Release Notes.

Improved

• Better retrieval quality for repo-wide code Q&A (code.ask / RAG): expanded retrieval results are now smart-reranked by default using code-structure relationships, keeping code.ask answers focused on the relevant content; incremental refresh scoping is also corrected, so the index refresh range after file changes is more accurate — no more over- or under-refreshing
• Enhanced IOC extraction (sec.ioc_extract): defanged notations such as hxxp://, 1[.]2[.]3[.]4 and user[at]example[.]com are restored before extraction; IPs get validity and classification checks, domains are recognized against public suffixes, and private ranges can optionally be excluded to cut false positives; new indicator types include CVE IDs, Windows registry paths, MAC addresses, SHA-512 and Ethereum / Monero wallet addresses; results carry occurrence counts and type labels with pagination support (consistent with sec.strings), plus optional defanged output and match context

Fixed

• Importing / rebuilding sessions no longer falsely reports “already exists”: fixed misplaced write targets and the dedup / ownership conflicts caused by reusing one session ID across workspaces (including sessions missing after import into a newly created workspace, and blank workspaces falsely reporting duplicates), and fixed the return structure of import conflict detection (previously failing because the Wails binding lost the model type); import / rebuild now completes normally
• Session-export redaction now covers nested JSON: redaction recursively descends into nested JSON structures, covering sensitive data in tool-call arguments and results so deeply nested fields are no longer missed

New

• Plugin marketplace update checks and upgrades: added marketplace sources can be refreshed from origin (git re-pull / local re-parse); after a refresh the view shows each plugin's installed version and an “upgradable” marker, with upgrades available right from the marketplace row or card; installed plugins get a “Check for updates” entry at the top of the list, and cards with updates show an “update available” badge and an upgrade button; upgrades compare by semver first with diff as fallback, the plugin directory is emptied before upgrading so stale components don't linger, and the marketplace cache swaps atomically with safe rollback on failure

Improved

• Real inline progress for marketplace downloads / updates: install / upgrade / preview / update / add operations show an inline progress bar fed by the real download progress parsed from git clone; the update / check-for-updates buttons become an in-place spinning refresh icon for a more consistent look

Fixed

• Session-export redaction now masks IPs at the start of code-block lines: rules are applied to decoded JSON text values, fixing line-leading IPs inside code blocks (affected by JSON escaping) that previously went unmasked
• Extra title bar on Kylin V11 (Wayland) fixed: when Kylin V11 2603 enables the Wayland display protocol by default, the app no longer shows an extra system title bar across the top after launch
• Sending no longer gets stuck when no model is configured: sending right after deleting the model provider in use previously hung; now an inline confirmation in the message flow guides you to switch to an available model first, and a failed start also resets the running state instead of staying stuck at “running”
• Real diagnostics when the child process (zAgent) exits abnormally: when writing the persona or feeding data to a subagent hits a broken pipe, the real zAgent exit reason is reported instead of a generic symptom message, making tricky issues easier to pin down
• Cleaner self-heal: result-type internal injection messages produced by auto self-heal (auto_self_heal) no longer appear in the conversation — what you see is still the normal output
• UI detail fixes: full-window overlays keep their rounded corners, consistent with the frameless window; the agent management area no longer overflows the window width from truncated text

New

• Intent restatement adds “Keep my original intent” plus choice statistics: the confirmation bar gains “Keep my original intent” — skip the AI's restatement and continue with your original message as-is; the count and character volume of each choice (confirm / edit / keep original / timeout auto-start / cancel) are recorded and shown as a table on the settings page; “Analyze intent and restart from here” gains a preview — messages after the restart point are hidden first and the confirmation bar sits below them, making the affected range easy to confirm
• New installer packages for domestic sovereign platforms: dedicated installers for two domestic desktop operating systems — Kylin V10 SP1 (Kylin-Desktop-V10-SP1-2503) and UOS V20 (Uos-Desktop-V20-1060) — join the download page alongside the Windows / macOS / Linux builds, ready to pick directly; links always point to the latest alpha-channel build

Improved

• Config directory and environment variables migrated to avlcode: internal identifiers move from zcode to avlcode across the board (global config directory, workspace .zcode directory, ZCODE_* environment variables and related YAML / frontmatter fields, with SSH remote workspaces handled as well); migration runs automatically at startup / workspace open, keeps dual-read fallback to the old names, shows progress along the way and leaves existing data untouched; leftover “zCode” mentions in the UI and outward-facing output are unified to “AVL Code”
• Real-time status for parallel subtasks (fan-out): the call card for parallel subtasks now shows during execution and reflects each subtask's running state live; spinners stop once stopped, and result blocks are no longer wrongly collapsed at minimal density

