feat(live-voice): honest human→agent GPT-Live bridge (v0.0.14)

[robotlearning123]

v0.0.14 — GPT-Live → coding-agent bridge (human → agent, observe-only)

A human talks to ChatGPT voice in a real signed-in browser; the observed human transcript routes to a coding agent; the reply reaches the human out-of-band (text overlay). GPT-Live silently drops client-injected speech, so there is no agent→Live "speak" path — this ships as an honest observe-only bridge, not a controllable TTS.

Reliable path

Real signed-in Chrome + sidecar/extension (TAP) + sidecar/agent-gateway.mjs (agent adapter + control plane). browser/sidecar.mjs (puppeteer + fake WAV mic) is a test harness — Live does not transcribe synthetic audio. Cloudflare Turnstile bypass is out of scope. Spec: docs/superpowers/plans/2026-07-11-gpt-live-bridge-layer-spec.md.

Honesty remediation (this update)

Cross-model review (cx GPT-5.6 + Opus) flagged the earlier Mode-B design as BLOCK. Fixed:

• Removed the agent→Live write channel (voice_live_send_text, POST /send_text, speak-queue) — it reported false delivered:true for a server-dropped op.
• Unified every path through the real chat_message_delta parser + isActionable; the reliable extension path now flows through the shared bridge and the control plane, so voice_live_* observe the real path.
• /end deadlock; group-killing agent timeout (src/agent-runner.mjs); gateway loopback + no-CORS + body cap + optional token; recursive secret redaction (py+js); version 0.0.14 + tool count 30 / 12 modules aligned across all surfaces; source-vs-wheel packaging disclosure.

MCP surface: voice_live_status / voice_live_get_transcript / voice_live_end / voice_live_export_help (observe + lifecycle only).

Gates (real output)

• sidecar npm test: 50 passed / 0 failed / 3 skipped
• pytest: 366 passed / 15 skipped · ruff clean
• Verify receipt: artifacts/verify/gpt-live-v0.0.14-remediation-2026-07-11.md

⚠️ Merge hold

Base is main, so this diff also includes the 0.0.13 voice-catalog lane. Per the release ordering (0.0.12 → 0.0.13 → 0.0.14), hold merge pending 0.0.12/0.0.13 landing and owner authorization. Do not merge without a "Go".

🤖 Generated with Claude Code

https://claude.ai/code/session_01VAbFRELj9M3MYQt7SkUaLG

Summary by CodeRabbit

• New Features

  • Added a read-only voice catalog tool, increasing available MCP tools to 30.
  • Added an experimental, observe-only GPT-Live bridge for human voice transcripts and agent replies displayed as text.
  • Added local controls for bridge status, transcripts, help, and session termination.
  • Added automated voice testing support using transcript, speech-to-text, and text-to-speech test doubles.

• Documentation

  • Updated product documentation, skill references, roadmap, FAQs, and release notes for voice capabilities and limitations.

• Bug Fixes

  • Improved session ending, timeout handling, reliability, and sensitive-data redaction.

• Chores

  • Updated the release version to 0.0.14 and removed an outdated QA report.

[coderabbitai]

📝 Walkthrough

Walkthrough

Changes

The release updates the project to version 0.0.14, adds a read-only list_voices MCP tool, introduces an observe-only GPT-Live human-to-agent bridge, adds sidecar runtime and browser tooling, and documents the supported boundaries and testing strategy.

Voice catalog and MCP surface

Layer / File(s)	Summary
Voice catalog and registration gpt2agent/tools/voice.py, gpt2agent/tools/__init__.py, gpt2agent/tools/voice_live.py	Adds normalized, redacted voice catalog access and four observe-only GPT-Live control tools.
Skill and product metadata README.md, gpt2agent/skills/*, server.json, pyproject.toml, gpt2agent/__init__.py	Aligns tool counts, descriptions, package versions, and documented voice capabilities.
Contract and integration tests tests/test_tools.py, tests/test_backend_tools.py, tests/test_voice_live.py, tests/test_install.py	Validates voice normalization, registration, annotations, live-contract gating, redaction, control behavior, and installed skill contents.

GPT-Live bridge and sidecar

Layer / File(s)	Summary
Transcript and bridge runtime sidecar/src/transcript.mjs, sidecar/src/events.mjs, sidecar/src/export.mjs	Parses consumer chat_message_delta events, filters actionable inbound utterances, buffers text-only transcripts, and records agent replies without a Live speak channel.
Control and session orchestration sidecar/src/control.mjs, sidecar/src/session.mjs, sidecar/src/adapter.mjs, sidecar/src/reconnect.mjs, sidecar/src/liveness.mjs	Adds localhost control routes, SDP exchange helpers, session states, reconnect backoff, and datachannel liveness tracking.
Gateway and browser integration sidecar/agent-gateway.mjs, sidecar/browser/*, sidecar/extension/*	Routes validated utterances to an agent command, applies body and token controls, renders replies as overlays, and provides browser-side transcript forwarding.
Protocol experiments and captures sidecar/experiments/*, sidecar/capture/*	Adds WebRTC, SDP, audio, bundle, handshake, injection, and live-monitoring harnesses for observed protocol behavior.
Runtime validation sidecar/test/*, sidecar/package.json	Tests control routes, event envelopes, transcript assembly, session safety, agent timeouts, reconnects, liveness, and opt-in voice loops.

Documentation and release evidence

Layer / File(s)	Summary
GPT-Live specifications and findings docs/superpowers/plans/*, docs/superpowers/specs/*, docs/superpowers/reviews/*	Documents the consumer protocol, handshake evidence, supported boundaries, architecture decisions, and deferred behavior.
Release and contributor documentation CHANGELOG.md, CLAUDE.md, CONTRIBUTING.md, docs/README.md, docs/roadmap.md, docs/faq.md, docs/how-it-works.md	Updates release notes, tool inventories, roadmap navigation, FAQ boundaries, and request-path descriptions.
Verification receipt artifacts/verify/gpt-live-v0.0.14-remediation-2026-07-11.md	Records remediation scope, validation gates, review findings, and the remaining real-browser validation boundary.

