Best possible solution:

Land a rebased, focused voice-call transcript fix that persists exactly one assistant turn, preserves current timer and BOOT.md safeguards, and splits or redesigns the boot-session memory work with separate security and upgrade proof.

Do we have a high-confidence way to reproduce the issue?

Yes. Source inspection on current main and v2026.6.11 shows final assistant realtime text is emitted as call.speaking.text while processEvent does not append that text to the call transcript; I did not run a live carrier call in this read-only review.

Is this the best way to solve the issue?

No. Persisting assistant text is the right direction, but this branch records too broad an event source, creates duplicate final paths, and mixes in boot/session-state changes that need a separate or corrected design.

Full review comments:

• [P1] Preserve BOOT.md echo protections when rebasing — src/gateway/boot.ts:38-49
  The PR head builds the boot prompt with raw BOOT.md content and omits current main's internal-runtime-context wrapper plus boot echo guard. Resolving conflicts by keeping this version can re-open prompt disclosure through final replies or message-tool sends.
  Confidence: 0.9
• [P1] Keep boot-run classification evidence after the run — src/gateway/boot.ts:139-142
  The new memory classifier only recognizes boot transcripts by finding a boot-run key in sessions.json, but runBootOnce deletes that same key immediately after agentCommand. The transcript file can remain while the classifier loses the only evidence that it was boot-generated, so BOOT.md content can still enter memory ingestion.
  Confidence: 0.9
• [P2] Avoid recording Telnyx TTS callbacks twice — extensions/voice-call/src/manager/events.ts:280-281
  This records bot text for every text-bearing call.speaking, but speak() already appends the bot utterance after provider.playTts, and Telnyx maps call.speak.started to call.speaking with the same text. Classic Telnyx TTS calls can therefore store duplicate bot transcript entries.
  Confidence: 0.9
• [P2] Emit one final Google transcript per utterance — extensions/google/realtime-voice-provider.ts:770-771
  When outputTranscription.finished is true, the bridge emits that chunk as a final callback and then flushes the accumulated aggregate as another final callback. The voice-call realtime handler persists every final assistant callback as call.speaking, so this can record both the last chunk and the aggregate as separate bot turns.
  Confidence: 0.88

Overall correctness: patch is incorrect
Overall confidence: 0.91

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning high; reviewed against 129f1bcf28e8.

Label justifications:

• P2: The PR targets a real bounded voice-call transcript persistence bug, but the remaining blockers are proof, correctness, conflicts, and session/security review rather than an emergency outage.
• merge-risk: 🚨 session-state: The diff changes persisted call transcripts and also adds boot-run session classification, memory ingestion, host SDK export, and reaper behavior.
• merge-risk: 🚨 security-boundary: The boot rewrite overlaps current BOOT.md prompt-disclosure protections and may let boot prompt content reach user-visible or memory paths if resolved incorrectly.
• merge-risk: 🚨 availability: The conflicting PR head lacks current call.speaking live-call timer refresh behavior, so conflict resolution must preserve the current availability guard.
• rating: 🧂 unranked krab: Overall readiness is 🧂 unranked krab; proof is 🧂 unranked krab and patch quality is 🧂 unranked krab.
• status: 📣 needs proof: The PR needs real behavior proof before ClawSweeper can clear the contributor ask. Needs real behavior proof before merge: The PR body describes live testing but provides no inspectable redacted rows, logs, terminal output, screenshots, recordings, or linked artifacts; contributor action is needed, with private details redacted, before ClawSweeper can clear proof. After adding proof, update the PR body; ClawSweeper should re-review automatically. If it does not, the PR author or someone with repository write access can comment @clawsweeper re-review.

PR surface:

Source +67, Tests +331, Docs +1. Total +399 across 17 files.