New

• Repo-wide code search and Q&A (code.search / code.ask): two new whole-repo semantic retrieval tools beyond the precise “find definition / find references” lookups — ask in natural language “where is the login logic and how does it flow”; purely local, zero vectors / zero embeddings, works offline (air-gap); hits are explainable with file:line plus keywords, structure-aware reranking uses call relationships and centrality, the index auto-refreshes incrementally as files change (with progress shown for first-time indexing), and SSH remote workspaces are supported too
• Clickable code references: code locations cited in AI answers and file:line references in tool results are now clickable, jumping straight to the source at that line — grounded evidence you can see and verify, no manual digging

Improved

• Continued polish for intent restatement: it now triggers only on the first message you typed yourself (automatic / programmatic triggers no longer interrupt), and only when the opening is short (under 30 characters) or fairly simple — say enough, clearly enough, and work starts right away; the confirmation bar's “Edit” can now edit both the goal and the plan, and the message menu adds “Analyze intent and restart from here”
• More polish: a lone .md / .mdx path in a code block (when the file exists) can be clicked to preview; supply-chain (SBOM) settings add a “Reset to defaults” button (changes the form only; takes effect after “Test and save”)

Fixed

• Fixed session-pane crosstalk when creating / switching workspaces (a briefly dangling current workspace wrongly fell back to the first workspace in the list); sessions no longer bleed across workspaces
• Stopping the main session now also stops the subagents it dispatched, leaving no subtasks still running in the background
• Building the code index / starting language services no longer floods the screen with a burst of toasts; progress is now a single in-place progress bar for a tidier interface

New

• Self-repair loop: on errors the AI switches strategy by error type and heals itself instead of grinding down one dead end — upstream errors are routed by category (auth / rate limiting / server 5xx·4xx / safety filtering, etc.), truncated replies auto-continue, circular thinking gets interrupted for a fresh approach, and repeated identical failures escalate the guidance; when a provider stays unavailable, the run can be suspended with a prompt to switch to a fallback provider (confirmation required)
• Parallel subtasks (fan-out): a new parallel fan-out tool splits one task into multiple subagents running at once, each working independently before results are merged — multi-way concurrency beats queuing tasks up one by one
• Closed-loop experience learning: task wrap-up can run an automatic retrospective, banking useful lessons for direct reuse the next time a similar situation comes up (toggle “auto-retrospective” in settings); self-repair also consults past wins and losses to dynamically adjust how many self-heal attempts this run allows
• Self-healing across network flaps: on a momentary disconnect / brief offline spell, the run waits for the network to come back and then resumes, and waiting for the network does not consume the retry budget (an elevator ride or a Wi-Fi blip won't burn through your retries); connection-level blips such as connection resets, transient DNS failures and interrupted handshakes are correctly recognized as network issues instead of a vague “model error”, sparing you pointless checks of the upstream config
• Intent restatement (align on the goal before work starts): when a request is vague, potentially destructive, or the opening is too short, the AI first restates the goal as it understands it and outlines its approach (multi-agent mode adds a phased plan), then waits for your confirmation before acting — no more veering off from the start; confirmation happens inline in the message flow (not a dialog) and auto-starts after 60 seconds by default (hover to pause the countdown), with on-the-spot rewriting before resend or outright cancel (cancel puts your original message back into the input box unchanged — nothing is lost); on by default, can be turned off under Settings → Agent → Introspection
• See the real cost of each round (subagents included): usage statistics now count the real consumption of parallel / subtask subagents; the GoalStrip and each round-end marker show true total token usage plus tool-call counts, persisted per session; every running subtask in the subagent drawer gets its own stop button, so you no longer have to stop the whole round

Improved

• Unified “Attachments” entry: session-attached directories / files and workspace samples / attachments are now managed together in a single “Attachments” bar above the input box, with attach progress and other states visible in real time; removing a session attachment only detaches it and never deletes the original file, while deleting workspace samples / attachments takes an inline double confirmation — clear semantics, hard to delete by mistake
• Safer bulk actions on workspace history: bulk “Open all” becomes “Open selected”, avoiding opening too many at once; past the concurrent-active limit, only the first N are opened and you are told clearly
• More usable “Attachments”: attached files auto-refresh when they change (no manual reload after the AI edits them), directories render as a multi-level tree for a clearer view, and macOS special files (such as .DS_Store) are ignored automatically
• Supply-chain and agent settings tidied: SBOM settings switch to “Test and save” (data-source availability is probed before saving, and offline fields such as the local vulnerability database move into an “Offline databases” area); a new “Agent / Introspection” settings page gathers the intent restatement / auto-retrospective / automatic task recovery toggles in one place

