` anchor. Data-image URIs now render inline under a strict allowlist (raster formats plus base64-only SVG, a safe payload charset, and a 2 MB cap), and `` routes through the media pipeline. Every other `data:` scheme (`data:text/html`, etc.) stays inert text and is never embedded; the sanitizer permits `data:` for `![]()
` only, and the streaming renderer enforces the same predicate so an image can't render mid-stream then vanish at settle. Base64 SVG renders via `![]()
` (scripts can't execute in that context). Thanks @ai-ag2026. (#6209)
- **A single configured model no longer appears multiple times in the model picker.** A configured model was registered under its bare id, `provider/model`, AND `@provider:model` forms, so the picker treated equivalent routing ids as separate entries and showed the same model more than once. Dedup now collapses only genuinely-equivalent entries (same normalized id **and** same provider), so two different providers offering the same bare model id (e.g. both offer `gpt-4o`) stay distinct while a truly-one-model-registered-three-ways case collapses to one badge — applied to both the live and static catalog builders. A picked provider badge resolves, sends, and persists as that provider (no snap-back to a first-match provider on a bare-id collision), and a colon-bearing model id (e.g. `model-a:free`) synthesized as a missing-catalog fallback reads its authoritative `data-provider` rather than mis-parsing at the last colon. Thanks @happy5318. (#6221, #3691)
- **Reopening a WebUI tab after a large active session is fast again, even when the session has few messages but large tool outputs.** The reconnect display-path tail optimization only fired based on message *count*, so a session with a handful of messages but multi-MB tool-call JSON forced a slow full-scan merge (bootstrap could take tens of seconds). The optimization now also fires when the session sidecar file exceeds 500 KB, regardless of message count — the same messages are returned (verified: identical message set, count, offset, and truncation metadata), just via the fast path. Thanks @webtecnica. (#6260)
- **OIDC allowlist values containing spaces (e.g. a group named `Hermes Users`) are no longer split into separate entries.** The allowlist parser split comma-separated values on all whitespace, so a multi-word group/user name became two spurious entries. Allowlist parsing now splits only on commas/newlines (multi-word values stay intact and match their exact signed claim — strictly tightening, no security loosening), while OAuth **scope** parsing keeps its space-delimited behavior per RFC 6749 §3.3. A blank allowlist entry no longer bricks an OIDC-only deployment. Thanks @webtecnica. (#6244)
- **`/sessions` and `/resume` typed in the composer now open the session browser instead of being sent as chat text.** Both commands appeared in the slash-command autocomplete (they're exposed by `/api/commands`) but the WebUI send-time dispatch didn't handle them, so they were posted to the agent as plain text. They now open the native WebUI session browser and clear the composer — on desktop and (via the mobile-aware opener) on phone-width layouts. Thanks @webtecnica. (#6245, #6224)
- **Z.AI GLM models now advertise the correct reasoning controls per version.** The WebUI offered the full `reasoning_effort` ladder (max/xhigh/high/medium/low/minimal) for every GLM model, but per Z.AI's API only GLM-5.2+ supports the effort ladder — GLM-4.5 through 5.1 support the binary `thinking` on/off toggle, and GLM-4.7 forces thinking on (not configurable). Reasoning controls are now gated correctly: GLM-5.2+ keeps the full ladder + None, GLM-4.5/4.6/5.0/5.1 get a two-way On/None thinking toggle (no effort levels), and GLM-4.7 shows forced thinking with no toggle. All three surfaces (config status, effort resolution, composer chip) agree across model tiers, provider aliases, and namespaces. Thanks @rh-id. (#6219)
- **Folder downloads now work when the WebUI is served under a subpath.** The file-tree "Download" context-menu action built an absolute `/api/folder/download?...` URL, which broke behind a reverse proxy that mounts the app under a base path (e.g. `/hermes/`). It now resolves the download URL against `document.baseURI` (matching the pattern the other subpath-aware download paths already use), so it works under any base path and is unchanged on root deployments. Thanks @steezypunk. (#6227)
