fix(chat): stabilize streaming interactions

AGENTS.md

@@ -35,6 +35,7 @@
 - Conventional commits enforced by hook: `type(scope): subject` ≤ 50 chars; types: `feat|fix|docs|dx|style|refactor|perf|test|workflow|build|ci|chore|types|wip|release`.
 - Do not include AI co-authoring footers in commits.
 - PRs: clear description, link issues (`Closes #123`), screenshots/GIFs for UI, pass lint/typecheck/tests. Keep changes focused.
+- Default PR base is `dev`; use `gh pr create --base dev` for routine feature, bugfix, docs, test, and refactor branches. Target `main` only for `release/<version>` branches following `docs/release-flow.md`.
 - UI changes: include BEFORE/AFTER ASCII layout blocks to communicate structure.
 
 ## Architecture Notes & Security

docs/issues/markdown-codeblock-session-scroll-regressions/plan.md

@@ -0,0 +1,101 @@
+# Plan
+
+## Diagnosis
+
+### Code Block Toolbar
+
+`src/renderer/src/assets/style.css` imports `markstream-vue/index.tailwind.css`, but Tailwind still
+needs to scan the package's generated class candidates. The current source points at:
+
+```css
+@source '../../../../node_modules/markstream-vue/dist/tailwind.ts';
+```
+
+The installed `markstream-vue@1.0.0-rc.0` package ships `dist/tailwind.js` and
+`dist/tailwind.d.ts`, not `dist/tailwind.ts`. Because the source target does not exist, Tailwind
+does not see the class candidates used by the package's code block shell, including
+`code-block-header`, `px-[var(--ms-inset-panel-x)]`, `py-[var(--ms-inset-panel-y)]`, and
+`p-[var(--ms-action-btn-padding)]`.
+
+The package CSS import still provides variables and base styles, so the failure appears as a partial
+style regression instead of a fully unstyled component.
+
+### Session Switch Scroll
+
+`src/renderer/src/pages/ChatPage.vue` restores a session by loading messages, waiting for
+`nextTick()`, syncing scroll metrics, and then calling `scrollToBottom(true)`.
+`scrollToBottom(true)` currently performs a single `requestAnimationFrame` measurement and sets
+`scrollTop` from the scroll height available in that frame.
+
+Message rows use `content-visibility: auto` with `contain-intrinsic-size: auto 180px`, and rendered
+message content can continue changing size after the first frame. Markdown blocks, code blocks,
+images, status rows, and input-area layout can all increase the final scroll height after the forced
+scroll has already run. Since no message revision necessarily changes after this late layout settle,
+the existing auto-follow watchers do not perform another corrective scroll.
+
+## Proposed Solution
+
+### 1. Restore `markstream-vue` Tailwind Scanning
+
+- Change the renderer Tailwind source from `dist/tailwind.ts` to `dist/tailwind.js`.
+- Keep the existing `@import 'markstream-vue/index.tailwind.css'` import for package CSS and design
+  variables.
+- Add a focused guard so future package path changes fail loudly. The guard should verify that
+  representative code block class candidates from `markstream-vue` are included in Tailwind's source
+  scanning or generated CSS.
+- Manually verify a rendered code block in light and dark themes after the implementation.
+
+### 2. Settle Bottom Scroll During Session Restore
+
+- Add a dedicated session-restore bottom-scroll helper instead of changing normal streaming
+  auto-follow behavior.
+- The helper should force bottom scroll immediately after session restore and then continue for a
+  short bounded settle window.
+- Recommended implementation:
+  - Use a session-local request id so pending settle work cancels when the user switches sessions
+    again.
+  - Run a small number of animation-frame retries and stop once `scrollHeight` has remained stable
+    for consecutive frames.
+  - Attach a temporary `ResizeObserver` to the scroll area or message root for roughly the first
+    few hundred milliseconds, forcing bottom again when late layout changes arrive.
+  - Disconnect the observer and cancel queued frames once the settle window ends, a spotlight jump is
+    requested, the session changes, or the user intentionally scrolls away.
+- Keep the current near-bottom logic for streaming updates so the previous bottom-shake fix remains
+  intact.
+
+### 3. Force Bottom Scroll After User Submit
+
+- User submit is an explicit intent to continue at the newest message, so submit and command-submit
+  paths should schedule a forced bottom scroll after input state has cleared.
+- This scroll should complement, not replace, the normal message-list watcher. The forced pass makes
+  the watcher robust when `isNearBottom` was stale after session restore or late layout changes.
+
+## Affected Interfaces
+
+- `src/renderer/src/assets/style.css`
+- `src/renderer/src/pages/ChatPage.vue`
+- Potential focused tests under `test/renderer/**`
+
+No main-process, IPC, persisted data, or i18n surfaces are expected to change.
+
+## Compatibility
+
+- The Tailwind path fix should be compatible with the current pnpm-linked package layout because it
+  targets the file that exists in the installed package.
+- Session scroll settling is renderer-only and should not change saved session data.
+- The bounded observer/retry design keeps the existing `content-visibility` optimization in place
+  and limits extra work to the initial restore window.
+
+## Test Strategy
+
+- Add or update a renderer-side guard that fails if `markstream-vue` code block utility candidates
+  are no longer visible to Tailwind scanning.
+- Add a focused `ChatPage` test that simulates session restore followed by a late `scrollHeight`
+  increase without a message revision change, then asserts the scroll position reaches the new
+  bottom.
+- Keep or extend existing streaming auto-follow tests to verify the session-restore helper does not
+  force bottom after the user scrolls away.
+- Add a focused submit test that first records a non-bottom scroll metric, sends a message, and then
+  asserts the submit path still forces bottom scroll.
+- After implementation, run the required project checks: `pnpm run format`, `pnpm run i18n`, and
+  `pnpm run lint`, plus targeted renderer tests.