Fixed

• Slash commands now join the input box's up / down history recall (in normalized /<name> <args> form)
• Deleting a companion messaging account now also clears the leftover “cooling down” red bar

New

• Software bill of materials (SBOM) and supply-chain vulnerability screening: sbom.generate recognizes 22 dependency manifest formats (npm / Go / Python / Maven / Rust / .NET and more) to produce SBOMs, with free conversion among the four major formats CycloneDX / SPDX / DSDX / SWID; sbom.audit screens known vulnerabilities against OSV / CVE / KEV, rates them by CVSS v2·v3·v4 and runs offline (air-gap); sbom.vex narrows by reachability and outputs OpenVEX, silencing the “installed but never called” noise
• Long-running tasks recover after a crash: background long-task progress is persisted to disk; after an unexpected exit / crash, relaunch scans for tasks that did not end cleanly and offers a recovery entry — recover manually or enable “auto-recover on startup”

Improved

• Faster, steadier loading for long sessions: very long sessions now load in pages on demand (the most recent stretch shows first, and scrolling up loads earlier messages in segments, with a centered progress hint); combined with underlying index and dedup optimizations, opening and scrolling feel noticeably more responsive

Fixed

• Scrollbar stuck / yanked back to the bottom while a reply was still streaming — dragging up now detaches following immediately, so you can scroll back to read and return to the bottom to resume following when needed
• “Scroll to bottom” no longer takes several clicks (it keeps calibrating until the content height settles, landing at the bottom in one go)
• Rapid repeated clicks on “Send” no longer send duplicates (synchronous in-flight interception — even a burst of clicks sends exactly one)
• Session outline jumps now land correctly even on earlier messages not yet loaded into the window
• First-paint theme flash caused by the security policy (CSP) blocking the theme preload script — fixed

New

• Syntax highlighting and one-click copy for code blocks: code blocks in the conversation are colored by language, labeled with the language in the top-right corner, and carry a “Copy” button to take the whole block in one click
• Unfinished todos trigger automatic follow-through: if todos remain when a round ends, a nudge is appended automatically urging the AI to verify / wrap up, so the list truly gets finished before stopping
• Session outline with one-click jumps: a hoverable “session navigation” control left of the session title pops up the current session's outline — click an entry to jump right to it; the highlight tracks your scrolling in real time (plain grayscale styling)
• In-place Markdown preview: for AI-generated .md files, clicking “Preview” on the “Open” bubble slides out a drawer from the right that renders the file; the drawer is wider and supports “Save as”, so no external app is needed
• Subtasks can now use the workspace's external tools (MCP): dispatched subagents pass through a tool-prefix allowlist to directly inherit and call the external MCP servers you configured, making complex tasks smoother

Fixed

• Subtasks failed to start with “model not found” under multi-provider configurations (the subtask inherited the parent session's model while the provider was always taken as the first in the list — the provider now follows the inherited model, preferring the one that actually offers it)
• Message list jittered / bounced rapidly (some message heights were re-measured over and over — rendered heights, including zero, are now cached; the list stops bouncing and scrolling is steadier)
• Authenticated HTTP MCP returned 401 after a cold start (mismatched keys for encrypted credential read/write led to an empty auth header being sent; fixed — cold starts now connect with credentials, no reconfiguration needed)

New

• Data redaction: before exporting HTML / sessions or before the AI writes files such as .md to disk, secrets and sensitive data are masked automatically by keyword / regex / IP range / domain rules (hits raise a toast; fail-safe throughout)
• Routines gain an “hourly” schedule: one more option beyond daily / weekdays / weekly / manual, firing at a chosen minute past each hour
• Two message-density levels, “normal / minimal”: minimal groups and counts process content such as tool calls and thinking blocks for a tighter view, while reply text stays complete as usual
• Exported / synced .zsession session bundles now include the work records of subtasks (subagents)

Improved

• Full logs kept when the child process crashes: an unexpected crash of the background conversation process (zAgent) captures native exceptions and preserves the pre-crash error output — no more silent “exit code 2”, so tough issues are easier to pin down
• Settings navigation regrouped from 3 into 6 sections: the overstuffed “Services” group is split apart, making individual settings easier to find
• Stuck tools / subtasks no longer wait forever (10-minute idle timeout by default, with the watchdog counting idle time)
• Cleaner upgrade cache (temporary files move into a hidden cache directory, and leftovers are cleaned up automatically)

