mcp-windbg/CHANGELOG.md

Changelog

All notable changes to the MCP Server for WinDbg Crash Analysis project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.5.1] - 2026-10-10

Fixed

• open_cdb_remote against a debug server whose target is running no longer waits out its whole timeout and fails with "Debugger initialization timed out". A running target reads no commands, so the prompt probe the open relied on could never be answered. The client now waits for its connect banner, which is where connection latency belongs, and then gives the target a short window to answer; silence means it is running, and the session opens in that state. The initial triage then breaks in, as any command on a running target does, the result says so under Target State, and the target stays stopped until it is resumed or the session is closed. A failed connect still reports the debugger's own error, and a connect that never completes now says so instead of blaming initialization (#130, item 2; the register-context approach was proposed in #127). Thanks to @xiaozhu1337.

Added

• A standalone mcp-windbg.exe, for anyone who wants the server without a Python toolchain. PyInstaller freezes the interpreter and every dependency into one file; cdb.exe remains a prerequisite. Importing cleanly does not prove a frozen build works - a dependency that imports lazily or reads a data file at runtime fails at the tool call, not at startup - so MCP_WINDBG_SERVER_EXE points the e2e harness at a built binary and the whole scenario suite re-runs against it.
• The executable is published on each release, code signed, with a SHA256SUMS file taken from the signed binary. Before the asset is attached, the signature is verified and the scenario suite runs once more against the signed binary, since signing rewrites the executable.
• Windows file metadata on the executable: a version resource generated from pyproject.toml, an application icon, and a note that the binary is authored by Sven Scharmentke and signed with a certificate sponsored by Liasoft GmbH, which is the publisher Windows shows.

[1.5.0] - 2026-10-09

Fixed

• A failed open_* no longer leaves behind a debugger session the caller cannot address. All four open tools own their debugger until initialization, triage and output filtering have all succeeded; on a failure or a cancelled request that call's session is rolled back and shut down, and unrelated sessions are left alone. A rollback never resumes a live target: an open that never returned a session id should leave the target as it found it, and resuming stays an explicit choice through close_kd_session's resume (#125, #136). Thanks to @xiaozhu1337 for the diagnosis, the reproduction and the design.
• Completion markers are now unique per session. They were numbered from a counter that restarts at 1 in every session, so several clients attached to one shared -remote debug server could complete each other's commands, and ..._1 also matched inside ..._10. Each session now carries a random nonce and the match is anchored at the end of the line (#137). Thanks to @xiaozhu1337 for finding it.
• A command that outruns its timeout on a dump session is now cancelled. Only live targets were broken into before, so on a dump the debugger kept executing the command while every later command queued behind it and timed out as well, wedging the session until it was closed. That is easy to reach now that open_kd_dump exists, since a !process 0 7 over a large kernel dump runs for minutes. If the cancel cannot resynchronize, a live target keeps its session and the manual break-in hint, because send_ctrl_break can still rescue it, while a dump session is closed rather than left wedged (#138). Thanks to @robster7674 for the analysis and the fix.
• The output retained for one command is now bounded, by line count and per line, with a notice saying how it was cut. The transport keeps being drained past the bound, so the command still completes and the session stays usable (#130, #139).
• .logopen /u quotes its path. Given a path containing a space the debugger opened a log at the truncated prefix instead, left that file behind and never created the intended one, which silently disabled the Unicode-log output channel that a multibyte code page depends on. Reachable whenever the resolved temp directory contains a space (#139).
• A close_* that cannot shut its debugger down reports the failure instead of success, so a debugger still holding its target is visible. A failed rollback no longer replaces the error explaining why the open failed (#139).
• --verbose writes to stderr rather than stdout, which on the stdio transport carries the JSON-RPC stream, and it now installs a logging handler at all, so the server's and a --filter-script's own messages actually appear (#139).
• Dump discovery runs off the event loop, so a glob over a large or remote share no longer blocks a break-in on another session. The order is unchanged (#139).

