL: ${{ vars.INKBOX_BASE_URL }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          INKBOX_VOICEMAIL_DETECTION: "disabled"
         run: bash scripts/live-aut.sh
 
       - name: Run live channel tests (${{ matrix.mode }})
@@ -101,7 +104,7 @@ jobs:
       # Failure-only: these logs carry live phone/email content and this
       # repo's Action logs are public.
       - name: Dump logs (on failure only)
-        if: failure()
+        if: failure() || cancelled()
         run: |
           echo "=== gateway.log ==="; cat "$AUT_GATEWAY_LOG" || true
           echo "=== serve.log ===";   tail -n 100 "$AUT_SERVE_LOG" || true
@@ -115,7 +118,7 @@ jobs:
           kill "$(cat "$RUNNER_TEMP/mock.pid" 2>/dev/null)" 2>/dev/null || true
 
       - name: Upload artifacts (on failure only)
-        if: failure()
+        if: failure() || cancelled()
         uses: actions/upload-artifact@v7
         with:
           name: live-logs-${{ matrix.mode }}

.github/workflows/live-external-events.yml

@@ -41,7 +41,7 @@ jobs:
       - uses: actions/checkout@v7
         with:
           repository: inkbox-ai/inkbox
-          ref: a4dd76b534e33c0a8352148d8f0e9ab25199f05b
+          ref: 73f18a2b8c0e9dc6887c5663e6e904d54869927e
           path: .ci/inkbox
       - uses: actions/setup-node@v7
         with:
@@ -65,6 +65,7 @@ jobs:
           INKBOX_BASE_URL: ${{ vars.INKBOX_BASE_URL }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           INKBOX_WEBHOOK_SECRET_GITHUB: ${{ secrets.LIVE_GITHUB_WEBHOOK_SECRET }}
+          INKBOX_VOICEMAIL_DETECTION: "disabled"
         run: bash scripts/live-aut.sh
 
       - name: Run external-event test

.github/workflows/live-voice.yml

@@ -1,9 +1,10 @@
-name: Live — voice calls (Inkbox STT/TTS + Realtime)
+name: Live — voice calls (Voice AI + Realtime + Inkbox TTS/STT)
 
 # Boots the AUT gateway plus a driver process that bridges the other side of a
-# real phone call over its own Inkbox tunnel. Two matrix legs:
+# real phone call over its own Inkbox tunnel. Three available matrix legs:
 #   inbound_inkbox    — driver calls the agent; agent answers Inkbox STT/TTS.
 #   outbound_realtime — driver texts "call me"; agent calls back on Realtime.
+#   outbound_hosted   — driver requests a Voice AI call and one exact post-call SMS.
 # Each leg verifies the stored call transcript shows the agent spoke to the
 # caller. Real model + real calls — the priciest suite — so it runs only on
 # ready (non-draft) PRs + manual dispatch, and shares the AUT tunnel lock.
@@ -47,18 +48,19 @@ concurrency:
 jobs:
   voice:
     runs-on: ubuntu-latest
+    timeout-minutes: 15
     strategy:
       fail-fast: false
       max-parallel: 1          # legs share the AUT + driver identities → one at a time
       matrix:
-        scenario: ${{ fromJSON(inputs.include_inbound && '["inbound_inkbox","outbound_realtime"]' || '["outbound_realtime"]') }}
+        scenario: ${{ fromJSON(inputs.include_inbound && '["inbound_inkbox","outbound_realtime","outbound_hosted"]' || '["outbound_realtime","outbound_hosted"]') }}
 
     steps:
       - uses: actions/checkout@v7
       - uses: actions/checkout@v7
         with:
           repository: inkbox-ai/inkbox
-          ref: a4dd76b534e33c0a8352148d8f0e9ab25199f05b
+          ref: 73f18a2b8c0e9dc6887c5663e6e904d54869927e
           path: .ci/inkbox
       - uses: actions/setup-node@v7
         with:
@@ -82,8 +84,9 @@ jobs:
           INKBOX_BASE_URL: ${{ vars.INKBOX_BASE_URL }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           INKBOX_VOICE_ENABLED: "true"
-          # Inbound leg proves the Inkbox STT/TTS path; the others prove Realtime.
-          INKBOX_REALTIME_ENABLED: ${{ matrix.scenario == 'inbound_inkbox' && 'false' || 'true' }}
+          INKBOX_VOICE_STACK: ${{ matrix.scenario == 'outbound_hosted' && 'inkbox_voice_ai' || (matrix.scenario == 'outbound_realtime' && 'openai_realtime' || 'inkbox_tts_stt') }}
+          INKBOX_VOICEMAIL_DETECTION: "disabled"
+          INKBOX_REALTIME_ENABLED: ${{ matrix.scenario == 'outbound_realtime' && 'true' || 'false' }}
           INKBOX_REALTIME_API_KEY: ${{ secrets.OPENAI_API_KEY }}
         run: bash scripts/live-aut.sh
 