Fixed

• Plugins still failed to install with signature verification set to “skip” (skills / plugins have two independent verification switches and the unified toggle previously only reached the skill side — it now controls both and self-heals legacy inconsistencies)
• Feedback without attachments failed to submit (an extra empty attachment field was rejected by the server — the field is no longer sent when there is no attachment, so text-only feedback goes through)
• Clicking “Delete” in settings did nothing (the native confirmation dialog was swallowed by the web view — replaced with inline double confirmation)
• The zMCP tool gateway failed to start after a restart (backends now connect asynchronously with capped retries, no longer dragged down by slow / unreachable backends)
• Analyzing high-volume packet captures stalled / froze (pcap_stream_list now paginates its output)

New

• Smarter memory: automatic weighting and ranking by hit frequency / success rate, contradictory memories held for review, a health dashboard, relevance-gated injection and enhanced retrieval; lessons from the AI's self-review are distilled into memory drafts automatically
• Security domain knowledge base sec.ontology: records supply-chain / exposure-surface facts and reasons over reachability (A depends on B, B depends on C → A can reach C)
• Unrestricted-mode shortcut command /unrestricted (Chinese /无限)

Fixed

• Deleting a default MCP backend did nothing (deletion now works)

New

• Self-check adds two new checks, “backfill tests” and “AI self-review” (code changes missing tests get them added automatically, and the AI gives everything a final review at wrap-up)
• “Trace” panel: replay on a timeline which tools an AI run called, in what order, with each one's inputs and outputs
• Self-check: every round gives the code an automatic checkup — tests + build diagnostics + lint + SAST (semgrep); failures flow back in structured form so the AI keeps fixing; blocking or report-only, configurable per work mode, overridable per session

Improved

• Self-check UI aligned with the unified data-table look
• Quota hints now tier by usage and reset time (running low / used up / time until next reset)

New

• Test gate: tests run automatically every round and failures get fixed right away (two-level workspace/session “inherit + override” switches)
• Test-failure feedback tool (test.run: a structured failure list plus a run→inspect→fix→verify loop)
• Python calls automatically honor the workspace virtual environment (venv / conda / poetry / pipenv)
• The AI can enumerate its own tools (tools.list / tools.describe)
• Code intelligence tool group (code.*, LSP-based: symbols / references / definitions / diagnostics / cross-file rename / quick fixes; 22 languages built in, downloaded on demand, works on SSH remotes)

Fixed

• Model IDs containing a slash (such as openrouter/free) got mangled, reporting “model not found”
• Whitespace-only replies were misjudged as “finished”, stopping runs early
• Overlong settings-card descriptions stretched the window wider (now wrap)

New

• OAuth sign-in for remote MCP (OAuth 2.1 + PKCE, automatic token renewal)
• MCP health dashboard (call counts / failure rate / P50·P95 latency)

Improved

• Redesigned MCP configuration UI (data-table look plus permission / dependency disclosure)
• More proactive skill matching (available skills are injected into the system prompt)

Security

• MCP credentials stored with national crypto encryption (VAULT; plaintext credentials migrated passively)

New

• Plugin marketplace (compatible with Claude Code plugins)
• Plugin skills / hooks now truly take effect, active in the workspace by default
• Hook execution visible in the conversation flow
• Connect local MCP servers (stdio channel)

Improved

• Links / paths in your own messages are no longer rendered clickable
• “Clear data” adds a dangerous “clear all workspace files” option

Fixed

• Auto-compaction summaries were falsely timed out on thinking models (switched to an idle watchdog)
• Two timing issues in /goal auto-continuation

New

• Custom providers can add custom request headers

Improved

• Work continues reliably after auto-compaction

Fixed

• Hashes of huge files were truncated and computed wrong (files over 128MiB are now fully covered by streaming)
• “Context overflow” cured at the root (tool-result clamping plus a pre-send estimation guardrail)

Improved

• Feedback can turn off “auto screenshot”, plus small account-page tweaks
• The trial-expired lock screen adds “Quit app”
• The feedback settings page now shows just a “my feedback” list, with new feedback created in a dialog

Fixed

• The first /goal in a clean workspace failed to enter goal mode
• Windows background-process stability cured at the root (liveness probes / startup / orphan-process cleanup)

New

• In-app product announcements
• User feedback (a feedback entry supporting screenshots and attachments)
• Automatic diagnostics when zMCP fails to start

Improved

• Steadier feedback submission and sounder attachments (upload progress / auto retry / dedup / encrypted upload)
• Better memory extraction quality (the Memory Palace adopts ontology-engineering principles)