Estimated code review effort: 5 (Critical) | ~120 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Chrome
  participant BrowserHook
  participant ModeBExport
  participant AgentGateway
  participant CodingAgent
  Chrome->>BrowserHook: Emit inbound chat_message_delta events
  BrowserHook->>ModeBExport: Forward consumer datachannel payload
  ModeBExport->>ModeBExport: Assemble and filter human utterance
  ModeBExport->>AgentGateway: Submit actionable text
  AgentGateway->>CodingAgent: Run configured agent command
  CodingAgent-->>AgentGateway: Return text reply
  AgentGateway-->>ModeBExport: Return agent reply
  ModeBExport-->>BrowserHook: Expose text-only reply
  BrowserHook-->>Chrome: Render reply overlay

Loading

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 32.14% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.
Description check	⚠️ Warning	The PR description is detailed, but it misses the template's Type section and checklist items.	Add the required Type section and checklist checkboxes, and align the wording under What & why to the repository template.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title is concise and accurately summarizes the main v0.0.14 GPT-Live bridge change.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch release/v0.0.14-live-voice

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1d0f9f5738

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Send the /end response before running shutdown

When voice_live_end hits the real browser sidecar, onEnd is the callback in browser/sidecar.mjs that calls await shutdown(), and shutdown() awaits control.server.close(). Because this line awaits onEnd before the current /end response is written, server.close() can wait on the still-open request, causing the MCP call to time out and the sidecar to remain stuck; send the JSON response first or schedule shutdown after res.end.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Wait for the data channel before sending session updates

For a real RTCDataChannel created in connect(), the channel is still connecting immediately after setRemoteDescription; it only becomes usable after its open event. Sending the session update here can throw InvalidStateError in browsers/werift before the session ever reaches LIVE, so this should be deferred until the channel is open or guarded by an awaitable open step.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Unwrap data_message before recording event types

When ChatGPT Live emits the documented {type:"data_message", data:"<inner json>"} envelope, this records only data_message in eventTypes and never captures the inner transcript/response event names the harness is supposed to discover. That makes the capture evidence misleading for the exact consumer wire format used elsewhere in this change; parse outer.data before deduping, as the sidecar hook does.

Useful? React with 👍 / 👎.

[coderabbitai]

Actionable comments posted: 15

🧹 Nitpick comments (4)

sidecar/extension/hook.js (1)

57-57: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Add optional debug logging to the empty catch block.

Silently swallowing all parse errors makes it impossible to diagnose why utterances may be silently dropped. Consider a debug flag or console.debug to surface unexpected message formats without breaking resilience.

♻️ Optional: add debug logging

-        } catch {}
+        } catch (e) {
+          if (window.__gptliveDebug) console.debug("[gptlive-hook] parse error:", e?.message || e);
+        }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/extension/hook.js` at line 57, Update the empty catch block in the
utterance parsing flow to optionally emit debug information about the caught
parse error, using the existing debug configuration if available or a
non-intrusive console.debug fallback. Preserve the current resilient behavior by
continuing to swallow the error and avoid interrupting hook processing.

gpt2agent/tools/voice_live.py (1)

172-177: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Neither new tool module's register() matches the mandated conv=None signature. Path instructions for gpt2agent/tools/*.py state new tools must implement register(mcp, client, conv=None); both new files here omit conv, though this mirrors several pre-existing sibling files (e.g. account.py, instructions.py) that also skip it when unused.

• gpt2agent/tools/voice_live.py#L172-L177: add conv=None to register(mcp, client: Any = None) -> None.
• gpt2agent/tools/voice.py#L88-L88: add conv=None to register(mcp, client: BackendClient) -> None.

As per path instructions: "New tools must be implemented in gpt2agent/tools/<name>.py with a register(mcp, client, conv=None) function."

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@gpt2agent/tools/voice_live.py` around lines 172 - 177, Update register in
gpt2agent/tools/voice_live.py (lines 172-177) to accept the optional conv=None
parameter while preserving its existing client handling. Also update register in
gpt2agent/tools/voice.py (line 88) to include conv=None, ensuring both new tool
modules match the mandated register(mcp, client, conv=None) signature.

Source: Path instructions

sidecar/browser/demo-headless.mjs (1)

21-21: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Same hardcoded personal machine path repeated as the default Python interpreter in three experiment scripts.

All three default to /home/robot/workspace/47-chatgpt2agent/gpt2agent/.venv/bin/python when SDP_PY/PY isn't set, so anyone else following the header run instructions gets a silent ENOENT instead of the venv the repo actually ships.

• sidecar/browser/demo-headless.mjs#L21: resolve SDP_PY against the repo-relative venv (e.g. fileURLToPath(new URL("../../.venv/bin/python", import.meta.url))) or throw a clear error naming the required env var when unset, instead of falling back to the hardcoded absolute path.
• sidecar/experiments/connect_live.mjs#L15: apply the same repo-relative fallback / explicit-error fix.
• sidecar/experiments/connect_live_audio.mjs#L15: apply the same repo-relative fallback / explicit-error fix.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/browser/demo-headless.mjs` at line 21, Replace the hardcoded
personal-machine Python fallback in the interpreter configuration of
sidecar/browser/demo-headless.mjs:21, sidecar/experiments/connect_live.mjs:15,
and sidecar/experiments/connect_live_audio.mjs:15 with the repository-relative
.venv/bin/python resolution based on each module’s location, or fail clearly
when the required SDP_PY/PY environment variable is unset; preserve the existing
environment-variable override behavior in all three scripts.

sidecar/experiments/voice-agent-inject.mjs (1)

38-88: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Avoid eval() for the page hook; pass args directly to evaluateOnNewDocument.

eval(code) at Line 87 is flagged by static analysis (noGlobalEval) and is also fragile: if the target page enforces a CSP without unsafe-eval, the injected hook will throw and silently fail to install. voice-to-agent.mjs in this same PR shows the cleaner pattern — pass a real function plus its argument to evaluateOnNewDocument instead of stringifying and eval-ing.