- **Reloading an active turn now prefers the durable run-journal over a stale browser cache, and never truncates recovered tool arguments.** On a hard reload/recovery of a live turn, `loadSession()` could keep a stale local cache as the Worklog authority over the server's newer run-journal snapshot — or, worse, mistake a legacy cache entry with no stream identity for the current stream and suppress replay of a newer run's events. Recovery now selects by stream identity first, then run-journal sequence (`_selectLiveRecoveryInflight`), and carries a run-aware `after_event_id` cursor through the whole reconnect pipeline (server-snapshot hydration, in-memory INFLIGHT, localStorage compaction, session-switch persistence, SSE reconnect), resetting the cursor on stream-ownership change. The durable snapshot also preserves full tool-call arguments (previously truncated to 120 chars) so a recovered command/input is never silently shortened, while a shared content-aware bound (`bound_run_journal_snapshot_args`) caps pathological payloads so the snapshot can't balloon. Fixes #6196. Thanks @franksong2702. (#6197, #6196)
- **The "Compressing context" card no longer flashes on turns that aren't actually compressing.** The card was driven by a broad status match (anything containing `compressing`), so skip/cooldown/defer notices and post-compression summaries (`Compressed X → Y`) could paint a phantom compression card mid-turn. A dedicated `_is_agent_compression_start_status()` now recognizes only the five genuine start-of-compression markers the Hermes agent emits, and the transcript renderer no longer lets an unrecognized `terminal_status` row paint a compression anchor. Real compression turns still show the card unchanged; verified against every current agent compression emitter. Thanks @carlotestor. (#6184)
- **Transparent Stream no longer shows the start of the final answer twice after a multi-segment turn settles.** In a turn with interim assistant prose interleaved with tool calls, the beginning of the final answer could render twice after the live→settled transition — once as a stale live-token accumulator snapshot, once as the real assistant segment — persisting until reload. The earlier suppression gate (#5758) computed its "after the last tool row" boundary over the combined row list, where settled per-message tool rows re-list tools that ran earlier in the turn, pushing the boundary past the final segment's accumulator so the dedupe never fired. Final-segment eligibility is now judged against the live projection's own chronology (a live-prose row belongs to the final segment only when no *projected* tool row follows it), so the stale prefix is dropped while pre-tool narration and every tool/thinking row are preserved. Thanks @ai-ag2026. (#6189, #5749)
- **The frontend now reads the `msg_limit` ceiling from `/api/session` metadata instead of hand-mirroring the constant.** Following the ceiling clamp shipped in exp-v0.52.98, the backend advertises its `_MAX_MSG_LIMIT` as `_msg_limit_max` in every `/api/session` response, and the frontend reads it dynamically (`_msgLimitMax`, a module-scope value defaulting to the static fallback), so the two can no longer drift and the mirrored constant's drift-guard test is no longer needed. Older backends that omit the field fall back to the built-in default, so mixed-version deployments are unaffected. Thanks @webtecnica. (#6214, #6177)
- **`GET /api/session?msg_limit=` is now bounded by a server-side ceiling, and the frontend paginates around it without dropping rows.** A client could request `msg_limit=1000000` (or an outline-jump path asked for `9999`) and force the server to assemble and serialize an unbounded message payload. The backend now clamps `?msg_limit=` to `[1, 500]` via a dedicated `_parse_msg_limit()` helper and sets the existing `_messages_truncated` signal when it clamps; the bare no-`msg_limit` path (branch/undo/jump-to-start) still returns the full transcript. The frontend mirrors the ceiling: `_loadOlderMessages` grows its tail window below the ceiling and switches to bounded `msg_before` backward paging once the server would clamp, and the outline jump uses the bare full-transcript path instead of the old `msg_limit=9999` hack. Two silent row-loss regressions found during review were fixed before ship: a raw-row-heavy `msg_before` page that textually repeated the current tail could be misclassified as the cumulative tail and wholesale-replace it (losing older rows), and a same-session refresh above the ceiling could shrink an already-loaded >500-row transcript to the last 500 — both now route through the correct paging/full-transcript path. Thanks @rh-id. (#6152, #6154, #6177)