Fixed

• Setting a goal with /goal in a new session falsely reported “session not found”
• Picking a parameterized command (such as /goal) from the command menu ran it immediately, with no way to fill in arguments
• Commands from a Windows client to a Linux remote workspace always failed
• Multiple info bars above the input box got clipped / overlapped each other when squeezed
• Corrected online documentation link targets

New

• Proactive “suggested tasks” from the AI, split off and started in one click

Improved

• Online documentation entry point

Fixed

• Race conditions in concurrent session-file writes cured at the root
• Floating panels (such as attached paths / suggested tasks) near the right edge expanded off-screen and got cut off

New

• Session-attached directories: give the AI temporary access to folders outside the workspace
• Phone-number sign-in / sign-up with SMS verification codes

Fixed

• Small toggles (such as the on/off switches in the routine list) did not move when flipped, hiding whether they were on
• Holiday easter-egg copy now states its token source clearly
• The file-changes panel mixed in files from other workspaces
• Typing /help with no session open did nothing
• Compaction-exclusion markers were written to disk, leaving sessions permanently “amnesiac”
• The settings sidebar was cut off in short windows

New

• Routines: put the AI to work automatically on a schedule
• Redemption codes on the account page

Fixed

• Full-screen overlays no longer break the window's rounded corners

New

• SSH remote workspaces (zero install on the remote end)

Improved

• History compaction is visible while it runs
• A more compact activity overview, plus instant heatmap tooltips
• Models left disabled are collapsed by default on the model configuration page

Fixed

• Clicking “Save provider” right after editing model parameters dropped unconfirmed changes
• Subagent / background-task execution now shows in real time in the main conversation

Security

• “Close and clear data” disabled for SSH remote workspaces

Fun

• Activity-overview text easter eggs for special days

New

• New “AVL Delta” high-efficiency transfer protocol
• Every call records the protocol and client version used
• Configurable per-pass extraction count for the Memory Palace
• “Remember window size” toggle
• New “wait for background tasks” tool (TaskWait)
• Built-in “VirusTotal AI” file threat scanning

Improved

• An extra fallback layer for subagent type detection
• Alpha upgrades verify login and usage first
• Steadier auto-compaction, with friendly diagnostics when it cannot compact

Fixed

• The real reason now shows when auto-compaction is cancelled
• Clicking “Check for updates” did nothing

Security

• Credential encryption upgraded to national crypto SM4 / SM3

Improved

• Better KV Cache effectiveness — long conversations run faster and cheaper
• Page-wide message buttons show or hide smartly by available width
• Running /clear in an empty session gives a gentle hint
• Sharper, better-targeted messages when calls fail
• Account-page copy and currency-symbol adjustments

Fixed

• On macOS, upgrading failed with a cryptic error when the app was launched directly from Downloads / Desktop

New

• Links and file paths in messages are now clickable
• Workspace “project instructions” (AGENT.md) injected automatically
• Built-in AGENTS.md editor

Fixed

• The default window was cut off on high-DPI / scaled displays

New

• Model-selection scope switch (global / follow the session)
• Workspace bindings show two-way “pending message” counts, plus a parking queue for inbound messages
• “Unrestricted mode” switch

Fixed

• Switching a session to a stricter model (such as Landi N2.5 → N2.5-Flash) instantly returned 400
• WeChat replies were silently lost on errcode=-2 (expired context_token)
• The confirmation-dialog overlay squared off the window's rounded corners

New

• The header command palette can search second-level settings under “Manage agents”
• Input history recall gains position labels; editing turns an entry into the latest draft

Fixed

• The Memory Palace's default tour path bar showed room names as English slugs
• Newly created workspace personas did not show up
• Anthropic-compatible providers self-heal a misconfigured Base URL / Type, with sharper error diagnostics

New

• “Result draft”: collect the useful fragments of a conversation into one deliverable document
• “Start a new session from here”

Improved

• /compact summaries keep their length in check automatically

Fixed

• The 30-second “false timeout” on tool calls cured at the root
• Deleting a defunct MCP backend was blocked by a runtime 400

Fixed

• False “trial expired” verdicts

Improved

• Release dates no longer guess the year from the version number
• Reworked time injection in the system prompt (optimal for prefix caching)

New

• Session goals advance continuously on their own
• Instant guidance (GuidanceStrip)
• Model settings page: operations by provider

Fixed

• The LLM occasionally wrote tool calls in the wrong place — two built-in fault-tolerance paths added
• Strict provider gateways silently swallowed “invalid tool name” 400s into an empty stream
• Ambiguous “this week / this month” labels in the day-bucketed session list / sub-session history

Security

• MCP management (the mcp. tool group)