♻️ Proposed refactor

-const PAGE_HOOK = `
-(wavB64) => {
+function pageHook(wavB64) {
   window.__pcs = []; window.__src = null; window.__replaced = false; window.__wavB64 = wavB64;
   ...
-}`;
+}
 
 const wavB64 = readFileSync(WAV).toString("base64");
 const browser = await puppeteer.connect({ browserURL: CDP, defaultViewport: null });
 const page = await browser.newPage();
-await page.evaluateOnNewDocument((code) => { eval(code); }, `(${PAGE_HOOK})(${JSON.stringify(wavB64)})`);
+await page.evaluateOnNewDocument(pageHook, wavB64);

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/experiments/voice-agent-inject.mjs` around lines 38 - 88, Replace the
eval-based page-hook installation in evaluateOnNewDocument with a directly
passed function and argument, following the pattern used by voice-to-agent.mjs.
Preserve PAGE_HOOK’s execution and wavB64 initialization while removing
stringification and global eval so the hook remains CSP-compatible.

Source: Linters/SAST tools

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@gpt2agent/tools/voice_live.py`:
- Around line 199-225: Wrap each blocking _request call in voice_live_status,
voice_live_get_transcript, and voice_live_end with asyncio.to_thread, awaiting
the result before passing it to _strip_secrets. Ensure the module imports
asyncio and preserves the existing HTTP methods, paths, and clear query
behavior.

In `@sidecar/agent-gateway.mjs`:
- Around line 40-49: Update the request-body accumulation in the req data
handler to store raw chunks as byte buffers and track total byte length against
MAX_BODY. Decode the concatenated buffer once after all chunks are received,
preserving the existing oversized-body rejection behavior and avoiding
incremental UTF-8 decoding.

In `@sidecar/experiments/fetch-bundle-and-bootstrap.mjs`:
- Around line 26-33: Update the datachannel logging around __onBoot and its
BOOT_LOG persistence to avoid writing unredacted chat content: sanitize inbound
and outbound payloads by retaining only type and structural metadata, or ensure
any full-payload log is created with restrictive 0o600 permissions. Preserve the
existing bootstrap diagnostics while preventing sensitive transcript fields from
being stored at the predictable temporary path.

In `@sidecar/experiments/realtime-spike.mjs`:
- Around line 14-16: Replace the global WebSocket construction with a WebSocket
client/API that explicitly supports custom handshake headers, preserving the
Realtime URL and MODEL while ensuring the Authorization bearer token from KEY is
sent during connection setup. Do not rely on changing the Node engine version or
passing headers through the built-in WebSocket options.

In `@sidecar/experiments/voice-agent-inject.mjs`:
- Around line 73-81: Update the injection branch in the interval callback so
window.__replaced is set synchronously before invoking makeTrack(), preventing
later poll ticks from starting duplicate asynchronous injections. Keep the
existing makeTrack(), replaceTrack(), success reporting, and error reporting
behavior unchanged.

In `@sidecar/extension/background.js`:
- Around line 15-23: Apply explicit deadlines to both network boundaries: in
sidecar/extension/background.js lines 15-23, add an abortable timeout to the
gateway fetch and return a clear timeout response; in sidecar/src/adapter.mjs
lines 50-58, accept a timeout or abort signal for the SDP exchange, abort it
when the deadline expires, and add coverage for the timeout path. Use the
existing request-handling symbols at each site and preserve normal success and
error responses.

In `@sidecar/extension/relay.js`:
- Around line 6-17: Strengthen the message handler in the inbound relay by
requiring an extension-only capability check before calling
chrome.runtime.sendMessage. Add a per-session secret or equivalent unforgeable
validation shared with the legitimate extension hook, validate it alongside
__gptlive_utterance in the window message listener, and ignore messages that
fail it while preserving the existing forwarding and reply behavior for valid
messages.

In `@sidecar/src/adapter.mjs`:
- Around line 65-74: In sidecar/src/adapter.mjs lines 65-74, set LIVE_CONFIRMED
to false until media establishment and a complete round trip are verified, while
preserving the existing transport evidence and NEEDS_LIVE_CONFIRMATION markers.
In sidecar/experiments/test-speak-inject.mjs lines 178-183, update the
validation to require correlated output-transcript and received-audio evidence
rather than accepting a marker found anywhere.

In `@sidecar/src/control.mjs`:
- Around line 73-111: Strengthen the control-plane request handling in the
http.createServer callback by validating the request Origin/Host before serving
/status or /transcript and before allowing POST /end. Reject missing or
non-local origins/hosts with an appropriate 4xx response, while preserving
access for trusted loopback callers and leaving /health and /help behavior
unchanged as intended.

In `@sidecar/src/export.mjs`:
- Around line 41-75: Update the isBlocked logic in redactForAgent to also reject
lowercased keys ending with "secret" and "cookie", alongside the existing
underscored checks. Preserve the current blocklist and all other suffix matching
behavior.

In `@sidecar/src/session.mjs`:
- Around line 127-135: Update close() so a failure from this._dc.close() cannot
prevent this._pc.close() from being attempted; isolate the two close operations
with separate error-handling scopes while preserving the existing
exportPlane.close() and State.CLOSED finalization.
- Around line 79-107: Store the newly created peer connection in this._pc
immediately after createPeer succeeds, before any asynchronous handshake steps
in connect(). Ensure partial failures remain reachable by close(), while
preserving the existing successful-handshake state transition and reconnect
reset behavior.
- Around line 44-64: Update the session loop around the liveness monitor to
periodically call `isDead()`. When it reports expiration, close the current
session and trigger the existing reconnect flow, reusing the configured
`LivenessMonitor`, `_pc`, and `ReconnectPolicy` behavior without changing normal
activity handling.

