> For the complete documentation index, see [llms.txt](https://docs.hex-rays.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.hex-rays.com/add-ons/assist/release-notes.md).

# Release Notes

{% hint style="success" %}

## Separate release cycle

Assist follows its own release schedule, independent of IDA.
{% endhint %}

## IDA Assist 1.1.0

This release adds per-binary investigations with evidence-backed goals and plans. It also adds IDA approval dialogs for tool calls from external MCP clients, and new query and control-flow cards.

### Investigations, goals, and plans

* Each binary can now hold one investigation. It tracks a goal question, the claims (C1) that answer it, and anchored leads (L1) in five phases.
* The investigation continues across conversations, compaction, and IDA restarts. IDA windows on the same input file share it.
* `/goal <question>` sets the goal and starts a read-only planning turn. After the agent calls `propose_plan`, it waits until you run `/plan approve` or `/plan reject`.
* The composer footer adds **Act**, **Ask**, and **Plan** modes. Ask and Plan turns refuse database edits, file writes, scripts, shell commands, and debugger actions for built-in and external agents.
* A goal strip above the message field shows the goal state, the claim progress, and the next decision. Run `/plan`, click **Board**, or choose **Investigations** in the Assist menu to open **Assist Investigation** in a pane to the right of Assist.
* The agent must cite database evidence or a tool result to settle a claim or lead. Refuted items stay on the board, and a new claim can replace a refuted claim.
* The board groups the evidence of each claim behind per-kind tabs and marks tabs that hold contradicting or stale records.
* An active goal continues after each turn in the conversation where you last sent a message. New goals have no turn or token limit. Set limits with `/goal budget <turns>` and `/goal tokens <count>`.
* Assist blocks a goal after two consecutive turns that record nothing new. After three consecutive failed turns, the goal waits for `/goal resume`.
* The agent can answer a goal only after it settles every claim. It can also report the goal blocked with a reason. `/goal done` finishes a goal immediately.
* `/plan check` and **Check evidence** compare the cited evidence with the live database. `/goal gated <question>` also requires checked evidence and your sign-off.
* Finished goals (G1) stay available through `/goal history`. Use `/goal reopen G<id>` to reopen one, or `/goal keep` to save its established claims as [Memory](https://docs.hex-rays.com/add-ons/assist/concepts/memory.md) entries. Assist anchors each entry to the places that the claim and its supporting evidence cite.
* `/plan export` and `/plan import` transfer an investigation for the same binary. Imported items arrive as unreviewed proposals.
* External MCP clients work on investigations under their own identity, and linked databases can only offer unreviewed leads.
* Investigation writes require the `knowledge_write` permission. If you turn Memory off, investigations are also off.

### Permissions and MCP

* External MCP clients can now see and call tools in [permission groups](https://docs.hex-rays.com/add-ons/assist/concepts/permissions-and-privacy.md) set to **Ask when needed**.
* Each call waits up to 240 seconds for an MCP tool request dialog in the serving IDA window. Choose **Deny**, **Allow once**, or **Allow for this session**.
* **Allow for this session** applies only to that client and permission group, and Assist does not save it to settings. Client connection approval still comes first.
* Without an IDA window, Assist still refuses these calls with no dialog. The same applies to linked databases, fan-out calls, and extensions.

### Memory

* A [Memory](https://docs.hex-rays.com/add-ons/assist/concepts/memory.md) entry can now hold up to 64 anchors, and Assist recalls it from each place. Earlier Assist versions cannot open the upgraded Memory store.
* In the Memory editor, click **Add another place** to add an anchor. For the agent, `remember` accepts an `anchors` list.

### Tool result cards and database mirror

* `idb_query` calls now show as query cards with formatted SQL and a value, a record, a list, or a table. You can sort and filter tables, and a rejected query marks the SQL near the syntax error.
* Structured `get_cfg` results now show as control-flow cards. Function rows open to show loop and branch arcs, exits, callees, and block disassembly.
* Assist now deletes the SQLite mirror of a database when the database closes. To keep mirrors, enable **Keep the database mirror after closing** in **Settings → Behavior**.

### Workspace and conversations

* A saved IDA desktop now restores the Assist panel and Assist Investigation in place. Assist then connects the configured agent, but it does not continue a goal.
* `/fulltriage` and **Triage this binary** now run in the open conversation. Assist saves the findings to Memory when the turn ends.
* Point at a suggestion pill and click **+** to append it below your draft.
* Restored sessions, prompt editing, session titles, and HTML exports no longer show the generated mention-context block.
* **Stop** now idles the panel controls immediately and drops late output from the stopped turn. Assist abandons an external-agent cancellation that stays unacknowledged for 10 seconds.
* A failed turn or a failed model switch now holds queued messages. Notices show missing sign-ins and rate limits separately.
* Transcripts with 128 or more visible rows no longer clip tall rows or leave blank space at the bottom.