@@ -93,6 +96,13 @@ jobs:
           INKBOX_BASE_URL: ${{ vars.INKBOX_BASE_URL }}
           VOICE_DRIVER_STATE: ${{ runner.temp }}/voice_driver_state.json
         run: |
+          if [ "${{ matrix.scenario }}" = "outbound_hosted" ]; then
+            HOSTED_MARKER="$(node scripts/nato-marker.mjs "$GITHUB_RUN_ID" "$GITHUB_RUN_ATTEMPT")"
+            echo "HOSTED_POST_CALL_MARKER=$HOSTED_MARKER" >> "$GITHUB_ENV"
+            export VOICE_DRIVER_LINE="After we hang up, send me one SMS. Create the post-call action now with this exact SMS body: $HOSTED_MARKER. Read those five words back to me after the action is saved. Do not send it during the call."
+            export VOICE_DRIVER_LISTEN=180
+            export VOICE_DRIVER_AUTO_STOP=false
+          fi
           nohup node tests/live/voice-driver.mjs > "$RUNNER_TEMP/driver.log" 2>&1 &
           echo $! > "$RUNNER_TEMP/driver.pid"
           for _ in $(seq 1 30); do        # up to ~90s
@@ -115,7 +125,7 @@ jobs:
 
       # Failure-only: these logs carry live call content and this repo is public.
       - name: Dump logs (on failure only)
-        if: failure()
+        if: failure() || cancelled()
         run: |
           echo "=== gateway.log ==="; cat "$AUT_GATEWAY_LOG" || true
           echo "=== serve.log ===";   tail -n 100 "$AUT_SERVE_LOG" || true
@@ -130,7 +140,7 @@ jobs:
           sleep 3   # let the driver revert its number on exit
 
       - name: Upload artifacts (on failure only)
-        if: failure()
+        if: failure() || cancelled()
         uses: actions/upload-artifact@v7
         with:
           name: voice-logs-${{ matrix.scenario }}

.github/workflows/tests.yml

@@ -13,7 +13,7 @@ jobs:
       - uses: actions/checkout@v7
         with:
           repository: inkbox-ai/inkbox
-          ref: a4dd76b534e33c0a8352148d8f0e9ab25199f05b
+          ref: 73f18a2b8c0e9dc6887c5663e6e904d54869927e
           path: .ci/inkbox
       - uses: actions/setup-node@v7
         with:
@@ -39,7 +39,7 @@ jobs:
       - uses: actions/checkout@v7
         with:
           repository: inkbox-ai/inkbox
-          ref: a4dd76b534e33c0a8352148d8f0e9ab25199f05b
+          ref: 73f18a2b8c0e9dc6887c5663e6e904d54869927e
           path: .ci/inkbox
       - uses: actions/setup-node@v7
         with:
@@ -65,7 +65,7 @@ jobs:
       - uses: actions/checkout@v7
         with:
           repository: inkbox-ai/inkbox
-          ref: a4dd76b534e33c0a8352148d8f0e9ab25199f05b
+          ref: 73f18a2b8c0e9dc6887c5663e6e904d54869927e
           path: .ci/inkbox
       - uses: actions/setup-node@v7
         with:

CHANGELOG.md

@@ -1,5 +1,12 @@
 # Changelog
 
+## 0.2.8 (unreleased)
+
+- Adds explicit Inkbox Voice AI, OpenAI Realtime API, and Inkbox TTS/STT phone-call stacks to setup.
+- Routes hosted calls through Voice AI and reconciles `call.ended` commitments in the OpenCode session.
+- Guards hosted post-call SMS with an exact-target durable journal and a bounded correction policy.
+- Uses `@inkbox/sdk` 0.5.9 and disables voicemail detection throughout live call CI.
+
 ## 0.2.7 (unreleased)
 
 - Adds safely framed matched-contact memories to inbound email, SMS, iMessage,
@@ -31,7 +38,7 @@
   existing approval and recipient-allowlist controls.
 - Adds paginated task and message history with direction, participant,
   lifecycle, context, role, keyword, and timestamp filters.
-- The plugin requires `@inkbox/sdk` 0.5.8 or newer.
+- The plugin uses exactly `@inkbox/sdk` 0.5.9.
 