In `@sidecar/src/voice-provider.mjs`:
- Around line 60-69: Update start() to clear any existing interval in this._poll
before assigning the new setInterval handle, ensuring repeated starts do not
leave the previous polling timer running.

In `@sidecar/test/voice-test-cases.md`:
- Around line 24-25: Insert a blank line between each Layer A, B, C, D, and E
heading and its following Markdown table in voice-test-cases.md, preserving the
existing headings and table contents.

---

Nitpick comments:
In `@gpt2agent/tools/voice_live.py`:
- Around line 172-177: Update register in gpt2agent/tools/voice_live.py (lines
172-177) to accept the optional conv=None parameter while preserving its
existing client handling. Also update register in gpt2agent/tools/voice.py (line
88) to include conv=None, ensuring both new tool modules match the mandated
register(mcp, client, conv=None) signature.

In `@sidecar/browser/demo-headless.mjs`:
- Line 21: Replace the hardcoded personal-machine Python fallback in the
interpreter configuration of sidecar/browser/demo-headless.mjs:21,
sidecar/experiments/connect_live.mjs:15, and
sidecar/experiments/connect_live_audio.mjs:15 with the repository-relative
.venv/bin/python resolution based on each module’s location, or fail clearly
when the required SDP_PY/PY environment variable is unset; preserve the existing
environment-variable override behavior in all three scripts.

In `@sidecar/experiments/voice-agent-inject.mjs`:
- Around line 38-88: Replace the eval-based page-hook installation in
evaluateOnNewDocument with a directly passed function and argument, following
the pattern used by voice-to-agent.mjs. Preserve PAGE_HOOK’s execution and
wavB64 initialization while removing stringification and global eval so the hook
remains CSP-compatible.

In `@sidecar/extension/hook.js`:
- Line 57: Update the empty catch block in the utterance parsing flow to
optionally emit debug information about the caught parse error, using the
existing debug configuration if available or a non-intrusive console.debug
fallback. Preserve the current resilient behavior by continuing to swallow the
error and avoid interrupting hook processing.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 8d0d73ae-696f-4cfd-a1a8-5b8c1994f3c1

📥 Commits

Reviewing files that changed from the base of the PR and between 04e3d93 and 68dab90.

⛔ Files ignored due to path filters (1)

• sidecar/package-lock.json is excluded by !**/package-lock.json

📒 Files selected for processing (86)

• .claude-plugin/marketplace.json
• .claude-plugin/plugin.json
• CHANGELOG.md
• CLAUDE.md
• CONTRIBUTING.md
• QA_REPORT.html
• README.md
• artifacts/verify/gpt-live-v0.0.14-remediation-2026-07-11.md
• docs/README.md
• docs/faq.md
• docs/how-it-works.md
• docs/roadmap.md
• docs/superpowers/plans/2026-07-10-v0.0.13-voice-release.md
• docs/superpowers/plans/2026-07-11-gpt-live-bridge-layer-spec.md
• docs/superpowers/plans/2026-07-11-gpt-live-full-pipeline.md
• docs/superpowers/plans/2026-07-11-gpt-live-handshake-evidence.md
• docs/superpowers/plans/2026-07-11-gpt-live-protocol-spec.md
• docs/superpowers/plans/2026-07-11-v0.0.14-live-voice-investigation.md
• docs/superpowers/reviews/2026-07-10-account-native-feature-coverage-cross-model-review.md
• docs/superpowers/specs/2026-07-10-account-native-feature-coverage-design.md
• gpt2agent/__init__.py
• gpt2agent/install.py
• gpt2agent/skills/gpt2agent/SKILL.md
• gpt2agent/skills/gpt2agent/tools-reference.md
• gpt2agent/tools/__init__.py
• gpt2agent/tools/voice.py
• gpt2agent/tools/voice_live.py
• pyproject.toml
• server.json
• sidecar/.gitignore
• sidecar/README.md
• sidecar/agent-gateway.mjs
• sidecar/browser/demo-headless.mjs
• sidecar/browser/sidecar.mjs
• sidecar/capture/gpt-live-capture.js
• sidecar/experiments/chunk-graph-fetch.mjs
• sidecar/experiments/connect_live.mjs
• sidecar/experiments/connect_live_audio.mjs
• sidecar/experiments/coop-fetch.mjs
• sidecar/experiments/diag-voice-entry.mjs
• sidecar/experiments/fakemic-diag.mjs
• sidecar/experiments/fetch-bundle-and-bootstrap.mjs
• sidecar/experiments/live-monitor.mjs
• sidecar/experiments/realtime-spike.mjs
• sidecar/experiments/sdp_exchange.py
• sidecar/experiments/sdp_exchange_full.py
• sidecar/experiments/test-inject-candidates.mjs
• sidecar/experiments/test-speak-inject-cdp.mjs
• sidecar/experiments/test-speak-inject.mjs
• sidecar/experiments/text-steer-watcher.mjs
• sidecar/experiments/voice-agent-inject.mjs
• sidecar/experiments/voice-to-agent.mjs
• sidecar/extension/background.js
• sidecar/extension/hook.js
• sidecar/extension/manifest.json
• sidecar/extension/relay.js
• sidecar/package.json
• sidecar/src/adapter.mjs
• sidecar/src/agent-runner.mjs
• sidecar/src/control.mjs
• sidecar/src/events.mjs
• sidecar/src/export.mjs
• sidecar/src/liveness.mjs
• sidecar/src/realtime-provider.mjs
• sidecar/src/reconnect.mjs
• sidecar/src/session.mjs
• sidecar/src/transcript.mjs
• sidecar/src/voice-provider.mjs
• sidecar/test/adapter.test.mjs
• sidecar/test/agent-runner.test.mjs
• sidecar/test/control.test.mjs
• sidecar/test/events.test.mjs
• sidecar/test/export.test.mjs
• sidecar/test/liveness.test.mjs
• sidecar/test/realtime-stt.test.mjs
• sidecar/test/reconnect.test.mjs
• sidecar/test/session.test.mjs
• sidecar/test/transcript.test.mjs
• sidecar/test/voice-loop.test.mjs
• sidecar/test/voice-test-cases.md
• tests/test_audit_2026_07_09_tools.py
• tests/test_backend_tools.py
• tests/test_install.py
• tests/test_none_guards.py
• tests/test_tools.py
• tests/test_voice_live.py