- **Background subagents no longer fail to persist when the parent's cached agent is reused mid-turn.** A parent session's `SessionDB` handle is shared by reference with the background subagents it spawns (`delegate_tool`), but the cached-agent reuse path unconditionally closed that handle when the parent got a new turn / server-side wakeup — so every still-running child then failed `append_message` with `'NoneType' object has no attribute 'execute'` (observed 600+ times under one parent). The reuse path now keeps a still-open handle (closing only the unused fresh one, preserving the #1421 FD-leak fix) and only replaces a dead/missing one. The credential self-heal path was hardened the same way: if rebuilding the handle fails it degrades cleanly to a lazy reinit instead of reusing a closed handle. Thanks @carlotestor. (#6143)
- **The CLI/cron session-list projection cache is now bounded.** `_CLI_SESSIONS_CACHE` was an unbounded plain dict of deep-copied session-list projections whose fingerprint-folded key advances on every streamed message, so orphaned heavy copies accumulated on a long-lived server until the next structural clear. It's now an `OrderedDict` capped at 8 entries with drop-oldest LRU eviction (recency refresh on hit), mirroring the existing `_CLAUDE_CODE_PARSE_CACHE` / `_SIDECAR_METADATA_CACHE` pattern; TTL stays the freshness control. Thanks @rh-id. (#6140)
- **Models selected from a list-shaped custom-provider allowlist now route to the right provider.** A `custom_providers[].models` allowlist can be written as either a mapping or a list, and the model picker accepted both — but runtime routing only read mapping keys, so a model chosen from a *list*-shaped allowlist could silently stay on the active/default provider instead of its own. The supported-model-ID extraction is now centralized in one helper used by both catalog construction and routing, so the picker and runtime can no longer drift. Existing mapping-shaped and list-of-strings configs are unaffected. Thanks @franksong2702. (#6127, #6121)
- **Stale sidebar approval attention badge now clears once its gateway approval has drained.** A parent session could keep showing the red approval "attention" dot in the sidebar after the underlying gateway approval was already resolved/gone — the live gateway queue drained but the sidebar summary kept the mirrored approval alive. `_session_attention_summary()` now reconciles the gateway pending-mirror before counting attention (the same self-healing repair the approval-pending route already does), so the badge clears correctly. Live approvals stay visible and actionable and the clarify fallback is unchanged. This is the WebUI read-path mitigation; the child-session context-propagation root cause stays tracked in #6100. Thanks @rodboev. (#6117, #6100)
- **The mobile composer model picker no longer opens off-screen.** On phones the model dropdown could open below the fold (looking like it "won't open", issue #6080) because it lived inside `.composer-footer`, whose `container-type` (and, under the Geist Contrast skin, `backdrop-filter`) made the footer the containing block for the dropdown's fixed positioning — so it anchored to the bottom-of-screen footer instead of the viewport. The picker now reparents to `` while open (the same idiom the profile dropdown already uses), so its fixed positioning resolves against the visual viewport on every skin; it flips above/below the anchor by available space, clamps below the titlebar, and re-positions as the on-screen keyboard / URL bar move the visual viewport. Desktop behavior is unchanged. Thanks @webtecnica. (#6105, #6080)
- **Approval and clarify popups no longer collapse to tiny blank windows on very narrow screens.** On very narrow viewports the clarify card could shrink to an unreadable sliver; it now gets a `min-width` and `overflow:visible` matching the approval card. Thanks @webtecnica. (#6104, #6094)
- **Stream event timestamps no longer break the row layout on narrow foldable screens.** On screens under 360px the per-event timestamps could spill out of tool/thinking card headers; the headers now clip overflow so the timestamp layout holds. (Tool names stay visible — an earlier draft hid them at ≤360px, which lost tool identity; the header clip alone prevents the spill.) Thanks @webtecnica. (#6106, #6099)