[1.4.0] - 2026-10-01

Added

• --init-command (repeatable), or MCP_WINDBG_INIT_COMMANDS with one command per line, runs debugger commands on every new session before triage - for example .load for a debugger extension. Their output leads the open result under Initialization. --kernel-init-command / MCP_WINDBG_KERNEL_INIT_COMMANDS do the same for kernel targets only: kd sessions and kernel dumps, which are recognized by their file header even when opened with open_cdb_dump (#121). Thanks to @RamanaIntel for the feature.
• An open_kd_dump tool that opens a kernel crash dump with kd.exe (-z): a complete, kernel, or bitmap memory dump (MEMORY.DMP) or a small memory dump from C:\Windows\Minidump. It runs the init commands, then vertarget and !analyze -v, and returns a kd session id for run_kd_command and close_kd_session (#111). Thanks to @robster7674 for the proposal and the analysis behind it, and to @RamanaIntel for testing it on a 66 GB dump.

[1.3.1] - 2026-09-29

Fixed

• Debugger startup, commands, and shutdown run on worker threads so slow I/O does not block other MCP calls. Break-in remains available even when the debugger worker pool is full. Concurrent close calls claim a session once and preserve the closing caller's kernel resume policy. Break-in keeps the validated session record so a concurrent close cannot cause a registry error (#109, #110). Thanks to @adity982 for the fix.
• A command that times out before reaching its completion marker keeps the output already read, so the timeout error still shows useful diagnostics from runaway commands such as SOS !clrstack. The retained output is capped at 2,000 lines and 64 KiB (#112, #114). Thanks to @Oscar-Williams for the fix.
• A debugger process that exits (q, a failed -remote or -k connect, or cdb/kd crashing) is now reported at once with its exit code and last output, instead of every pending command, wait_for_break, or session open sitting out its full timeout and then misreporting a timeout (#115). Thanks to @RamanaIntel for the fix.
• Debugger errors (timeouts, a debugger that exited, failed connects) reach the caller as their message alone, without a Python stack trace. The trace repeated the message, so a timeout with partial output returned that output twice.

[1.3.0] - 2026-09-10

Added

• A mcp-windbg-skills Claude Code plugin with four optional debugging skills for the uvx plugin or independently configured MCP servers, without a bundled server or agent.
• A mcp-windbg-agents Claude Code plugin with the optional crash-analyst agent, using an existing MCP connection without a bundled server or skills.

Changed

• The uvx plugin now ships only the MCP server configuration. Install mcp-windbg-skills@mcp-windbg and/or mcp-windbg-agents@mcp-windbg separately to keep those workflows. Skills use /mcp-windbg-skills: instead of /mcp-windbg:, and the agent is now mcp-windbg-agents:crash-analyst.
• Skills resolve tools from the configured server; setup diagnosis checks its actual launcher and host instead of requiring uv for every installation.
• Release validation checks the version of every marketplace entry.
• README and setup documentation explain all three plugins, their workflow names, switching and updating, and the separate built-in MCP prompts.

[1.2.2] - 2026-09-05

Fixed

• Unicode command output on Chinese, Japanese and Korean Windows (#102) - on a multibyte system code page (936, 932, 949, or the "Use Unicode UTF-8" setting 65001) the debugger truncates its text output over a pipe, dropping the tail of any line that contains non-ASCII text; a du of a Chinese string came back empty or partial, other Unicode output was cut short, and a line split in the middle of a character could leave the session unresponsive. On these code pages the session now mirrors output to a UTF-16 log and reads each command's output from it, keyed by the markers it already uses to synchronize; the result is complete for every command. Single-byte code pages (such as Western 1252), where the pipe is lossless, are detected and left on the pipe path unchanged.

[1.2.1] - 2026-08-27

Changed

• The README install section is organised by client, not by install method - Claude Code has its own heading, with the plugin and a manual claude mcp add beneath it, and other clients follow. Python is no longer listed as a prerequisite, because it is not one for the plugin route: each route now states what it needs next to the command it applies to. The note about enterprise policy moves to the top, where it can still change whether you proceed.

Fixed

• A corrupted path in the README example prompts - C:\dumps\app.dmp shipped with a literal BEL byte where the \a should have been, so anyone copying that line got a control character instead of a path. Every markdown, workflow and JSON file in the repository was swept for stray control characters; this was the only one.