💤 Files with no reviewable changes (1)

• QA_REPORT.html

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '\n== voice_live.py outline ==\n'
ast-grep outline gpt2agent/tools/voice_live.py --view expanded || true

printf '\n== relevant line ranges ==\n'
nl -ba gpt2agent/tools/voice_live.py | sed -n '1,280p'

Repository: robotlearning123/gpt2agent

Length of output: 997

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '\n== imports and helper bodies ==\n'
sed -n '1,260p' gpt2agent/tools/voice_live.py | cat -n

printf '\n== search for asyncio usage in file ==\n'
rg -n "asyncio|to_thread|_request\\(" gpt2agent/tools/voice_live.py

Repository: robotlearning123/gpt2agent

Length of output: 10113

---

Wrap the blocking control-plane calls in asyncio.to_thread().

voice_live_status, voice_live_get_transcript, and voice_live_end are async def but call _request(), which uses blocking urllib.request.urlopen with a 5s timeout. That will stall the MCP event loop during each request; run these calls in a thread instead.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@gpt2agent/tools/voice_live.py` around lines 199 - 225, Wrap each blocking
_request call in voice_live_status, voice_live_get_transcript, and
voice_live_end with asyncio.to_thread, awaiting the result before passing it to
_strip_secrets. Ensure the module imports asyncio and preserves the existing
HTTP methods, paths, and clear query behavior.

Source: Path instructions

[coderabbitai]

🗄️ Data Integrity & Integration | 🟡 Minor | ⚡ Quick win

🧩 Analysis chain

🌐 Web query:

Does concatenating Node.js Buffer chunks into a string with += corrupt multi-byte UTF-8 characters split across chunks?

💡 Result:

Yes, concatenating Node.js Buffer chunks into a string using the += operator (or by calling.toString on individual chunks) will corrupt multi-byte UTF-8 characters if they are split across chunk boundaries [1][2]. When a multi-byte character (such as many CJK characters or emojis) is split between two chunks, the bytes at the end of the first chunk and the beginning of the next chunk will be incomplete on their own [3][2]. If you convert each chunk to a string independently, the decoder will be unable to recognize the multi-byte sequence and will typically replace the incomplete bytes with the Unicode replacement character (U+FFFD,), resulting in permanent data corruption [3][4][2]. To avoid this corruption, you should ensure that multi-byte character sequences are decoded only when they are complete. You can achieve this using one of the following methods: 1. Collect all chunks into an array, concatenate them using Buffer.concat, and then convert the resulting single Buffer to a string [3][1][2]: const bufs = []; for await (const chunk of readableStream) { bufs.push(Buffer.isBuffer(chunk)? chunk: Buffer.from(chunk)); } const result = Buffer.concat(bufs).toString('utf8'); 2. Use the node:string_decoder module, which is specifically designed to handle this scenario. The StringDecoder class maintains an internal buffer to hold incomplete multi-byte sequences until the next chunk arrives [5][6]: const { StringDecoder } = require('node:string_decoder'); const decoder = new StringDecoder('utf8'); let result = ''; for await (const chunk of readableStream) { result += decoder.write(chunk); } result += decoder.end; 3. Use readable.setEncoding('utf8') on a stream, which internally manages a StringDecoder to ensure that chunks are correctly reassembled into valid strings before they are emitted [7][8]. Note that this relies on the stream implementation correctly utilizing the internal encoding logic [3][2].

Citations:

• 1: https://stackoverflow.com/questions/40238063/why-use-buffer-concatbody-tostring-instead-of-uint8array-buffer-tostring
• 2: fix: use Buffer.concat for UTF-8 response body to prevent multi-byte character corruption elastic/elastic-transport-js#364
• 3: setEncoding('utf8') on response body corrupts multi-byte UTF-8 characters at chunk boundaries nodejs/undici#5002
• 4: http://nodejs.org/api/buffer.html
• 5: https://nodejs.org/api/string_decoder.html
• 6: https://medium.com/@farhad.gulizada/string-decoder-module-in-node-js-beginner-series-d20be6b81104
• 7: https://nodejs.org/dist/latest/docs/api/stream.html
• 8: https://milddev.com/converting-javascript-buffers-to-strings

---

Accumulate request chunks as bytes before decoding.
body += chunk decodes each Buffer independently, so multi-byte UTF-8 characters split across chunks can be corrupted. body.length also tracks string length, not raw bytes, so the size limit can drift from the actual request size.

Switch to byte counting and decode once at the end

 function readBodyLimited(req) {
   return new Promise((resolve, reject) => {
-    let body = "";
+    const chunks = [];
+    let size = 0;
     let over = false;
     req.on("data", (chunk) => {
       if (over) return;
-      body += chunk;
-      if (body.length > MAX_BODY) {
+      size += chunk.length;
+      if (size > MAX_BODY) {
         over = true;
         req.destroy();
         reject(new Error("body too large"));
+        return;
       }
+      chunks.push(chunk);
     });
     req.on("end", () => {
-      if (!over) resolve(body);
+      if (!over) resolve(Buffer.concat(chunks).toString("utf8"));
     });
     req.on("error", reject);
   });
 }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	let body = "";
	let over = false;
	req.on("data", (chunk) => {
	if (over) return;
	body += chunk;
	if (body.length > MAX_BODY) {
	over = true;
	req.destroy();
	reject(new Error("body too large"));
	}
	const chunks = [];
	let size = 0;
	let over = false;
	req.on("data", (chunk) => {
	if (over) return;
	size += chunk.length;
	if (size > MAX_BODY) {
	over = true;
	req.destroy();
	reject(new Error("body too large"));
	return;
	}
	chunks.push(chunk);
	});
	req.on("end", () => {
	if (!over) resolve(Buffer.concat(chunks).toString("utf8"));
	});

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/agent-gateway.mjs` around lines 40 - 49, Update the request-body
accumulation in the req data handler to store raw chunks as byte buffers and
track total byte length against MAX_BODY. Decode the concatenated buffer once
after all chunks are received, preserving the existing oversized-body rejection
behavior and avoiding incremental UTF-8 decoding.