- **Streaming sidebar polls no longer re-run the expensive CLI/cron projection every few seconds.** The per-stream cache TTL was pinned at 30s while the frontend's streaming sidebar poll cadence could approach it, so the state.db CLI/cron projection (and its ~200 sidecar reads) could be re-paid on nearly every poll during a live turn (the #4842/#4808 high-CPU `get_cli_sessions` case). The streaming TTL is now held strictly above the sidebar poll interval (45s) and paired with the stable streaming cache key, so repeated polls reuse the projection; a new test asserts the TTL stays above the poll cadence so this can't silently regress. Thanks @starship-s. (#6102, #4842)
- **Delegated subagent sessions in the sidebar now show a meaningful title instead of "Subagent Session".** Generic delegated child sessions previously all displayed the placeholder "Subagent Session" in the conversations list, making them indistinguishable. Their sidebar title is now derived (view-only) from the child's first user message during the existing batched sidebar-metadata pass — without mutating the stored session title, so the read-only boundary is preserved. Only generic "Subagent Session" placeholders are affected; custom titles are left untouched. Thanks @rodboev. (#6056, #6033)
- **Concurrent remote-gateway health checks no longer stampede the gateway.** When many requests needed a remote-gateway health probe at once, each fired its own probe (a thundering herd). Probes are now single-flighted: one "leader" thread probes while latecomers wait (bounded) and share the result, guarded by a single condition/lock with a `finally` that always releases waiters even if the probe errors. Thanks @ai-ag2026. (#5798, #5455)
- **Appending run-journal events no longer gets slower as a session's journal grows.** The next sequence number was recomputed by scanning existing entries on every append (O(n²) over a session's lifetime); it's now cached per journal path (guarded by a dedicated lock, evicted on journal delete), making each append O(1). Thanks @ai-ag2026. (#5799)
- **The login page's connectivity retry timer is now cleared correctly.** The retry poller is started with `setInterval` but was cleared with `clearTimeout`, so once the server came back the interval kept firing (a zombie timer) instead of stopping and reloading once. It now uses `clearInterval`. Thanks @ai-ag2026. (#5801)
- **A background wakeup no longer eats the assistant reply that preceded it.** When a background-process wakeup prompt arrived after an assistant response, the response could disappear from the interactive transcript (it remained in the DB and export). Background wakeups now render as their own distinct "Background wakeup" status row aligned to the message rail — with attachment support — leaving the prior reply intact. Thanks @Isla-Liu. (#5766)
- **Service-tier (fast-mode) support is now derived from the agent's model metadata instead of a hardcoded list.** The service-tier control appears for a model only when the agent reports it supports the fast tier, so newly-supported models light up automatically and unsupported ones don't show a dead control. Falls back safely (no control) when the metadata is absent. Thanks @starship-s. (#5762, #4536)
- **Live reasoning no longer briefly renders twice while a reply streams.** The streaming path updated the reasoning display through two renderers at once (the anchored reasoning block and the live thinking card), which could momentarily show the reasoning twice. It now routes through a single renderer — the live thinking card only updates when the anchored path didn't handle it. Thanks @rodboev. (#5773, #5720)
- **A wide or tall code block in the chat no longer thrashes the surrounding layout while a reply streams.** Code blocks now use CSS `contain: content`, so a growing/overflowing `` is isolated to its own rounded container (with its own horizontal scroll) instead of reflowing the message column during streaming. Thanks @rodboev. (#5770, #5760)
- **Workspace switcher is now navigable with a screen reader.** The workspace switcher exposes proper roles/labels and announces the "New Chat" workspace state transiently, and the closed workspace dropdown is kept out of screen-reader navigation so it isn't read while collapsed. Visual layout and click/keyboard behavior are unchanged. Thanks @rodboev. (#5721, #5671)
- **A background child stream no longer bumps its parent session to the top of the sidebar.** When a child (branched/background) stream was active, its activity re-sorted the parent session's sidebar time-bucket, yanking the parent around while you were working elsewhere. The sidebar sort no longer keys the parent's bucket off child-stream activity, so parents stay put. Thanks @rodboev. (#5729, #5699)