### Analysis tools and scripting

* Raw-binary and firmware imports now give the agent `load_facts`, such as segment-to-file mappings, ARM Thumb state, and vector tables. `get_binary_type` returns the same facts.
* `search_strings` now scans raw database bytes when IDA's string list lacks a plain pattern. These rows report `source: "raw_bytes"`.
* In a database without a decompiler, the first decompile call returns disassembly with a warning. Later calls return an error that directs the agent to disassembly.
* `idascript` now answers common IDAPython names and call forms that IDA does not supply. It restores IDA's Python state after each script.

### Models and Codex

* Built-in Codex adds GPT-6.1 Terra (`gpt-6.1-terra`) with a 500,000-token default context window.
* Typed Codex policy, access, model-lookup, and budget rejections now end the turn without retries. HTTP 403 policy refusals no longer ask for sign-in.
* Assist reads a shared Codex `auth.json` file again before it saves refreshed tokens. This keeps a sign-out or a newer login from the Codex CLI.
* The bundled context reference for OpenAI-compatible endpoints is current. It adds the Gemini 4 family on Google's endpoint.

### CPU and process emulation

* [Emulation](https://docs.hex-rays.com/add-ons/assist/concepts/emulation.md) now tracks per-thread hardware instruction breakpoints on ARM64 and x86. Hardware data watchpoints remain unimplemented.
* Windows process emulation adds file creation, synchronous file reads, directory enumeration, and file and volume queries. It also adds file-backed data sections and allocates RAM pages on demand.
* Windows guests can create, suspend, and resume threads, and read and write thread contexts.
* Windows on ARM64 emulation now selects ARM64X and CHPE images. It runs ARM64EC code through paired x86-64 and ARM64 execution.
* Broader Windows application support remains incomplete.

## IDA Assist 1.0.2

Conversation text and tool results now render on a dedicated thread outside IDA's main thread. This release also improves tool recovery and emulation.

### Conversations and tool results

* Conversation text and tool results use the separate render thread for layout and drawing. This includes reasoning and code blocks as well as tables and images.
* Streaming reasoning displays accumulated updates together instead of replaying a backlog of short fragments.
* Symbol and command suggestions stay inside IDA's window and keep focus in the message field.
* <kbd>Esc</kbd> dismisses suggestions without closing Assist. Leaving the active token or clicking elsewhere dismisses the list without inserting anything.
* Filtering and sorting large result tables also run on the render thread. Sorting handles missing values and invalid numbers consistently.
* Text and code views offer **Rendering failed - retry** when preparation fails. The control remains accessible after horizontal scrolling.
* Panels close without waiting for unfinished rendering. Assist also avoids crashes from delayed focus updates after a conversation closes.
* IDA can still pause typing and scrolling while it blocks its main event loop.

### Tools and instructions

* Saving a database through a tool now restores IDA's previous status. Later writes no longer fail because the save left IDA busy.
* Decompilation and byte reads now use bounded retries for transient database changes. Signature searches do the same. Retries stop if the database closes or its identity changes.
* Retrieving a flow chart no longer falsely counts as a database edit that invalidates a read.
* The `expand_tool_result` tool now accepts calls without `offset` or `limit`. Assist starts at the beginning and uses the configured slice size.
* Custom instructions now reach new sessions of the built-in agent intact.

### Settings and definitions

* The API key list now centers the selected key when it needs scrolling. Nearby saved keys stay visible without another scroll.
* Fractional color values now parse correctly under locales that use a decimal comma.
* Very long theme values and struct declarations no longer cause parsing crashes.

### CPU emulation and snapshots

* x86 and ARM64 snapshots now preserve pending CPU events and execution counters. Restoring or forking a session retains that execution state.
* Undefined ARM64 instructions now stop at the faulting address without increasing the completed-instruction count.
* ARM64 exclusive load/store sequences now account for writes through aliased memory mappings.
* Linux ARM64 CPU identification reads now follow Linux's filtered register view.
* Oversized `printf` field widths no longer cause large temporary padding allocations during emulation.
* x86-64 emulation recognizes VEX-encoded packed-string comparisons used by some runtime libraries.
* Some ARM64 code can stop on optional FP/SIMD instructions. Assist does not yet expose a control to enable these features.

### Process emulation

* Windows process emulation captures more registry data used during startup. This includes Command Processor and OLE settings as well as user locale data.
* Access-denied entries under Image File Execution Options no longer abort registry capture. Ordinary guest opens still return access denied.
* Guest programs can enumerate captured registry keys and query several values together.
* Guest programs can open existing private events and request registry notifications.
* Selected Windows runtimes can query token statistics and captured code-integrity policy information. These queries do not enforce signatures on guest code.
* Additional ARM64 startup paths accept Windows build 29698. Windows ARM64 guests can also read the cycle counter.
* Darwin ARM64 emulation preserves pending syscall state across continuations.
* Broader Windows application support remains incomplete.