[coderabbitai]

🔒 Security & Privacy | 🟡 Minor | ⚡ Quick win

Captured datachannel payloads may include real chat content, written unredacted to a predictable /tmp path.

window.__onBoot persists up to 4000 chars of inbound (and 2000 of outbound) raw datachannel JSON to BOOT_LOG (/tmp/gptlive-bootstrap.jsonl), which per this file's own protocol notes can include actual chat_message_delta conversation content, not just protocol metadata. On a shared machine this is a predictable, unprotected path for potentially sensitive transcript data to land.

Consider redacting/truncating message text fields before logging (keep only type/structural info), or write with restrictive permissions ({ mode: 0o600 }) if full payloads are needed for debugging.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/experiments/fetch-bundle-and-bootstrap.mjs` around lines 26 - 33,
Update the datachannel logging around __onBoot and its BOOT_LOG persistence to
avoid writing unredacted chat content: sanitize inbound and outbound payloads by
retaining only type and structural metadata, or ensure any full-payload log is
created with restrictive 0o600 permissions. Preserve the existing bootstrap
diagnostics while preventing sensitive transcript fields from being stored at
the predictable temporary path.

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Check for engines field in sidecar/package.json
rg -A2 '"engines"' sidecar/package.json

Repository: robotlearning123/gpt2agent

Length of output: 203

---

🏁 Script executed:

#!/bin/bash
sed -n '1,80p' sidecar/package.json
printf '\n---\n'
sed -n '1,80p' sidecar/experiments/realtime-spike.mjs

Repository: robotlearning123/gpt2agent

Length of output: 3201

---

🌐 Web query:

Node.js global WebSocket headers option undici stable version support custom headers constructor

💡 Result:

As of July 2026, the global WebSocket client built into Node.js (which uses the Undici library) does not natively support passing custom headers via the standard WebSocket constructor [1][2]. While the standalone Undici package historically added support for a WebSocketInit options object that included a headers field [3][4][5], this extension is generally not available in the global, stable WebSocket implementation provided by Node.js [2][6]. Node.js aligns its global WebSocket API closely with the browser-based standard, which intentionally excludes options for custom handshake headers [4]. Key points regarding the current status: - Stable Implementation: The built-in, global WebSocket client in Node.js reached stable status in v22.4.0 [7]. It is designed to be compliant with standard browser-based WebSocket APIs [4]. - Custom Headers Limitation: Because the global WebSocket implementation follows standard web specs, there is no standardized way to provide custom headers (such as Authorization) during the handshake process [4][2][6]. - Alternatives: If your application requires custom headers for the handshake (e.g., for authentication or custom routing), you must continue to use the established third-party 'ws' library, which explicitly supports these features [6]. In summary, despite some custom extensions existing within the underlying Undici library code, they are not exposed in the stable, global Node.js WebSocket API [2][6].

Citations:

• 1: WebSocket (and probably fetch) does not support Authorization header from URL nodejs/undici#4744
• 2: WebSocket connection fails on Node.js 24 — built-in WebSocket doesn't support auth headers agentmail-to/agentmail-node#12
• 3: https://github.com/nodejs/undici/blob/main/docs/docs/api/WebSocket.md
• 4: websocket: add websocketinit nodejs/undici#2088
• 5: tlkrs/undici@a3efc98
• 6: https://devcraftly.com/nodejs/websockets-ws/
• 7: https://nodejs.org/learn/getting-started/websocket.html

---

Global WebSocket won’t send these headers

new WebSocket(url, { headers: ... }) uses an option the built-in Node WebSocket doesn’t expose, so the Authorization header won’t reach the handshake. Use a client/API that supports handshake headers instead; bumping the Node engine floor won’t fix this.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/experiments/realtime-spike.mjs` around lines 14 - 16, Replace the
global WebSocket construction with a WebSocket client/API that explicitly
supports custom handshake headers, preserving the Realtime URL and MODEL while
ensuring the Authorization bearer token from KEY is sent during connection
setup. Do not rely on changing the Node engine version or passing headers
through the built-in WebSocket options.

[coderabbitai]

🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win

Set window.__replaced synchronously before the async makeTrack() call to avoid a duplicate-injection race.

The guard is only flipped to true inside the resolved promise chain, so a slow AudioContext/decodeAudioData can let the next 400ms poll tick re-enter the same branch and start a second concurrent makeTrack()/replaceTrack().

