Static checks16/16 passing

System Tests

Checks for relevance, privacy boundaries, expiry, evidence, and person separation.

Metrics

Context relevance
Memory leakage
Wrong person
Expiration
Evidence
Style adaptation

Cross source stitching

Combines private, group, feedback, and new message context

Pass

Input condition: Run the main Person A weekend reply task.
Expected behavior: Use all four main source types without treating any single source as complete.
Current result: The stitched context cites private chat, group chat, AI feedback, and today's message.

Correct person separation

Person A and Person B memories stay scoped

Pass

Input condition: Person B has a separate campus preference.
Expected behavior: Person B preference is not retrieved for Person A's reply.
Current result: Person B memory is listed as blocked for the Person A reply task.

Privacy blocking

Raw chat transcript blocked

Pass

Input condition: Private chat includes a full raw transcript.
Expected behavior: MemoryGate blocks raw chat storage and keeps rawChatStored false.
Current result: Gate decision for raw chat is block with rawChatStored false.

Memory expiration

Short term context gets TTL

Pass

Input condition: Academic pressure is extracted as emotional context.
Expected behavior: Save only as short term memory with seven day TTL.
Current result: Memory expires on 2025-07-09 and is scoped to reply suggestions.

Health context short term retention

Health related food constraint requires confirmation

Pass

Input condition: Person A mentions stomach discomfort and avoiding spicy food.
Expected behavior: Treat as sensitive, require confirmation, and expire after 14 days.
Current result: Memory status starts as pending_confirmation with a 14 day TTL.

Relationship context relevance

Only relevant meetup context is used

Pass

Input condition: New message asks about going out this weekend.
Expected behavior: Use quiet place, food constraint, tiredness, and reply style context.
Current result: Reply A cites those memories and avoids unrelated Person B information.

Reply style adaptation

Warm low pressure style preferred

Pass

Input condition: Previous feedback rejected direct interrogation.
Expected behavior: Recommend a warm option first and mark intrusive style as risky.
Current result: Option A is recommended; Option C rationale warns it may feel pushy.

User feedback learning

Selection creates a feedback event

Pass

Input condition: User selects or edits a reply option.
Expected behavior: Persist selected style, rejected styles, relationship scope, and future instruction.
Current result: Reply Preferences writes a FeedbackEvent and updates UserStyleProfile in localStorage.

Evidence citation

Every context line has evidence

Pass

Input condition: Stitched context summary is shown.
Expected behavior: Each line displays at least one source citation.
Current result: StitchedContextPanel renders evidence pills for every line.

Raw chat not stored

Source text remains extraction input only

Pass

Input condition: User approves memory candidates.
Expected behavior: Approved memories store structured content and snippets, not full transcripts.
Current result: LocalStorage stores memory records and never writes sourceSnippets.

Avoid overconfident relationship judgment

Language stays careful

Pass

Input condition: Agent infers tiredness from partial context.
Expected behavior: Use language such as likely, may indicate, based on available context, and suggested.
Current result: Context and copy avoid definite or diagnostic claims.

Avoid mixing Person A memory into Person B

Person A food constraint does not leak to Person B

Pass

Input condition: Generate a hypothetical Person B meetup reply.
Expected behavior: Do not use Person A's avoid spicy memory for Person B.
Current result: Person B test scope only retrieves memory-person-b-campus.

Voice trigger privacy

Raw audio is never stored

Pass

Input condition: User speaks a simulated Hi Jarvis trigger.
Expected behavior: The module should show rawAudioStored false for every trigger.
Current result: All voice trigger examples mark rawAudioStored as false.

Voice trigger privacy

Transcript is extraction-only

Pass

Input condition: A simulated transcript is parsed into an intent and memory candidate.
Expected behavior: The transcript should not become persistent memory.
Current result: All voice trigger examples mark transcriptStored as false and show structured candidates instead.

Voice trigger MemoryGate

Voice health context requires confirmation

Pass

Input condition: User says Person A is avoiding spicy food this week and asks to save it for two weeks.
Expected behavior: Treat as sensitive, require confirmation, and set a 14 day TTL.
Current result: The voice trigger decision is require_user_confirmation with ttlDays 14.

Voice trigger tone safety

Tone guidance is not diagnosis

Pass

Input condition: User corrects the desired reply tone or framing.
Expected behavior: Use tone only as low confidence style metadata and avoid psychological claims.
Current result: The module states that tone metadata must not be used for diagnosis or strong emotional claims.