docs/issues/markdown-codeblock-session-scroll-regressions/spec.md

@@ -0,0 +1,62 @@
+# Markdown Codeblock Session Scroll Regressions
+
+## User Need
+
+Chat markdown rendering should keep code blocks visually readable, and switching sessions should land
+at the actual bottom of the restored conversation. Two regressions currently make the chat view feel
+unfinished:
+
+- Fenced code block chrome, especially the toolbar, renders too compact because the expected
+  `markstream-vue` Tailwind utility classes are not generated.
+- Switching conversations often scrolls close to the bottom but stops slightly short after message
+  content finishes laying out.
+
+## Goals
+
+- Restore the intended `markstream-vue` code block toolbar spacing, background, border, and action
+  button styles.
+- Make session restore scroll to the final settled bottom of the message list when no message
+  spotlight jump is requested.
+- Preserve existing streaming auto-follow behavior, scroll-away behavior, and message rendering
+  performance optimizations.
+
+## Acceptance Criteria
+
+- Generated renderer CSS includes representative `markstream-vue` code block utility candidates,
+  including `py-[var(--ms-inset-panel-y)]`, `px-[var(--ms-inset-panel-x)]`,
+  `p-[var(--ms-action-btn-padding)]`, `bg-[var(--code-header-bg)]`, and
+  `text-[var(--code-action-fg)]`.
+- Code block headers, language labels, copy buttons, and overflow controls render with the intended
+  spacing in light and dark themes.
+- Switching to an existing session without a spotlight target scrolls to the real bottom after
+  markdown, code blocks, images, status rows, and input-area layout settle.
+- Switching sessions does not reintroduce bottom shaking or overscroll during streaming updates.
+- Sending a new message forces the conversation back to the bottom even if the previous bottom
+  proximity metric was stale.
+- If a spotlight target is requested, the message jump remains the winning scroll behavior.
+- User-initiated scroll-away from the bottom is respected after the initial session restore has
+  completed.
+
+## Constraints
+
+- Keep the fix scoped to renderer markdown styling and chat scroll restoration.
+- Keep `markstream-vue` as a package dependency; do not fork or patch the package unless the package
+  path fix proves insufficient.
+- Keep the message row `content-visibility` performance optimization unless a later benchmark shows
+  it is the actual blocker.
+- Use bounded scroll settling so the renderer does not keep observers or animation-frame loops alive
+  after session restore.
+- Do not introduce new runtime dependencies.
+
+## Non-goals
+
+- Redesign the markdown renderer or code block component.
+- Rewrite chat virtualization, message storage, or streaming message flow.
+- Change the composer layout, sticky input behavior, or session loading UX.
+- Add a new user-facing setting for scroll behavior.
+
+## Discussion Points
+
+- The recommended scroll-settling approach is a short `ResizeObserver` window plus bounded animation
+  frame retries. A smaller bounded-rAF-only fix is possible, but it is less robust when late content
+  changes arrive outside the first few frames.

