[Mixed-source orchestration] Phase 5: Preserve source context across follow-up turns

Parent: #1055

Phase 5 - Conversation Continuity and Selection Semantics

Planning baseline version: 0.250.061

Status: Proposed

Depends on: Phases 1-4 for full behavior

Objective

Make source selection and follow-up behavior predictable across turns without unnecessary reprocessing.

Explicit selections are authoritative. When no current selection exists, conversation history should be used first; previously grounded sources are reauthorized and rerun only when fresh evidence is required.

Current Behavior to Reuse

Conversation metadata already persists last_grounded_document_refs, and Chat already has a history-only sufficiency assessment followed by authorized fallback search.

The remediation should extend that contract to native tabular evidence rather than add a second memory mechanism.

Selection Precedence

1. Current explicit selection (selection_mode: selected)
2. Analyze All Documents (selection_mode: all, Analyze only)
3. History grounding (selection_mode: history)
4. Relevance candidates (selection_mode: relevance, Chat/Search)

Current explicit selection must never be silently merged with stale prior sources unless the user asks to include both.

Follow-Up Decision Flow

flowchart TD
    A[New prompt] --> B{Current explicit selection?}
    B -- Yes --> C[Resolve and process selected manifest]
    B -- No --> D{History sufficient?}
    D -- Yes --> E[Answer from saved conversation evidence]
    D -- No --> F[Load last grounded refs]
    F --> G[Reauthorize current access and source versions]
    G --> H[Classify sources by native engine]
    H --> I[Run only engines needed for fresh evidence]
    I --> J[Update grounding metadata]

Scope

1. Extend grounded references

Persist safe document-level fields needed for native routing:

• document ID
• scope and scope ID
• file name/extension
• source kind
• source version when available
• last engine used
• evidence status

Do not persist raw query results or secret-bearing source descriptors in conversation metadata.

2. History sufficiency

Retain the existing assessment, but include whether the new question requires:

• no document access
• narrative retrieval
• tabular recomputation
• both engines

Examples requiring tabular recomputation:

• new filters
• new grouping/dimensions
• fresh totals or comparisons
• a question referring implicitly to the previously used table

Examples that may use history:

• rephrasing the prior conclusion
• asking for a shorter summary
• asking what the previous result means

3. Authorization and source freshness

Before any historical source reuse:

• revalidate conversation ownership
• revalidate group membership/public visibility
• resolve the current source version
• fail or disclose stale evidence when the source changed

4. Selection UI/backend consistency

Send selection_mode explicitly. Panel visibility must not change source semantics.

Display a compact source-context indicator for the current turn when useful, but do not add instructional clutter to the chat response.

5. All Documents semantics

• Chat/Search: relevance-bounded
• Analyze: exhaustive authorized catalog within limits
• Compare: explicit Source and Targets only

Reject ambiguous all requests for Compare rather than guessing.

Acceptance Criteria

• A follow-up such as "now break that down by region" reuses and reauthorizes the prior table when no current selection exists.
• A follow-up answerable from the previous response does not rerun tabular tools.
• Selecting a new PDF supersedes the previously grounded table.
• A changed or unauthorized prior source is not reused silently.
• Current selection semantics are identical whether the panel is open or closed.
• Analyze All Documents is exhaustive; Chat/Search All Documents is relevance-bounded.
• Compare without explicit Source/Targets returns a clear validation error.

Tests

• history-sufficient tabular follow-up
• fresh tabular calculation follow-up
• mixed prior grounding requiring both engines
• explicit selection supersedes history
• group membership revoked between turns
• public workspace visibility revoked
• source version changed between turns
• no current selection and no grounded refs
• standard/streaming parity
• model/agent parity
• selection panel open/closed parity

Rollout

Use enable_mixed_source_conversation_continuity. Start with explicitly grounded tabular follow-ups, then enable mixed prior-source reruns.

Rollback

Disable native-engine reruns while retaining existing history-only and search fallback behavior.

Exit Gate

Proceed when follow-up turns make deterministic, observable decisions between history reuse and engine rerun, with authorization tests for every scope.

Not in This Phase

• Long-term semantic memory of raw table results
• Automatic inclusion of sources older than the latest grounded set
• Unbounded source catalogs
• Many-to-many Compare