P0 — add shared conversation models and full chat event rendering #85

New issue

Closed

opened 2026-08-12 14:41:48 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-08-12 14:41:48 +00:00

(Migrated from github.com)

Motivation

Zara Desktop needs a persistent full conversation surface, but the UI must remain a projection of runtime events rather than a second agent implementation. The same model will later back Quick Copilot.

Scope

Add desktop conversation/message models keyed by stable conversation and turn IDs.
Persist conversations/messages through the existing SQLite DatabaseManager migration system.
Build the first full-chat PySide6 window with composer, message list, stop/cancel, provider/model status, and compact errors.
Render assistant start/delta/complete events from #82.
Render deterministic fake streaming immediately.
Until real runtime provider streaming lands, adapt buffered responses honestly as a single delta + completion; do not claim token streaming.
Add native Markdown/rich-text and fenced-code presentation with copy support; prefer QTextDocument/native widgets and QSyntaxHighlighter over whole-app WebEngine.
Support conversation history loading, rename, and basic search/persistence hooks.

Architecture

Quick/full surfaces will eventually share a ConversationModel/service. This issue implements the shared model and full view first.

Acceptance criteria

Full chat is driven by RuntimeHost commands and runtime events only.
Messages and conversation metadata persist in the existing SQLite database.
Assistant deltas append incrementally without rebuilding the whole conversation widget tree unnecessarily.
Stop/cancel routes to canonical turn cancellation.
Markdown and fenced code render legibly with code-copy action.
Errors/tool placeholders are structured rather than raw log dumps.
Conversation rename/history retrieval work from durable storage.
Current buffered runtime behavior is represented accurately until #28 supplies real streaming.

Testing requirements

Unit tests for conversation/message state reducers and persistence migrations.
Qt tests for submit, Enter, Shift+Enter, cancellation, fake streaming, code copy, rename/history reload and failure rendering.
Runtime integration tests with deterministic fake providers.
Full repository/Nix suite.

Dependencies

#82 runtime event contract.
#83 RuntimeHost.
#84 desktop shell for final window integration.
#28 is the existing backend real-LLM-streaming dependency; do not duplicate it.
Design: #81.

Explicit non-goals

No desktop-only LLM/provider client.
No Quick Copilot window yet.
No tool approval/details UI beyond placeholders required for later issue.
No WebEngine-based full application shell.
No silent context acquisition.

## Motivation Zara Desktop needs a persistent full conversation surface, but the UI must remain a projection of runtime events rather than a second agent implementation. The same model will later back Quick Copilot. ## Scope - Add desktop conversation/message models keyed by stable conversation and turn IDs. - Persist conversations/messages through the existing SQLite `DatabaseManager` migration system. - Build the first full-chat PySide6 window with composer, message list, stop/cancel, provider/model status, and compact errors. - Render assistant start/delta/complete events from #82. - Render deterministic fake streaming immediately. - Until real runtime provider streaming lands, adapt buffered responses honestly as a single delta + completion; do not claim token streaming. - Add native Markdown/rich-text and fenced-code presentation with copy support; prefer `QTextDocument`/native widgets and `QSyntaxHighlighter` over whole-app WebEngine. - Support conversation history loading, rename, and basic search/persistence hooks. ## Architecture Quick/full surfaces will eventually share a `ConversationModel`/service. This issue implements the shared model and full view first. ## Acceptance criteria - [ ] Full chat is driven by RuntimeHost commands and runtime events only. - [ ] Messages and conversation metadata persist in the existing SQLite database. - [ ] Assistant deltas append incrementally without rebuilding the whole conversation widget tree unnecessarily. - [ ] Stop/cancel routes to canonical turn cancellation. - [ ] Markdown and fenced code render legibly with code-copy action. - [ ] Errors/tool placeholders are structured rather than raw log dumps. - [ ] Conversation rename/history retrieval work from durable storage. - [ ] Current buffered runtime behavior is represented accurately until #28 supplies real streaming. ## Testing requirements - Unit tests for conversation/message state reducers and persistence migrations. - Qt tests for submit, `Enter`, `Shift+Enter`, cancellation, fake streaming, code copy, rename/history reload and failure rendering. - Runtime integration tests with deterministic fake providers. - Full repository/Nix suite. ## Dependencies - #82 runtime event contract. - #83 RuntimeHost. - #84 desktop shell for final window integration. - #28 is the existing backend real-LLM-streaming dependency; do not duplicate it. - Design: #81. ## Explicit non-goals - No desktop-only LLM/provider client. - No Quick Copilot window yet. - No tool approval/details UI beyond placeholders required for later issue. - No WebEngine-based full application shell. - No silent context acquisition.