Security concerns:

• [medium] BOOT.md echo guard can be lost on rebase — src/gateway/boot.ts:38
  Current main wraps BOOT.md content in internal-runtime-context markers and registers a boot echo guard for message-tool sends; PR head lacks both protections on its boot prompt path.
  Confidence: 0.9
• [medium] Boot prompt transcript can remain unclassified — src/gateway/boot.ts:139
  The PR's memory classifier depends on a boot-run sessions.json key, but the boot runner deletes that key after the run, so BOOT.md prompt content can remain in a transcript without being marked as boot-generated for memory ingestion.
  Confidence: 0.9

What I checked:

• Repository policy read: Read full root AGENTS.md plus scoped extension and gateway AGENTS.md files; their proof, changelog, plugin-boundary, session-state, and security guidance affected this review. (AGENTS.md:1, 129f1bcf28e8)
• Live PR state: GitHub reports this PR open at head 578b502, mergeable CONFLICTING, merge state DIRTY, with failing Real behavior proof checks and no closing issue references. (578b502cefd2)
• Current main drops speaking text: Current main refreshes the live-call max-duration timer and transitions call.speaking, but it does not append event.text to the persisted transcript. (extensions/voice-call/src/manager/events.ts:283, 129f1bcf28e8)
• Current main emits assistant speaking text: Current main forwards final assistant realtime transcripts to the manager as text-bearing call.speaking events, making the missing persistence path source-reproducible. (extensions/voice-call/src/webhook/realtime-handler.ts:744, 129f1bcf28e8)
• Latest release has the same voice-call gap: v2026.6.11 still emits assistant realtime text as call.speaking.text while the manager's call.speaking handler does not persist that text. (extensions/voice-call/src/manager/events.ts:283, e085fa1a3ffd)
• Telnyx duplicate transcript risk: Telnyx maps call.speak.started to a text-bearing call.speaking, while speak() already appends the same bot utterance after provider.playTts. (extensions/voice-call/src/providers/telnyx.ts:188, 129f1bcf28e8)

Likely related people:

• steipete: Path history shows repeated recent work on voice-call manager/realtime handling and Google realtime behavior near the central transcript surfaces. (role: voice-call and Google realtime area contributor; confidence: high; commits: 2a39c217c87f, 5dccba7405c4, da7fb64aa485; files: extensions/voice-call/src/manager/events.ts, extensions/voice-call/src/webhook/realtime-handler.ts, extensions/google/realtime-voice-provider.ts)
• jalehman: Path history shows recent work on boot session mapping, memory transcript corpus metadata, and memory ingestion paths that overlap the PR's boot-session side changes. (role: recent session and memory area contributor; confidence: high; commits: acc2a0ee7297, c24d266b2d09, da50a450d271; files: src/gateway/boot.ts, packages/memory-host-sdk/src/host/session-files.ts, extensions/memory-core/src/dreaming-phases.ts)
• stainlu: Commit 95b2f9c added the current BOOT.md internal-context wrapping and echo suppression that this PR must preserve. (role: security-boundary contributor; confidence: high; commits: 95b2f9c6f9d4; files: src/gateway/boot.ts, src/gateway/boot-echo-guard.ts)
• vincentkoc: Path history shows recent voice-call path normalization, voice model catalog, replay-key, and Windows transcript-path work adjacent to the PR's voice and memory surfaces. (role: recent adjacent contributor; confidence: medium; commits: 905c9759a716, c8cee2dce4e3, 27b15a19e84c; files: extensions/voice-call/src/manager/events.ts, extensions/voice-call/src/webhook/realtime-handler.ts, extensions/google/realtime-voice-provider.ts)
• dynamite-bud: Path history shows Telnyx Media Streaming work adjacent to the PR's Telnyx/realtime transcript behavior. (role: Telnyx realtime feature contributor; confidence: medium; commits: bb1b075118e5; files: extensions/voice-call/src/webhook/realtime-handler.ts, extensions/voice-call/src/providers/telnyx.ts)

• 🦀 challenger crab: rare, exceptional readiness with strong proof, clean implementation, and convincing validation.
• 🦞 diamond lobster: very strong readiness with only minor maintainer review expected.
• 🐚 platinum hermit: good normal PR, likely mergeable with ordinary maintainer review.
• 🦐 gold shrimp: useful signal, but proof or patch confidence is still limited.
• 🦪 silver shellfish: thin signal; proof, validation, or implementation needs work.
• 🧂 unranked krab: not merge-ready because proof is missing/unusable or there are serious correctness or safety concerns.
• 🌊 off-meta tidepool: rating does not apply to this item.

Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

• ClawSweeper keeps one durable marker-backed review comment per issue or PR.
• Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
• A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
• PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
• Maintainers can also comment @clawsweeper review to request a fresh review only.
• Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
• Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
• Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.

• The egg game starts only after the PR passes the real-behavior proof check.
• Before that, no creature or rarity is rolled. The treat waits for real proof.
• This is still just collectible flavor: proof affects review readiness, not creature quality.