Once done has been set after a [DONE] frame, this unconditional ctx.Done() check can still return context.Canceled before the trailing blank line/EOF is consumed. In that case a successfully completed stream is reported as canceled and message_done is never emitted. A regression test that cancels after [DONE] but before EOF would reproduce this.

a.transcript.Height now uses max(lay.rightHeight/2, availableHeight), but renderWaterfall() still renders the transcript with max(6, height-activity-menu-prompt). When those diverge, the visual transcript/prompt positions no longer match the coordinates used by inputBounds() / activityBounds(), so mouse hit-testing and scrolling become offset on layouts where the activity pane or command menu is visible.

a.transcript.Height is now forced to at least half of the right pane, even when the activity panel, command menu, and prompt leave less space than that. The new renderWaterfall() path recomputes a smaller visible transcript height, so the viewport's internal height diverges from what is actually rendered. On shorter terminals this throws off AtBottom/scroll calculations and can clip or jump the transcript. The viewport height should stay in sync with the rendered container instead of enforcing a larger minimum here.

This is the new token-threshold auto-compact branch, but it still calls runCompactForSession(..., contextcompact.ModeManual, ...). As a result every compact_* event from an automatic compact is reported as manual, so the UI/docs added in this PR cannot distinguish an auto-compact from an explicit /compact. If auto-compaction is now a first-class path, it needs its own trigger mode instead of reusing manual.

"requested too many tokens" / "too many tokens" are too broad for context_too_long detection. Providers also use those phrases for bad-request cases like oversized max_tokens / requested output, which compaction cannot fix. In those cases this PR will route the error into the new reactive-compact retry path instead of surfacing the real configuration problem.

renderWaterfall now recomputes its own transcript height instead of using a.transcript.Height, but applyComponentLayout, inputBounds, and transcriptBounds still use a.transcript.Height. On short panes those values diverge (applyComponentLayout enforces a half-pane minimum, this code does not), which will misplace the rendered prompt/activity area relative to the viewport and break mouse hit-testing/scroll positioning.

Once [DONE] has been seen, this pre-read ctx.Done() check can still turn a successfully completed stream into context.Canceled if the request is canceled before the trailing blank line/EOF is consumed. In that case message_done never fires even though the model already finished. The cancel fast-path should be skipped after done == true, or handled after the pending completion is flushed.

autoCompacted is flipped before we know whether compacting actually succeeded. Because runCompactForSession(..., failOnError=false) swallows runner/save errors and can also return Applied=false, a failed or no-op auto-compact permanently disables further threshold-triggered compaction for the rest of the run. That leaves later oversized turns with no second chance to compact.

If the provider has already sent [DONE], canceling the request before the trailing blank line / EOF now returns context.Canceled here instead of finishing the stream. That skips message_done, so the last assistant turn and token usage can be dropped during normal shutdown races. Once done is true, this pre-read cancel check should stop treating the stream as incomplete.

The new "requested too many tokens" / "too many tokens" fragments are broad enough to catch ordinary 400 validation errors that are not caused by prompt/context size (for example, asking for too many completion tokens). Those requests will now be upgraded to context_too_long and sent through the reactive-compact retry path even though compact cannot fix them.

a.transcript.Height now uses max(minTranscriptHeight, availableHeight), but renderWaterfall still renders the transcript with just availableHeight. On short panes the viewport model thinks it is taller than the box actually drawn on screen, which throws off bottom detection/scroll math and causes clipping or jumpy scrolling.