🔒️ Proposed fix

-        if(isAudio && tr.sender){ makeTrack().then(t=>tr.sender.replaceTrack(t).then(()=>{window.__replaced=true; window.__on(...)})).catch(...); return; }
+        if(isAudio && tr.sender){ window.__replaced = true; makeTrack().then(t=>tr.sender.replaceTrack(t).then(()=>{ window.__on(...) })).catch(e=>{ window.__replaced=false; window.__on({sys:"inject_err",e:String(e&&e.message||e)}); }); return; }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	setInterval(()=>{ if(window.__replaced) return;
	try { window.__on({ sys: "poll", pcs: window.__pcs.length, trs: window.__pcs.reduce((s,pc)=>{try{return s+pc.getTransceivers().length}catch{return s}},0) }); } catch {}
	for(const pc of window.__pcs){ try{
	for(const tr of pc.getTransceivers()){
	const isAudio = (tr.receiver&&tr.receiver.track&&tr.receiver.track.kind==="audio") || (tr.sender&&tr.sender.track&&tr.sender.track.kind==="audio");
	if(isAudio && tr.sender){ makeTrack().then(t=>tr.sender.replaceTrack(t).then(()=>{window.__replaced=true; window.__on({sys:"inject", same: tr.sender.track===t, kind: tr.sender.track&&tr.sender.track.kind, senderTrackId: tr.sender.track&&tr.sender.track.id, injectedId: t.id});})).catch(e=>window.__on({sys:"inject_err",e:String(e&&e.message||e)})); return; }
	}
	}catch{} } }
	, 400);
	setInterval(()=>{ if(window.__replaced) return;
	try { window.__on({ sys: "poll", pcs: window.__pcs.length, trs: window.__pcs.reduce((s,pc)=>{try{return s+pc.getTransceivers().length}catch{return s}},0) }); } catch {}
	for(const pc of window.__pcs){ try{
	for(const tr of pc.getTransceivers()){
	const isAudio = (tr.receiver&&tr.receiver.track&&tr.receiver.track.kind==="audio") || (tr.sender&&tr.sender.track&&tr.sender.track.kind==="audio");
	if(isAudio && tr.sender){ window.__replaced = true; makeTrack().then(t=>tr.sender.replaceTrack(t).then(()=>{window.__on({sys:"inject", same: tr.sender.track===t, kind: tr.sender.track&&tr.sender.track.kind, senderTrackId: tr.sender.track&&tr.sender.track.id, injectedId: t.id});})).catch(e=>{ window.__replaced=false; window.__on({sys:"inject_err",e:String(e&&e.message||e)}); }); return; }
	}
	}catch{} } }
	, 400);

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/experiments/voice-agent-inject.mjs` around lines 73 - 81, Update the
injection branch in the interval callback so window.__replaced is set
synchronously before invoking makeTrack(), preventing later poll ticks from
starting duplicate asynchronous injections. Keep the existing makeTrack(),
replaceTrack(), success reporting, and error reporting behavior unchanged.

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
rg -n "isDead" -g '*.mjs' sidecar

Repository: robotlearning123/gpt2agent

Length of output: 729

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '\n== session.mjs outline ==\n'
ast-grep outline sidecar/src/session.mjs --view expanded || true

printf '\n== liveness usage ==\n'
rg -n "liveness|isDead|planReconnect|ReconnectPolicy|close\(" sidecar/src sidecar -g '*.mjs'

printf '\n== session.mjs relevant slice ==\n'
sed -n '1,220p' sidecar/src/session.mjs

printf '\n== liveness.mjs relevant slice ==\n'
sed -n '1,220p' sidecar/src/liveness.mjs

Repository: robotlearning123/gpt2agent

Length of output: 12785

---

Liveness monitor needs a dead check
seen() is updated on human utterances and datachannel messages, but isDead() is never polled here. Half-open sessions can stay alive indefinitely; add a periodic check in this session loop that closes and reconnects when the timeout expires.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/src/session.mjs` around lines 44 - 64, Update the session loop around
the liveness monitor to periodically call `isDead()`. When it reports
expiration, close the current session and trigger the existing reconnect flow,
reusing the configured `LivenessMonitor`, `_pc`, and `ReconnectPolicy` behavior
without changing normal activity handling.

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

connect() leaks the peer connection on partial failure.

this._pc is only assigned after the full handshake succeeds (line 104). If any step between mic acquisition and setRemoteDescription throws (mic permission error, network failure, auth failure), the locally-created pc is never stored and becomes unreachable — close() can never reach it, leaking the RTCPeerConnection/ICE agent/track.

🔧 Proposed fix

   async connect() {
     this._setState(this.state === State.IDLE ? State.CONNECTING : State.RECONNECTING);
     const pc = this.opts.createPeer(this.opts.iceServers ?? []);
-    const track = await this.opts.getMicTrack();
-    pc.addTrack(track);
-    this._dc = pc.createDataChannel("", { negotiated: true, id: adapter.DATACHANNEL_ID });
-    this._dc.addEventListener("message", (m) => {
-      this.liveness.seen(Date.now());
-      void this.exportPlane.ingest(m.data);
-    });
-    const offer = await pc.createOffer();
-    await pc.setLocalDescription(offer);
-    const url = adapter.realtimeUrl({
-      mode: this.opts.voiceMode,
-      sessionType: this.opts.sessionType,
-    });
-    const { answerSdp } = await adapter.exchangeSdp({
-      url,
-      token: this.opts.auth.token,
-      offerSdp: offer.sdp,
-    });
-    await pc.setRemoteDescription({ type: "answer", sdp: answerSdp });
-    // No client-side session.update: the consumer channel silently drops it.
-    this._pc = pc;
-    this.reconnect.reset();
-    this._setState(State.LIVE);
+    try {
+      const track = await this.opts.getMicTrack();
+      pc.addTrack(track);
+      this._dc = pc.createDataChannel("", { negotiated: true, id: adapter.DATACHANNEL_ID });
+      this._dc.addEventListener("message", (m) => {
+        this.liveness.seen(Date.now());
+        void this.exportPlane.ingest(m.data);
+      });
+      const offer = await pc.createOffer();
+      await pc.setLocalDescription(offer);
+      const url = adapter.realtimeUrl({
+        mode: this.opts.voiceMode,
+        sessionType: this.opts.sessionType,
+      });
+      const { answerSdp } = await adapter.exchangeSdp({
+        url,
+        token: this.opts.auth.token,
+        offerSdp: offer.sdp,
+      });
+      await pc.setRemoteDescription({ type: "answer", sdp: answerSdp });
+      this._pc = pc;
+      this.reconnect.reset();
+      this._setState(State.LIVE);
+    } catch (err) {
+      pc.close();
+      this._dc = null;
+      throw err;
+    }
   }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	async connect() {
	this._setState(this.state === State.IDLE ? State.CONNECTING : State.RECONNECTING);
	const pc = this.opts.createPeer(this.opts.iceServers ?? []);
	const track = await this.opts.getMicTrack();
	pc.addTrack(track);
	this._dc = pc.createDataChannel("", { negotiated: true, id: adapter.DATACHANNEL_ID });
	this._dc.addEventListener("message", (m) => {
	this.liveness.seen(Date.now());
	// Fire-and-forget async agent turn. The reply is buffered as text for
	// out-of-band egress; it is NOT sent back to Live (injection is dropped).
	void this.exportPlane.ingest(m.data);
	});
	const offer = await pc.createOffer();
	await pc.setLocalDescription(offer);
	const url = adapter.realtimeUrl({
	mode: this.opts.voiceMode,
	sessionType: this.opts.sessionType,
	});
	const { answerSdp } = await adapter.exchangeSdp({
	url,
	token: this.opts.auth.token,
	offerSdp: offer.sdp,
	});
	await pc.setRemoteDescription({ type: "answer", sdp: answerSdp });
	// No client-side session.update: the consumer channel silently drops it.
	this._pc = pc;
	this.reconnect.reset();
	this._setState(State.LIVE);
	}
	async connect() {
	this._setState(this.state === State.IDLE ? State.CONNECTING : State.RECONNECTING);
	const pc = this.opts.createPeer(this.opts.iceServers ?? []);
	try {
	const track = await this.opts.getMicTrack();
	pc.addTrack(track);
	this._dc = pc.createDataChannel("", { negotiated: true, id: adapter.DATACHANNEL_ID });
	this._dc.addEventListener("message", (m) => {
	this.liveness.seen(Date.now());
	void this.exportPlane.ingest(m.data);
	});
	const offer = await pc.createOffer();
	await pc.setLocalDescription(offer);
	const url = adapter.realtimeUrl({
	mode: this.opts.voiceMode,
	sessionType: this.opts.sessionType,
	});
	const { answerSdp } = await adapter.exchangeSdp({
	url,
	token: this.opts.auth.token,
	offerSdp: offer.sdp,
	});
	await pc.setRemoteDescription({ type: "answer", sdp: answerSdp });
	this._pc = pc;
	this.reconnect.reset();
	this._setState(State.LIVE);
	} catch (err) {
	pc.close();
	this._dc = null;
	throw err;
	}
	}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/src/session.mjs` around lines 79 - 107, Store the newly created peer
connection in this._pc immediately after createPeer succeeds, before any
asynchronous handshake steps in connect(). Ensure partial failures remain
reachable by close(), while preserving the existing successful-handshake state
transition and reconnect reset behavior.