docs/issues/markdown-codeblock-session-scroll-regressions/tasks.md

@@ -0,0 +1,14 @@
+# Tasks
+
+- [x] Investigate the compact code block toolbar regression.
+- [x] Investigate the session-switch bottom scroll regression.
+- [x] Record proposed fixes for discussion before implementation.
+- [x] Confirm the implementation approach with the reviewer.
+- [x] Update the `markstream-vue` Tailwind source path from `dist/tailwind.ts` to
+  `dist/tailwind.js`.
+- [x] Add a focused guard for representative `markstream-vue` code block utility candidates.
+- [x] Add bounded session-restore scroll settling with cancellation and user-scroll guards.
+- [x] Add renderer coverage for late layout growth after session restore.
+- [x] Add renderer coverage for forced bottom scroll after submit.
+- [ ] Manually verify code block rendering and session switching in the app.
+- [x] Run `pnpm run format`, `pnpm run i18n`, `pnpm run lint`, and targeted renderer tests.

docs/issues/pr-base-dev-default/plan.md

@@ -0,0 +1,11 @@
+# Plan
+
+## Approach
+
+- Add an explicit PR base rule to `AGENTS.md` under commit and pull request guidelines.
+- Keep the rule aligned with `docs/release-flow.md`, where `dev` is the long-lived integration
+  branch and `main` is the stable release mirror.
+
+## Validation
+
+- Run formatting and lint checks for documentation consistency.

docs/issues/pr-base-dev-default/spec.md

@@ -0,0 +1,24 @@
+# Default PR Base Branch
+
+## User Need
+
+Contributors and coding agents need a clear repository-level instruction that routine pull requests
+target `dev` by default instead of accidentally targeting `main`.
+
+## Problem
+
+`docs/release-flow.md` defines `dev` as the integration branch and `main` as the release mirror, but
+`AGENTS.md` did not state the default PR base branch in the commit and pull request guidelines.
+Automation can therefore fall back to `main` when creating PRs.
+
+## Acceptance Criteria
+
+- `AGENTS.md` states that routine PRs default to `dev`.
+- `AGENTS.md` states that `main` is only for `release/<version>` PRs following the release flow.
+- The instruction is located in the section that coding agents read before creating commits and PRs.
+
+## Non-goals
+
+- Change the release flow.
+- Change GitHub repository settings.
+- Change CI branch filters.

docs/issues/pr-base-dev-default/tasks.md

@@ -0,0 +1,5 @@
+# Tasks
+
+- [x] Inspect existing branch and release guidance.
+- [x] Add the default PR base rule to `AGENTS.md`.
+- [x] Run required checks.

docs/issues/reasoning-heading-font-size/plan.md

@@ -0,0 +1,27 @@
+# Plan
+
+## Diagnosis
+
+`ThinkContent.vue` renders reasoning content through `markstream-vue` `NodeRenderer`. The component
+sets compact body variables on `.think-prose`, but `markstream-vue` also defines heading-specific
+variables such as `--ms-text-h1`, `--ms-text-h2`, and `--ms-text-h3`. Without overriding those
+heading variables, markdown headings inside thinking content can render larger than the surrounding
+text.
+
+Because `ThinkContent.vue` uses scoped styles and `NodeRenderer` is a child component, selector
+fallbacks that target rendered heading elements must use `:deep(...)`.
+
+## Approach
+
+- Override heading font-size and line-height CSS variables in `.think-prose` so markstream headings
+  inherit the compact thinking body size.
+- Add a scoped `:deep(...)` fallback for rendered `h1` through `h6` and `.heading-node` elements.
+- Add a small source-level guard test to keep the reasoning heading overrides from being removed
+  accidentally.
+
+## Test Strategy
+
+- Run a focused renderer test that verifies `ThinkContent.vue` contains the heading variable
+  overrides and deep heading fallback.
+- Run existing thinking block tests.
+- Run formatting and renderer quality checks.