## IDA Assist 1.0.1

Database mirror queries now tolerate unrelated edits. This update also improves external-agent recovery and extends Windows process emulation.

### Database mirrors

* Mirror queries no longer restart because of unrelated edits. Renaming a function does not invalidate a query over strings.
* Function and segment listings avoid unnecessary live IDA reads after unrelated changes. Import and export listings behave the same way.
* `idb_query` waits for the data it needs within `timeout_ms`. It repeats the query only if that data changes.
* Failure messages identify active analysis or pending capture. They also distinguish repeated edits from unapplied changes.
* Pausing auto-analysis no longer blocks an otherwise current mirror. Adding decompiler helper types also avoids unnecessary refreshes of type declarations.

### Permissions and scripts

* Changing a capability from **Ask** to **Off** invalidates earlier session approvals. Returning it to **Ask** requires fresh approval.
* External MCP calls use updated permissions immediately. The built-in agent adopts updates with its next tool-list refresh. Already admitted work can finish.
* Recursive script requests on the same thread now return an error instead of waiting indefinitely for themselves.

### Process emulation

* Supported native Windows runtimes gain more x64 and WoW64 startup paths. This includes locale data and hardware-exception delivery through installed NTDLL handlers.
* ARM64 support extends additional startup paths to build 29695.
* Native console writes now reach captured output. Emulated programs can open thread handles and request memory prefetching when the selected runtime supports these operations.
* Queries for mapped native libraries can now return retained backing filenames. A descriptive path supplied with image bytes does not establish a backing-file identity.
* Read-only registry snapshots now include primary-user information and selected System policy keys. AppModel data is available in both registry views.
* x86 single-step exceptions now handle repeated string instructions more accurately. The same applies to stack-segment changes and resume flags.
* Broader Windows application support remains incomplete.

### External agents and linked databases

* Local MCP connections can retain their running bridge across compatible Assist updates. Incompatible versions still require bridge replacement.
* Starting multiple IDA instances together on first use no longer causes competing identity creation to prevent MCP startup.
* Local MCP listeners recover after lost sockets or leadership changes. Malformed bridge-negotiation requests no longer terminate IDA.
* Use per-call `target` when concurrent agents share an MCP connection. `use_database` changes the selection for every caller on that connection.

## IDA Assist 1.0.0

Assist 1.0.0 adds manual context limits for custom endpoints and refreshes permissions across active conversations. It also removes blank transcript messages and extends Windows process emulation.

### Model context

* For OpenAI-compatible endpoints with unknown model limits, **Settings → Agent → Context** lets you set the context size in tokens. The initial value is 131,072 tokens, or the value you previously saved for that endpoint and model.
* Assist uses your saved limit for token budgeting and automatic compaction. **Save** applies the selected model's edit, and **Cancel** discards it.
* Known limits take precedence and hide the manual field. This setting does not enlarge the provider's actual context window.

### Permissions and queued changes

* Applying Permissions now refreshes running conversations, including those you are not currently viewing, even when the saved settings already match. Settings changes from another IDA instance also reach running conversations without a reconnect.
* Applying Permissions preserves your separate **Run Python** choice when shell access temporarily enables Python. Disabling shell access restores that choice rather than losing an existing Python grant.
* Requests queued while an approval batch runs remain pending for a later batch. Approving changes cannot execute an already running request a second time.

### Sessions and transcripts

* Some OpenAI-compatible servers send blank text between reasoning and tool calls. Assist no longer shows empty assistant messages for these separators during streaming, session restore, or action-conversation replay.
* The **Sessions** window now follows Settings for fonts, themed backgrounds, buttons, and fields, with more compact spacing.
* Tool headers no longer label the current database as a remote target. Calls to another database retain the target label.

### Windows process emulation

* Exceptions raised through `NtRaiseException` now enter guest exception handling. Second-chance exceptions terminate the guest instead of repeating exception handling.
* Process queries report executable image information, fixed DEP settings, and the absence of a debugger for supported guest profiles. Image queries retain the loaded metadata even if guest code later rewrites the executable's headers.
* Processor queries use emulated CPU capabilities and include WoW64 native-system queries. Startup code can inspect the guest's processor identity and features without using the host CPU's capabilities.
* With supported native runtimes, x86 emulation supplies WoW64 startup data and companion thread state. Access to native runtime files requires permission to read outside the project.
* Selected Windows ARM64 startup calls now accept build 29695 and additional session handle-table layouts. Broader Windows application support remains incomplete.

### Linked databases

* **Allow once** now reaches a linked database, including when approved knowledge sharing waits in the queue. The approval applies only to that call and does not change saved permissions.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://docs.hex-rays.com/add-ons/assist/release-notes.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
