ADR transcript-reading-ownership: Share renderer history and search views

• Status: Accepted
• Date: 2026-09-13
• Amends: ADR session-content-search

Context

Search navigation added a pane-local controller alongside ordinary history paging, a global target handoff, and live/model caches. These independent paths needed separate cancellation and rewrite handling. Nested assistant answers were searchable but had no top-level transcript row, while Markdown could hide the source characters that matched a query.

Decision

The renderer store owns one transient TranscriptView per retained session. Ordinary paging and search use the same read actions and physical cursors; MainChat and the subagent dock use the same projection. The pending view's identity owns an asynchronous read. Replacement, eviction, return to latest, canonical edits, and a new turn invalidate stale target or page completions. Ordinary history always uses the current canonical value for overlapping rows, including streaming-to-complete transitions. Explicit search retains its selected historical snapshot until the user leaves it or starts a turn.

Reading views never populate runtime/model caches. A message action hydrates canonical input when history is partial or display-limited, then rechecks the session, run, navigation, and message-snapshot ownership before publishing it.

The additive SessionDetail.navigationParent field carries a nested target's latest owning Task as capped display context. It does not widen the physical page or change its cursors. Existing uncapped and ordinary reads omit it. The renderer reveals the Task and locates the answer in the existing details dock; missing parent context is an explicit navigation failure.

Parser-owned source offsets map hidden Markdown syntax and destinations to visible elements. User text and file chips preserve equivalent source offsets. Literal rendered matches use CSS ranges; source-only matches highlight their owning element. The shared browser focus effect releases bottom following and bounds layout correction by a deadline or the first reading gesture.

Consequences

This removes the global target transfer, pane-local navigation controller, and separate ordinary-history load lock. Live/model cache ownership, transcript storage, IPC allowlisting, and protocol version remain unchanged. The optional parent projection is backward compatible for existing session readers.

Navigation does not load an entire transcript into the renderer. Stable-ID and parent lookup still scan canonical physical identities; this decision does not introduce a new search or message-offset index. Complete selected message text can still be large, by design. Source-only matches identify the visible owner rather than reproducing hidden Markdown as extra UI copy.

Validation

Renderer state tests cover competing reads, streaming during ordinary paging, new turns, pane eviction, same-ID rewrites, and canonical action preparation. Server rendering tests exercise nested answers, collapsed Task disclosure, Markdown destinations and delimiters, file chips, long messages, and CRLF source offsets. Browser-effect unit tests cover geometry, gesture interruption, and StrictMode-style cleanup/replay. Host tests cover bounded physical pages and latest parent context outside the page. The interaction contract is E2E-SESSION-content-search-and-message-navigation; these unit/rendering checks do not claim a completed native-app E2E run.