docs/issues/reasoning-heading-font-size/spec.md

@@ -0,0 +1,20 @@
+# Reasoning Heading Font Size
+
+## User Need
+
+Reasoning/thinking blocks should feel like compact diagnostic text. Markdown heading syntax inside a
+thinking block must not enlarge the text, because model reasoning often uses `#`, `##`, or `###` as
+internal outline markers rather than user-facing document headings.
+
+## Acceptance Criteria
+
+- `h1` through `h6` rendered inside a reasoning/thinking block use the same font size as the rest of
+  the thinking text.
+- The fix is scoped to `ThinkContent` and does not change normal assistant markdown heading styles.
+- The thinking block continues to render markdown, lists, links, and code blocks.
+
+## Non-goals
+
+- Redesign the thinking block.
+- Change how normal assistant message markdown headings are rendered.
+- Disable markdown parsing inside reasoning content.

docs/issues/reasoning-heading-font-size/tasks.md

@@ -0,0 +1,7 @@
+# Tasks
+
+- [x] Diagnose why reasoning headings inherit large markdown heading styles.
+- [x] Document the scoped fix.
+- [x] Override thinking heading font-size and line-height styles.
+- [x] Add a focused style guard test.
+- [x] Run targeted renderer tests and required checks.

docs/issues/stop-pauses-pending-queue/plan.md

@@ -0,0 +1,16 @@
+# Plan
+
+## Approach
+
+- Track sessions whose pending turn queue was paused by an explicit user stop.
+- Set that pause in `cancelGeneration` when a queue drain is active or pending turn input exists.
+- Prevent automatic queue drains for `enqueue` and `completed` while the pause is active.
+- Clear the pause when the user explicitly calls `resumePendingQueue`, when the session is
+  destroyed, and when all pending inputs are gone.
+
+## Test Strategy
+
+- Add a main-process `AgentRuntimePresenter` regression test that starts a queued pending item,
+  makes `processStream` return `aborted`, and verifies the item is released but not immediately
+  claimed again.
+- Keep existing queue and cancellation tests passing.

docs/issues/stop-pauses-pending-queue/spec.md

@@ -0,0 +1,27 @@
+# Stop Pauses Pending Queue
+
+## User Need
+
+When a user stops an active generation, DeepChat must stop the current turn and must not immediately
+continue queued pending inputs. The pending queue should remain visible so the user can resume it
+explicitly.
+
+## Problem
+
+If the active turn was launched from the pending queue, stopping the stream aborts that turn and
+releases the claimed queue item back to `pending`. `drainPendingQueueIfPossible` then sees the
+session is idle and still has pending input, so it automatically drains the same item again.
+
+## Acceptance Criteria
+
+- Stopping an active queued turn releases the queued input back to the waiting lane but does not
+  auto-start it again.
+- Stopping a normal active turn while queued items exist pauses automatic queue draining.
+- Clicking resume queue clears the pause and allows pending items to drain.
+- Destroying or emptying a session clears any stale pause state.
+
+## Non-goals
+
+- Remove the pending queue feature.
+- Change rate-limit provider queues.
+- Change normal stream cancellation behavior when no pending inputs are involved.

docs/issues/stop-pauses-pending-queue/tasks.md

@@ -0,0 +1,7 @@
+# Tasks
+
+- [x] Diagnose repeated restart after stopping a queued turn.
+- [x] Document expected stop and resume behavior.
+- [x] Add pending queue pause state in `AgentRuntimePresenter`.
+- [x] Add regression coverage for stop-paused queue drain.
+- [x] Run focused and required checks.

package.json

@@ -176,7 +176,7 @@
     "katex": "^0.16.47",
     "lint-staged": "^16.4.0",
     "lucide-vue-next": "^0.544.0",
-    "markstream-vue": "1.0.0-rc.0",
+    "markstream-vue": "1.0.1-beta.4",
     "mermaid": "^11.15.0",
     "minimatch": "^10.2.5",
     "monaco-editor": "^0.55.1",
@@ -186,7 +186,7 @@
     "pinia": "^3.0.4",
     "reka-ui": "^2.9.7",
     "simple-git-hooks": "^2.13.1",
-    "stream-monaco": "^0.0.40",
+    "stream-monaco": "^0.0.41",
     "tailwind-merge": "^3.6.0",
     "tailwind-scrollbar-hide": "^4.0.0",
     "tailwindcss": "^4.3.0",