- Owner: Codex verification agent
- Accountable product: V6
- Priority: P0 live operational seam
- Linked task: V6-002, test durable memory continuity, welfare/safety policy escalation, voice consent, latency, failover, and incident response
- Evidence captured: 2026-07-25
Exercised boundary#
This acceptance pass exercised the repository's canonical production seams rather than introducing a second test-only implementation.
Durable memory continuity#
The shell assistant's Iris bridge preserves consent, memory scope, retrieval, promotion, and disclosure state across session behavior. The BFF's canonical Iris adapter persists memory, consent records, opt-outs, deletion tombstones, and revisions through the shared durable snapshot store.
The live local PostgreSQL exercise granted the required storage and processing consents, wrote member and adjacent-member memories, reconstructed the adapter from a fresh process/store instance, and recovered the exact member content. It also proved inactivity expiry, exact-subject erasure, adjacent-subject isolation, restart after erasure, and rejection of a later write for the erased subject.
Welfare and safety escalation#
Lilith's V6 policy rejects steward ownership operations and forced continuation
after refusal. Chronic low Care, repeated coercion, a stalled arc, and sustained
distress produce a Lilith review trigger and safety-transfer Departed event.
Player crisis content routes to the V1 regional resource flow, suppresses agent
counseling, suspends persona memory writes, and requests a crisis-trained
reviewer when severity requires it.
The live local crisis exercise invoked the authenticated activation route, published the crisis event through Redis, and verified that the worker durably projected a non-overridable frame to every contemplative surface in PostgreSQL. Every projected frame halted synthesis and suspended memory writes.
Voice consent#
The Tara TTS scope lock is the fail-closed runtime and publication authority for synthetic voice. Permission requires an active consent record whose voice signature, signature owner, playback instructor, tenant, purpose, and surface match the requested use. Missing configuration, missing locks, cross-tenant or cross-instructor use, unsupported purpose, revoked consent, expired consent, and malformed bindings deny use.
The adjacent admin route tests also verify grant-window classification and explicit denial for absent, expired, pending, revoked, or wrong-scope consent.
Latency and failover#
The canonical Psyche session suite measures and enforces:
- ASR final-text p50;
- first-token p50 and p95;
- voice first-audio p50, p95, and p99;
- avatar-frame p50;
- end-to-end pipeline ceilings; and
- transport round-trip p50 and p95.
Budget pressure produces deterministic background shedding, pipeline trimming, reconnect, or text fallback. Mid-turn provider failure preserves the stable assistant identity, ordered transcript, last acknowledgement, interruption state, memory scope, and user-visible degradation disclosure. Reconnect rejects identity drift, cross-session replay, duplicate sequence numbers, sequence gaps, and unsafe resurrection of interrupted audio.
Incident response#
The durable operator incident authority preserves acknowledgement, monotonic severity escalation, assignment, snooze restrictions, classified resolution, and full history. Fresh PostgreSQL-backed reconstruction recovered committed triage exactly. The exercise also proved operator pseudonymization, confidential rationale removal, adjacent-operator isolation, post-redaction fencing, and retention-bound pseudonym purge.
Verification#
| Command / gate | Result |
|---|---|
| Canonical Psyche embodiment suite | PASS — 24 files, 383 tests |
| V6 Lilith welfare and crisis policy suite | PASS — 1 file, 20 tests |
| Tara TTS voice contract/consent/scope-lock suite | PASS — 16 files, 69 tests |
| Iris bridge, engine integration, and disclosure continuity | PASS — 3 files, 44 tests |
| Focused BFF Iris durability, incident, crisis suppression, and voice-consent routes | PASS — 4 files, 56 tests |
| Live local PostgreSQL Iris restart and incident restart exercises | PASS — 2 files, 2 tests |
| Live local PostgreSQL + Redis crisis propagation exercise | PASS — 1 file, 1 test |
Total: 51 test files and 575 passing tests. The BFF integration run emitted non-fatal missing source-map warnings for prebuilt shared HTTP client JavaScript; all selected tests completed successfully.
Source lineage#
| Boundary | Source commit |
|---|---|
| Durable canonical Iris memory | a64a03415eda5b11a2edb283a7b2f2e2622326cf |
| V6 Lilith welfare and crisis policy | 8d3784c2ef2e7312be1732eca74b398705cef412 |
| Tara TTS voice scope lock | 5b9344bc2cc07969eea1e9177708759a2d3c33f2 |
| Psyche latency SLO dashboard | 5b6b14cdc1e217c053593d62c124103ba436ead4 |
| Psyche provider failover | 951e507f45f721e701531825fd5f8fa7b267be79 |
| Durable operator incident lifecycle | f74af22b7e296edf23827c6155d5ebeed1d5d376 |
| Durable crisis-frame runtime | 08395cc2c98483e5b973a1549c1578003e1b49bf |
Artifact hashes#
| Artifact | SHA-256 |
|---|---|
apps/oshun/bff/src/adapters/iris-memory-adapters.ts |
820701fd5fcef208417ad5f19e5b54c4269e4c7d933c4310aa5ed30eb69ac6d2 |
libs/oshun/shell-assistant/src/iris-memory-bridge.ts |
3adc0db16148a3f530f0e2ed46e8e12310551d6924270a1e8186fdf093e551e7 |
libs/v6/lilith-agent-welfare/src/index.ts |
8d28ed406750e54d2b9492754e15e8f242649a0c02ac242d108ec2d562aa270c |
libs/v3/tara-studio/src/tts-voice-scope-lock.ts |
532e64efc2159387aa75e1a788b47961daadf266a2afc8816e01da487b410cec |
libs/oshun/embodiment-psyche/src/latency-dashboard.ts |
e54f2a24f40c6448e48b2892a5fcce8b49944c2a25abdb763d0489e2cac8f8cb |
libs/oshun/embodiment-psyche/src/provider-failover.ts |
4600131e641b0008c481e1f0e86ea8817ca564cb49ee64cef3eda9e4c41ef6db |
apps/oshun/bff/src/admin/operator-incident-store.ts |
d54928135609e097fe64edf3848e22acb630a1ebc69bc1adc9507ef60e2f417f |
apps/oshun/bff/src/safety/crisis-frame-runtime.ts |
7418d56e50d78ebf0a19b81823ba91c263064c2c0c38bd5cb3d5bad8c4a96860 |
Honest boundary#
The persistence and propagation exercises used the live local PostgreSQL and Redis services, while provider failure and latency observations used controlled deterministic inputs at the production policy seams. This evidence does not claim production-region load, a real vendor outage, real customer crisis traffic, real voice-owner participation, longitudinal human assessment, or the separately open WebTransport/SFU, vendor ASR/TTS, V1 promotion, OTLP, and live Themis exercise.