[1.2.0] - 2026-08-27

Added

• A Claude Code plugin, and a marketplace to serve it from this repo - installing is now two lines with no pip install and no MCP configuration to hand-edit:

  /plugin marketplace add svnscha/mcp-windbg
  /plugin install mcp-windbg-uvx@mcp-windbg
  

  The plugin launches the server with uvx, which fetches the pinned version from PyPI on first use, so the only prerequisites are Windows, CDB, and uv.

  It ships the ten tools, four skills - analyze-dump, debug-remote, kernel-debug, and windbg-doctor for diagnosing a machine that cannot debug - and a crash-analyst agent that investigates a dump on its own and returns a verdict with its evidence, what it ruled out, and what to do next. Always-on context cost is about 400 tokens; a skill or agent costs more only when it fires.

  Symbols work out of the box: the plugin supplies the Microsoft symbol server as a default for _NT_SYMBOL_PATH, and an existing value wins over it, so a symbol path you tuned yourself is never overwritten.

Changed

• Dependency floors raised to current releases - mcp 2.1.1, starlette 1.6.0, uvicorn 0.52.4, and pymdown-extensions 11.0.2 for the docs build. The major-version caps are unchanged, and uv.lock is regenerated to match; it had been left at 1.0.1 since the 1.1.0 release, because release-please has no updater for it.

Fixed

• The crash-analyst agent no longer hard-codes the plugin's installed name - its tools: list named MCP tools as mcp__plugin_mcp-windbg_mcp-windbg__*, and that middle segment is the name the plugin is installed under. Under any marketplace entry not called exactly mcp-windbg the list would have matched nothing, leaving the agent with no debugger tools at all. It now denies the mutating tools instead, which keeps it read-only without naming the plugin.

• The version-consistency check reports mismatches instead of crashing on them - it looped with foreach ($error in $errors), and $error is a read-only PowerShell automatic variable, so the moment it had something to report it threw "Cannot overwrite variable Error" and lost the diagnostic. The build still failed, but never said which files disagreed. It now also covers the plugin manifest, the marketplace entry, and the server version pinned in the plugin's .mcp.json.

[1.1.0] - 2026-08-22

Changed

• Runs on the mcp 2.x SDK - the requirement is now mcp>=2.0.0, up from the <2.0.0 cap 1.0.1 shipped as a stopgap. 2.0.0 removed the low-level @server.list_tools() / @server.call_tool() decorators this server registered its tools through, so the handlers are now passed to Server(...) directly and return the SDK's result types (ListToolsResult, CallToolResult) instead of bare lists. McpError is MCPError, and it takes a code and a message rather than an ErrorData. None of this is visible to clients: the same nine tools with the same schemas, over both stdio and streamable-http (#78).

  Installing mcp-windbg pulls the SDK it needs, so there is nothing to do on upgrade. Python support is unchanged - the 2.x SDK requires 3.10+, exactly as this project already did.

• Every runtime dependency is capped at the next major - mcp<3.0.0, pydantic<3.0.0, starlette<2.0.0, uvicorn<1.0.0. #76 happened because mcp>=1.28.1 had no upper bound, so the 2.0.0 release landed in fresh installs and broke them at import; moving the floor to 2.x without a ceiling would have left the identical trap for 3.0.0. mcp-windbg is an application rather than a library, so an upper bound cannot cause a diamond conflict for anyone, and it turns the next breaking SDK release into a Dependabot PR that fails CI instead of an install everybody has to work around by hand.

• A weekly canary tests the versions users will actually get - CI installs from uv.lock, so it only ever proved the server works against the exact versions pinned there, while a PyPI install resolves to whatever the ranges allow. dependency-canary.yml closes that gap: it resolves to the newest allowed versions, ignoring the lock, and runs the hermetic suite against them every Monday. It is scheduled rather than attached to pull requests on purpose - an upstream release breaking the build should page the maintainer, not block someone's unrelated PR.

Added

• wait_for_break - block until a target you resumed stops again (bugcheck, breakpoint, or a CTRL+BREAK from elsewhere) and return everything the debugger printed when it did. If the wait expires the target is left running, so waiting never halts a machine behind your back.