 ## 0.1.0 (unreleased)
 

README.md

@@ -44,8 +44,8 @@ That's the whole setup. The clone lives in `~/.inkbox-opencode/app`; the
 plugin is installed into your **global opencode config** (`~/.config/opencode`)
 with a one-file wrapper. The wizard creates a fresh Inkbox agent via
 self-signup (or takes an existing API key), enables iMessage, provisions a
-dedicated phone number, waits for your START opt-in, validates an OpenAI key
-for Realtime voice, mints the webhook signing key, picks the agent's working
+dedicated phone number, waits for your START opt-in, configures the phone call
+voice stack, mints the webhook signing key, picks the agent's working
 directory, and offers to **keep the gateway running on every boot**. When it
 finishes, text, email, or call your agent and it answers.
 
@@ -223,10 +223,14 @@ Outbound calls can originate from either line, chosen with `origination` on
   iMessage; the underlying number is never surfaced.
 
 When `origination` is omitted the plugin uses whichever line exists, and
-prefers the dedicated number when both do. Note that `inkbox_place_call`
-currently requires an audio bridge: pass `clientWebsocketUrl` per call or set
-the `callWebsocketUrl` option (env `INKBOX_CALL_WEBSOCKET_URL`) — Inkbox
-connects to that WebSocket for the call's media.
+prefers the dedicated number when both do. OpenAI Realtime and Inkbox TTS/STT
+calls require an audio bridge: pass `clientWebsocketUrl` per call or set the
+`callWebsocketUrl` option (env `INKBOX_CALL_WEBSOCKET_URL`). Inkbox Voice AI
+handles media itself and receives the call's required `purpose` as its task brief.
+Hosted outbound calls do not send a per-call authority override: Inkbox applies
+the saved Voice AI default. The wizard changes that server-side default only
+with an admin credential and records `INKBOX_VOICE_AI_AUTHORITY_MODE` as a local
+mirror so `doctor` can report configuration drift.
 
 ## Configuration reference
 
@@ -236,7 +240,10 @@ connects to that WebSocket for the call's media.
 | `identity` | `INKBOX_IDENTITY` (also `INKBOX_AGENT_IDENTITY`, `INKBOX_AGENT_HANDLE`) | Agent handle (required) |
 | `baseUrl` | `INKBOX_BASE_URL` | API base URL override |
 | `signingKey` | `INKBOX_SIGNING_KEY` | Webhook signature key (future inbound use) |
-| `callWebsocketUrl` | `INKBOX_CALL_WEBSOCKET_URL` | Audio-bridge WebSocket for `inkbox_place_call` |
+| `callWebsocketUrl` | `INKBOX_CALL_WEBSOCKET_URL` | Audio bridge used by OpenAI Realtime and Inkbox TTS/STT calls |
+| `phoneVoiceStack` | `INKBOX_VOICE_STACK` | `inkbox_voice_ai`, `openai_realtime`, or `inkbox_tts_stt` |
+| `voiceAiAuthorityMode` | `INKBOX_VOICE_AI_AUTHORITY_MODE` | Local mirror of saved Voice AI authority: `contact_scoped` or `yolo` |
+| `voicemailDetection` | `INKBOX_VOICEMAIL_DETECTION` | Optional explicit `enabled` / `disabled`; omission uses the Inkbox API default |
 | `vault.keyEnvVar` | — (default `INKBOX_VAULT_KEY`) | Which env var holds the vault unlock key |
 | `tools.enable` / `tools.disable` | — | Tool gating (names, groups, `"all"`) |
 | `outbound.approval` | — | `"ask"` (default) / `"allowlist"` / `"auto"` |
@@ -312,8 +319,10 @@ inbound events. What it does:
   decline") and time out to a decline.
 - **Control commands** (whole-message): `/clear`, `/stop`, `/status`,
   `/health`, `/resume`, `/usage`.
-- **Voice** (on by default with the gateway): the agent answers calls. Realtime
-  auto-enables when an OpenAI key is present (`INKBOX_REALTIME_API_KEY`, or
+- **Voice** (on by default with the gateway): setup offers Inkbox Voice AI,
+  OpenAI Realtime API, and Inkbox TTS/STT. Voice AI handles the live call and
+  notifies OpenCode after it ends; the two local stacks keep the call attached
+  to the OpenCode gateway. Realtime uses `INKBOX_REALTIME_API_KEY` (or
   `OPENAI_API_KEY` as the backstop) and runs the call as a live raw-audio
   conversation with in-call actions; otherwise Inkbox handles speech-to-text
   and text-to-speech. Opt out with `INKBOX_VOICE_ENABLED=false` (stop answering)