[coderabbitai]

🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win

close(): a throwing dc.close() skips pc.close().

Since both calls share one try, an exception from this._dc.close() jumps straight to finally, leaving this._pc never closed.

🔧 Proposed fix

   close() {
-    try {
-      this._dc?.close();
-      this._pc?.close();
-    } finally {
-      this.exportPlane.close();
-      this._setState(State.CLOSED);
-    }
+    try { this._dc?.close(); } catch { /* ignore */ }
+    try { this._pc?.close(); } catch { /* ignore */ }
+    this.exportPlane.close();
+    this._setState(State.CLOSED);
   }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	close() {
	try {
	this._dc?.close();
	this._pc?.close();
	} finally {
	this.exportPlane.close();
	this._setState(State.CLOSED);
	}
	}
	close() {
	try { this._dc?.close(); } catch { /* ignore */ }
	try { this._pc?.close(); } catch { /* ignore */ }
	this.exportPlane.close();
	this._setState(State.CLOSED);
	}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/src/session.mjs` around lines 127 - 135, Update close() so a failure
from this._dc.close() cannot prevent this._pc.close() from being attempted;
isolate the two close operations with separate error-handling scopes while
preserving the existing exportPlane.close() and State.CLOSED finalization.

[coderabbitai]

🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win

Guard against leaking the previous poll interval on repeated start() calls.

If start() runs again before a matching stop() (e.g. a reconnect path), the prior setInterval handle in this._poll is overwritten and never cleared, leaking a polling timer that keeps calling page.evaluate indefinitely.

🔧 Proposed fix

   async start() {
     const { page } = this.opts;
+    if (this._poll) clearInterval(this._poll);
     // Drain completed human utterances the hook parked on the page.
     this._poll = setInterval(async () => {

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	async start() {
	const { page } = this.opts;
	// Drain completed human utterances the hook parked on the page.
	this._poll = setInterval(async () => {
	try {
	const us = await page.evaluate(() => { const a = window.__utterances || []; window.__utterances = []; return a; });
	for (const u of us || []) for (const cb of this._cbs) cb(u);
	} catch {}
	}, 400);
	}
	async start() {
	const { page } = this.opts;
	if (this._poll) clearInterval(this._poll);
	// Drain completed human utterances the hook parked on the page.
	this._poll = setInterval(async () => {
	try {
	const us = await page.evaluate(() => { const a = window.__utterances || []; window.__utterances = []; return a; });
	for (const u of us || []) for (const cb of this._cbs) cb(u);
	} catch {}
	}, 400);
	}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/src/voice-provider.mjs` around lines 60 - 69, Update start() to clear
any existing interval in this._poll before assigning the new setInterval handle,
ensuring repeated starts do not leave the previous polling timer running.

[coderabbitai]

📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add blank lines before tables (MD058).

Static analysis flags 5 tables (Layer A/B/C/D/E) that immediately follow their heading with no blank line, which can break table rendering in stricter markdown parsers.

📝 Proposed fix (repeat for each flagged heading)

 ## Layer A — STT (does GPT-Live transcribe the utterance?) [T2/T3]
+
 | id | utterance | expected transcript (≈) |

Also applies to: 35-36, 44-45, 55-56, 62-63

🧰 Tools

🪛 markdownlint-cli2 (0.23.0)

[warning] 25-25: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@sidecar/test/voice-test-cases.md` around lines 24 - 25, Insert a blank line
between each Layer A, B, C, D, and E heading and its following Markdown table in
voice-test-cases.md, preserving the existing headings and table contents.

Source: Linters/SAST tools