Fixed

• g no longer freezes the target it was supposed to release - go-class commands (g, gh, gn, gN, gc, gu) hand the CPU back to the target, after which the debugger stops reading its stdin. The marker protocol queued an .echo the debugger could not answer, so the command hit its timeout and the cancel-on-timeout CTRL+BREAK halted the target again - the exact opposite of what was asked. They are now written bare and return immediately. The step family (p, t, pa, ta, ...) is unaffected: it returns to the prompt on its own and keeps the marker round-trip (#74).
• An ordinary command after g breaks in by itself - no need to interleave send_ctrl_break manually, and whatever the target printed on the way to stopping leads that command's output instead of being discarded.
• A break-in issued the instant g returns is no longer swallowed - resuming wrote the g and returned without waiting for the debugger to read it, so for a moment the debugger was still sitting at its prompt with the resume unread. A CTRL+BREAK arriving in that window was answered by the prompt instead of reaching the target, and the break was simply lost: the caller then waited out a full wait_for_break timeout on a machine nothing was going to stop. Measured against a live KDNET target, a break sent with no gap after g was lost every single time. The resume is now confirmed consumed before it is reported. A go-class command that stops again at once - a gu returning in microseconds, a breakpoint hit immediately - reports what it printed instead of claiming the target is running.
• Break-in asks before it signals - a CTRL+BREAK aimed at a target that has in fact already stopped queues a break request a kernel target honours later, re-halting a machine we thought we had released. The session probes for a prompt first and only signals if that goes unanswered, which also makes send_ctrl_break safe to issue speculatively.
• wait_for_break trusts the debugger, not a flag - it is right about a target running for reasons this server never caused, and a second g while the target really is still running is refused rather than queued behind the first.
• bp nt!X; g reports whether the breakpoint was set - the part before the g runs with its output returned, so a typo'd symbol surfaces as Couldn't resolve error rather than as a wait for a break that can never arrive. A ; inside a quoted command string is left alone.
• Output is no longer lost at a deadline - a marker that lands in the moment between a timeout expiring and being handled now counts as arrived, and the break-in output a timed-out command was carrying rides along on the error instead of vanishing with it.
• One operation at a time per session - wait_for_break parks for minutes on a worker thread so the rest of the server keeps answering; a second call against the same session is refused at once, with a message pointing at send_ctrl_break, rather than interleaving markers and returning each other's output. The refusal never waits, so it cannot stall the server it exists to keep responsive.
• Closing a session ends a wait parked on it - instead of leaving a thread sitting out its full timeout on a debugger that no longer exists and then reporting the target as still running.
• Kernel sessions over a named pipe or serial cable connect - kd announces a COM/pipe link with Kernel Debugger connection established, not the KDNET Connected to target .... Only the latter was matched, so open_kd_session timed out on a target that was in fact attached (#47, #74).

[1.0.1] - 2026-08-21

A single-line dependency fix, but an important one: every fresh install of 1.0.0 was broken.

Fixed

• Installs no longer pull an incompatible mcp - the runtime requirement is now mcp>=1.28.1,<2.0.0. mcp 2.0.0 renames McpError to MCPError and drops the low-level Server.list_tools() decorator this server registers its tools with, so an unpinned install resolved to 2.x and died at import with ImportError: cannot import name 'McpError' from 'mcp.shared.exceptions' (#76). Thanks to @aphroteus for the fix, and to @arjunarjun07 for pinning down the cause.

  If you hit this on 1.0.0, upgrading is the fix: pip install --upgrade mcp-windbg. The manual workaround (pip install "mcp<2") is no longer needed.

  Moving to the 2.x server API is tracked in #78.

[1.0.0] - 2026-07-16

First stable release, and the one where kernel debugging actually works.

The tool surface is redesigned around opaque session ids and split by debugger engine: user-mode (cdb.exe) and kernel (kd.exe). This is a breaking change for clients of the 0.x tools. Every *_windbg_* tool has a new name, listed under Changed.