- **The embedded terminal now recovers from a dropped connection instead of silently freezing.** A transport-level failure on the terminal's output stream (session expired, gateway restarted, network drop) fired an EventSource `error` with no handler, so the terminal froze with no feedback. It now shows a "connection lost, reconnecting…" / "disconnected" line, lets the browser auto-reconnect a recoverable connection, and tears down a permanently-closed one so a restart can reconnect — notifying once per outage so a flapping connection can't flood the pane. Thanks @ai-ag2026. (#5786)
- **The OpenCode Go model picker now lists the current flat-rate catalog.** The static `opencode-go` model snapshot was stale; it's been re-synced from the public opencode.ai/go docs + models endpoint (adds MiniMax M3, Kimi K2.7 Code, GLM-5.2, Qwen 3.7 Max/Plus, and more), keeping preview/free-only Zen models out of the Go picker. Thanks @ai-ag2026. (#5761)
- **Passkey login no longer risks a hung or mis-framed response on keep-alive connections.** The passkey-login success `200` was written without a `Content-Length` header, so a client on a persistent connection could wait for a body boundary that never came or mis-frame the next response. The response now sends an explicit `Content-Length` matching the body's byte length. Thanks @ai-ag2026. (#5785)
- **Deleting a session now evicts its run-journal writer locks instead of leaking them.** The per-run-journal writer-lock cache (`_WRITER_LOCKS`) kept an entry after a session's journal was deleted, so the cache grew unbounded, and a later journal reusing the same path could be handed a stale lock. `delete_run_journal` now evicts that session's lock-cache entries — but only after confirming the directory was actually removed, so a Windows/permission transient that leaves files (locks still live) never has its lock wrongly dropped out from under a writer. Thanks @ai-ag2026. (#5784)
- **The background reaper and drain daemon threads can no longer be started twice under a race.** `start_session_channel_reaper` and `start_drain_thread` used a check-then-act (`is_alive()` then create) with no lock, so two concurrent callers could both see "not running" and each spawn a thread — the loser's thread was never referenced by the module and ran forever, un-joinable. Both start paths now perform the liveness check, thread creation, and global assignment atomically under a dedicated lifecycle lock. Stop/join still run outside the lock, so there's no deadlock. Thanks @ai-ag2026. (#5783)
- **Session-lineage reads and the gateway session-list watcher now open the agent's `state.db` read-only.** Reading session lineage/parent-child metadata and the watcher's cheap change-fingerprint opened a *writable* SQLite handle on the live, potentially multi-GB WAL `state.db` while the agent was streaming into it — adding needless checkpoint/lock surface to a pure read. These three read paths now open a read-only (`mode=ro`) connection, falling back to a writable one (with a warning) only if the read-only open fails on an exotic filesystem, and raising rather than letting the fallback create an empty "ghost" `state.db` at a missing path. Same rationale as the #5455 session-listing read-only path. Thanks @ai-ag2026. (#5780, #5455)
- **The MCP servers/tools list now reflects the active profile instead of the default one.** When viewing MCP servers or tools while a non-default profile was active, the list was read from the ambient (default) profile's config rather than the profile you were actually in, so it could show the wrong profile's MCP servers. Both the server-list and tool-list reads are now scoped to the active profile's config home, matching where the MCP write handlers save. Deployments that point config at an external path via `HERMES_CONFIG_PATH` now also read that override consistently on the active-home path. Thanks @rodboev. (#5772, #5619)
- **Mobile/touch: the chat no longer jumps backward while a reply streams and you're scrolled up reading history (non-virtualized transcripts).** Companion to the desktop and virtualized fixes (#5637/#5638/#5742) for the one code path they didn't cover: with transcript virtualization turned off (#4325 opt-out), a re-render rebuilt every row from scratch, and a fresh off-screen tall user row under `content-visibility: auto` (active on coarse-pointer devices) reserved only its flat placeholder height instead of its real height — so `scrollHeight` shrank, the browser clamped `scrollTop`, and the viewport jumped toward the top. The rebuild now reserves each user row's real height, measured from the still-laid-out rows just before the wipe (with a CJK-aware estimate as the backstop for never-painted rows and a `max(remembered, estimate)` floor so a partial paint can't under-reserve). Desktop is unchanged (the placeholder is inert there). Thanks @allenliang2022. (#5751, #5744)