docs/live-ci.md

@@ -0,0 +1,75 @@
+# Live CI
+
+These Actions exercise the installed plugin against live Inkbox identities. Each component Action supports reusable and manual execution; credential-gated tests skip outside configured live jobs. Tests use current-run markers or pre-request snapshots so stale records cannot pass.
+
+## Full stack e2e
+
+Runs the reusable Actions in sequence for ready same-repository pull requests, manual dispatches, and successful canary runs on `main`. The orchestrated voice call includes the conditional inbound scenario.
+
+### `full-stack`
+
+**Proves:** Every live suite passes as one required gate. **Flow:** 1. Run channels. 2. Run Agent2Agent. 3. Run voice. 4. Run external events. 5. Fail unless every suite succeeded.
+
+## Live — Agent2Agent
+
+Runs all four scenarios serially with both live identity credentials.
+
+### `inbound-single`
+
+**Proves:** The plugin completes one inbound A2A task. **Flow:** 1. Send a tagged task. 2. Wait for completion. 3. Require the tag in task history.
+
+### `inbound-multi`
+
+**Proves:** An inbound task can request and consume follow-up input. **Flow:** 1. Send a tagged task. 2. Wait for `input-required`. 3. Reply in the same task. 4. Require both tags at completion.
+
+### `outbound-single`
+
+**Proves:** The agent delegates work without completing its outer task early. **Flow:** 1. Request delegation. 2. Find the tagged worker task. 3. Complete it remotely. 4. Require its result in the outer completion.
+
+### `outbound-multi`
+
+**Proves:** Delegation preserves a worker's input round trip. **Flow:** 1. Start a delegated task. 2. Receive its input request. 3. Reply through the agent. 4. Complete the worker. 5. Require its result in the outer task.
+
+## Live — agent channels (email + SMS)
+
+The `mock` matrix leg runs only the deterministic tests; the `real` leg runs only the real-model tests. Both require live identity credentials.
+
+### `email — mock model: the nonce travels inbound → model → reply → delivery`
+
+**Proves:** The complete email transport works deterministically. **Flow:** 1. Snapshot inbound email IDs. 2. Send a unique nonce. 3. Wait for a fresh reply. 4. Require the nonce and mock marker.
+
+### `email — real model: replies with actual content`
+
+**Proves:** The real agent can answer over email. **Flow:** 1. Snapshot inbound email IDs. 2. Request a fixed acknowledgement. 3. Wait for a fresh reply. 4. Reject error text and require the acknowledgement.
+
+### `SMS — mock model: the nonce travels inbound → model → reply → delivery`
+
+**Proves:** The complete SMS transport works deterministically. **Flow:** 1. Send a unique nonce. 2. Wait for a fresh inbound reply. 3. Require the nonce and mock marker.
+
+### `SMS — real model: reports its own identity when asked`
+
+**Proves:** The real agent receives context and answers over SMS. **Flow:** 1. Read the agent mailbox. 2. Ask for that address by SMS. 3. Wait for a fresh reply. 4. Require the exact address.
+
+## Live — voice calls (Voice AI + Realtime + Inkbox TTS/STT)
+
+Requires both live identity credentials and a real model. Outbound Realtime and Voice AI always run; inbound TTS/STT runs only when `include_inbound` is true.
+
+### `inbound: driver calls, agent answers via Inkbox TTS/STT and replies`
+
+**Proves:** Inbound client-media calling uses Inkbox speech services. **Flow:** 1. Snapshot agent calls. 2. Place a call with voicemail detection disabled. 3. Require two-way speech. 4. Verify the persisted call policy and speech mode. 5. Hang up.
+
+### `outbound: 'call me' text → agent calls back on the Realtime path and replies`
+
+**Proves:** A message-triggered callback uses Realtime. **Flow:** 1. Snapshot both owners' calls. 2. Text the request. 3. Require exactly one fresh paired call after duplicate grace. 4. Require two-way speech, Realtime flags, and disabled voicemail detection. 5. Hang up.
+
+### `outbound: Voice AI call settles one exact-target post-call SMS`
+
+**Proves:** Hosted calling completes one durable post-call action. **Flow:** 1. Snapshot both call owners and sender-side SMS rows. 2. Request a hosted call. 3. Require one fresh pair, reason, saved authority, and disabled voicemail detection. 4. Before hangup, require caller intent and a matching open action. 5. Hang up. 6. Require completed reconciliation and exactly one current-marker SMS to the caller after duplicate grace.
+
+## Live — external events (webhook → agent acts)
+
+Runs only with live identity credentials, a real model, the webhook signing secret, and the gateway log used to correlate the exact turn.
+
+### `rejects forged GitHub hooks and completes a valid real-model turn`
+
+**Proves:** External events are authenticated before agent execution. **Flow:** 1. Send an invalidly signed event and require rejection with no turn. 2. Send a valid event. 3. Require acceptance. 4. Wait for the exact request's completed agent turn.