- **Desktop: the reading position no longer jumps up while a streaming reply grows above it.** When you scrolled up to read earlier history during an active streaming reply, each live activity refresh restored a stale absolute scroll offset (captured before the reply grew content above the viewport), yanking the reader up by a few hundred pixels. The desktop restore now holds the reader on the same content by realigning to a surviving above-viewport row's measured offset (the app's own `scrollTop += delta` idiom) rather than the stale absolute value — with a top-spacer-delta fallback when the anchor row is gone, and a raw fallback when neither is measurable. Content growing *below* a scrolled-up reader no longer drifts them down either. Touch/tablet behavior is unchanged (the browser's native scroll anchoring holds there). Thanks @allenliang2022. (#5742, #5637)
- **Cache-hit rate and per-session usage counters now populate correctly in Insights.** WebUI-tracked usage counters (prompt-cache read/write tokens and API-call count) were kept only in the in-memory session sidecar and never mirrored to `state.db`, so the Insights/reporting views that read from `state.db` showed a false 0% cache-hit rate and missing per-session call counts. Those counters are now synced to `state.db` at every usage-sync site. The API-call count uses a monotonic floor (`max(incoming, existing)`) so a rebuilt agent (server restart, session-cache eviction) whose in-memory counter reset to 0 can never clobber the durable running total downward. Thanks @kosta963. (#5463)
- **A normal message no longer gets stuck after the server restarts mid-reply.** If the server restarted or crashed while a reply was streaming, the browser could hold a stale "streaming" state and misroute your next normal send into the steer path, where it failed and left the composer stuck busy. The client now recognizes a proven-dead stream and clears that stale busy state (also clearing the leftover status line, elapsed timer, and streaming badge) so you can send again — while preserving your live turn, draft text, and staged files when steering is merely temporarily unavailable. Thanks @rodboev. (#5723)
- **The message composer no longer hides behind the on-screen keyboard on iPad.** On touch tablets (iPad Safari), the soft keyboard shrinks the visual viewport but not the layout viewport, so the bottom-docked composer was left covered by the keyboard. The composer now lifts to sit just above the keyboard, tracking it as it shows and hides, and drops back to normal when the keyboard closes. It's a no-op on desktop, on devices with a hardware keyboard/trackpad, and while the page is pinch-zoomed. Thanks @rodboev. (#5701)
- **Turns no longer fail with an HTTP 400 from strict providers on empty tool-call history.** Some stored assistant messages carry an empty `tool_calls: []` array, which DeepSeek v4 and newer OpenAI models reject outright ("empty array. Expected an array with minimum length 1"). The model-facing history sanitizer now drops an empty `tool_calls` key on every path that builds the API payload, so a session that accumulated such a message can still send. Populated tool-call chains and their linked results are untouched. Thanks @Swanzb. (#5737)
- **Switching profiles no longer occasionally sticks a session to the wrong provider.** When you switched profiles or tabs while the model dropdown was mid-rebuild, the app could read the *previous* profile's selected model and stamp its provider (e.g. `ollama`) onto a model that doesn't belong to it. That wrong provider got saved to the session and re-sent on every turn, bricking it with a "Provider 'X' is set… but no API key was found" error for a provider the session never used. The provider is now resolved from the dropdown option whose value actually matches the model being sent, so the mismatch can't happen. Credit @b3nw for the root-cause trace. (#5567)
- **Clicking a session link that belongs to another profile now switches to that profile instead of failing.** A valid `session://` deep link pointing at a session owned by a different Hermes profile used to look identical to a deleted session — the UI showed "Session not available in web UI" and self-healed away. Now the WebUI recognizes a valid-but-wrong-profile session, offers to switch to the owning profile, and retries the load once; truly missing/deleted sessions still self-heal as before. No foreign-profile transcript is ever returned — the mismatch response carries only the owning profile name. Thanks @harcek. (#5419)
- **Windows: git operations no longer leave hidden console windows piling up.** On Windows, every workspace git command (`_run_git`) and git-config probe spawned a short-lived `git.exe` whose console window could flash and accumulate as orphaned `conhost.exe` processes over a long session. Those subprocess calls now pass `CREATE_NO_WINDOW` so no window is created. Safe no-op on macOS/Linux, and it uses a small in-module helper (mirroring the update path) rather than a hard dependency on the optional agent package, so a standalone WebUI keeps full git functionality. Thanks @jin89ho. (#5692)
- **The agent's closing summary now shows when it hits the iteration limit.** When a turn exhausted its tool/iteration budget, the agent produces a graceful closing message (a summary of what it got done, or a fallback like "I reached the iteration limit…"), but the WebUI showed a bare `tool_limit_reached` error card instead — most visible with reasoning-only models (extended-thinking Claude, DeepSeek, Codex Responses) whose summary arrives empty. The closing text is now surfaced as a proper assistant message. Purely additive and guarded — it never fires on a turn that already produced a final answer, never double-injects, and degrades to the old behavior when no fallback text is available. Thanks @claw-io (co-authored by @b3nw). (#5717, #5494)
- **The workspace file tree no longer jumps to the top when you expand a folder.** Every folder expand/collapse, breadcrumb navigation, refresh, and hidden-files toggle re-rendered the tree by wiping its container, which collapsed its height and made the browser reset the scroll position to the top — teleporting you away from where you were in a long tree. The scroll position is now captured before the re-render and restored after. Reported by @claw-io. (#5664, #5657)
- **Self-update recovers when a `.git/index.lock` is blocking it.** If "Update Now" failed because a stale `.git/index.lock` was present (a previous git process left it behind), the WebUI gave no in-UI way out — you were stuck on an un-updatable install. The update banner now detects that specific failure, shows the exact `rm -f …/.git/index.lock` command to run on the host, and offers a "Retry update" button that re-runs the normal update once the lock is gone. The server never deletes anything under `.git` itself (fail-closed — git's lock can't be safely auto-cleared without racing), so this is purely a detect-and-guide recovery. Thanks @b3nw. (#5688, #5687)
- **Profile list no longer serves stale rows after a change.** The session-load perf pass (v0.51.908) bumped the profile-list cache TTL to 60s, but profile-row mutations (default model / providers / skills / gateway config) don't invalidate that cache, so a changed profile could show stale details for up to a minute. Reverted the TTL to 4s — frequent enough that mutation-to-refresh staleness is negligible, while rapid dropdown re-opens stay cheap. Thanks @Kopamed. (#5696)
- **Faster session loads on long conversations.** Opening a session with thousands of messages was taking multiple seconds because `Session.compact()` re-walked the whole message list on every response (an O(N) user-message count plus a full reverse scan for the last-message timestamp), and the slow-request watchdog didn't cover the `/api/session` path. The user-count and last-message-timestamp lookups are now bounded (a tail-window scan with an exact full-scan fallback), and slow-request instrumentation covers the session-load path so future regressions are caught. Multi-second loads on large sessions drop to sub-second. Thanks @Kopamed. (#5696, #5455)
- **Sidebar conversation grouping stays stable across refreshes.** Fork / compaction / delegated-subagent clusters could visibly reshuffle on every sidebar refresh — the same group flipping between "N children" and "N prior turns," message counts and nesting jumping around — because the lineage-report cache was keyed only by the group root and reused stale grouping after the authoritative tip or segment count changed. The cache key now includes the authoritative lineage tip and evicts on a segment-count mismatch, and the backend exposes a stable parent-lineage-tip id so grouping is deterministic across live refreshes. Thanks @rodboev. (#5674, #5598)
- **Root fix for mobile scroll jump-back: off-screen tall messages keep their real height.** The underlying cause of the mobile jump-back family (#5637/#5638) was that a virtualization rebuild recreated user rows as fresh elements, discarding `content-visibility:auto`'s remembered size — so a tall message (e.g. a long paste) rebuilt off-screen collapsed to the flat 96px estimate, shrinking `scrollHeight` and forcing the browser to clamp or re-anchor the viewport. Each user row's real height is now remembered (keyed by a stable per-session index) and reserved on rebuild, with a content-length estimate covering rows not yet measured (floored at 96px so short rows never reserve less). Refreshed each measure pass so edits self-heal; inert on desktop. Thanks @allenliang2022. (#5672, #5638)
- **Mobile: reading back in history during a stream no longer drifts backward each tick.** On Android, while a reply was streaming and you'd scrolled up into history, the viewport was nudged backward a few hundred pixels on each streaming tick (a residual of the earlier mobile jump-back). The stale viewport-anchor restore is now refused during streaming when the browser's native scroll-anchoring will hold the position instead — scoped to Android (where overflow-anchor actually compensates), leaving iOS and desktop on their existing, correct paths. Thanks @allenliang2022. (#5666, #5637)
- **Scrolling up into history during a live stream no longer throws you to the top.** While a reply was streaming, a mid-stream re-render (tool completion, activity-scene refresh, clarify echo) wipes and rebuilds the transcript, momentarily collapsing its height so the browser clamps the scroll position to the top. For a reader who had scrolled up into history, that stranded them at the top (a jump of thousands of pixels). The pre-wipe viewport is now restored for an unpinned reader instead of being left clamped; readers following the tail are unaffected. Thanks @allenliang2022. (#5681)
- **No more mid-stream scroll jitter while following a streaming reply.** When you were pinned to the bottom of a live-streaming answer, every mid-stream re-render (tool completion, activity-scene refresh, clarify echo) caused a brief ~1-row up-and-back bounce, so a long streaming reply visibly shuddered. The tail is now re-anchored to the settled bottom in a microtask after the re-render (after layout flushes, before paint), so the short intermediate never reaches the screen. Only a reader already following the tail is affected — an unpinned reader parked up in history is never moved. Thanks @allenliang2022. (#5685)
- **Mobile drawer shows the dashboard link and extension nav-actions again.** On narrow screens (≤800px) the sidebar drawer was hiding the dashboard link and any extension-added nav-action icons, so you couldn't reach them on mobile. They now mirror into the drawer and stay in sync as they appear/disappear. Thanks @rodboev. (#5605, #5583)
- **Inline Mermaid diagrams no longer collapse to zero height on mobile.** An inline diagram could render invisibly on narrow screens before its SVG laid out; it now has a min-height on mobile (scoped so the lightbox and desktop are unaffected). Thanks @nankingjing. (#5560, #5525)
- **Steer file uploads no longer leak their progress across sessions.** When you steer a live turn with an attachment and switch to another conversation mid-upload, the progress bar and composer status now stay scoped to the session that owns the upload (and the steer indicator shows your original text). Thanks @ruizanthony. (#5630)
- **The chat message header scales with your text-size preference.** The model name, icon, TPS badge, and timestamp in the assistant header were ignoring the font-size setting, so they looked disproportionately small at large/xlarge text; they now scale with the body. Thanks @nankingjing. (#5633)
- **The three-panel desktop layout keeps the conversation readable when you resize.** With the sidebar and workspace panel both open, shrinking the window used to crush the center conversation; it now has a readable floor and the side rails yield first (desktop ≥901px; the workspace panel still fully collapses when closed). Thanks @rodboev. (#5594, #5545)
- **Markdown tables render even when a row has trailing whitespace or a small leading indent.** A table whose header or separator row ended in a stray space (something LLMs emit routinely) previously fell through to raw literal pipes instead of a real `