Added

• Kernel debugging - open_kd_session, run_kd_command, and close_kd_session attach to a kernel target over KDNET (net:port=,key=), a named pipe (com:pipe,port=\\.\pipe\...), or serial, driven by kd.exe. This was impossible in 0.x: every remote connection was launched with -remote, which is user-mode only, so a kernel target could never connect (#62, #47).
• Kernel sessions arrive already stopped - open_kd_session waits for the target's connect banner and breaks in for you, so there is no separate break-in step.
• Closing a kernel session sets the machine running again - close_kd_session sends g by default (resume: true). Pass resume: false only to leave the target halted on purpose.
• --kd-path - point the server at a specific kd.exe, the counterpart to --cdb-path. Kernel sessions need their own option because cdb.exe cannot drive a kernel cable.
• Session ids - every open_* returns an opaque session_id (cdb-… / kd-…) on its first output line, and run_*, close_*, and send_ctrl_break address a session by that id. The kind is enforced: a cdb id passed to run_kd_command (or the reverse) returns an error naming the right tool.
• Per-call timeouts - timeout_seconds on any open_* / run_* overrides that tool's default (open_cdb_dump 180s, connects 60s, run_cdb_command 60s, run_kd_command 120s). The server-wide --timeout (default 60s) is the floor.
• remote-triage prompt - a guided investigation of a live user-mode target: break in, orient, then track down the hang or crash.
• kernel-triage prompt - the same for a kernel target, including telling a real bugcheck apart from a plain break-in, and releasing the machine at the end.

Changed

• BREAKING - list_windbg_dumps is now list_dumps.
• BREAKING - open_windbg_dump is now open_cdb_dump.
• BREAKING - open_windbg_remote is now open_cdb_remote.
• BREAKING - run_windbg_cmd is now run_cdb_command (user mode) or run_kd_command (kernel). The connection_type discriminator is gone; the tool name says which engine.
• BREAKING - close_windbg_dump and close_windbg_remote are both now close_cdb_session.
• BREAKING - send_ctrl_break takes a session_id.
• BREAKING - sessions are addressed by id, not by dump path or connection string, and run_* no longer opens one implicitly: open a session first. Reopening the same dump gives you an independent session instead of silently reusing one.
• Module layout - kernel sessions live in kd_session.py and user-mode in cdb_session.py, over a shared debug_session.py that owns the subprocess and marker protocol.
• Docs - README and the docs site rewritten around the new tools and the session-id flow, with a "Debug a kernel target" guide, a Development page, and client guides for Claude Code and Autohand Code (#63). Setup instructions standardize on pip install mcp-windbg and python -m mcp_windbg.
• Dependencies - runtime floors refreshed (mcp>=1.28.1, pydantic>=2.13.4, starlette>=1.3.1, uvicorn>=0.51.0), plus pytest>=9.1.1, and uv.lock regenerated. The streamable-http transport was smoke-tested against the Starlette 0.x to 1.x bump.

Fixed

• A slow command no longer wedges a live session. When a command outruns its timeout the debugger is still busy running it, so the server now breaks in with CTRL+BREAK and resynchronizes before reporting the timeout. The session is usable again for the next command instead of stranded.
• A slow command's output is no longer attributed to the next command. Every command waits on its own unique completion marker, so late output from one command can never be mistaken for the next one's result.
• Debugger processes are cleaned up on close. cdb.exe launched through the Microsoft Store execution aliases spawns a child that a plain terminate left behind, still holding the target. Shutdown now does a Windows process-tree kill.

Removed

• All *_windbg_* tool names, and the connection_type parameter on run_windbg_cmd and send_ctrl_break, superseded by the cdb / kd split.
• Session reuse by dump path or connection string. open_* always creates a fresh session and returns its id.

[0.15.0] - 2026-06-08

Added

• Documentation site: New MkDocs (Material) user guide under docs/, deployed to GitHub Pages via pages.yml. Covers getting started, the use cases, and a reference for the command-line options, tools, and client configuration. Content migrated and trimmed from the project wiki.
• Prompts reference: Documented the built-in dump-triage MCP prompt and its dump_path argument (docs/reference/prompts.md).
• Usage guide coverage: New use-case pages for running the server over HTTP (Debug from another machine) and scrubbing tool output (Redact sensitive data), plus WER auto-capture setup in the triage guide and the dump-triage prompt in the crash-dump guide. Documented the HTTP transport's lack of authentication, that attach-by-PID is unsupported, and that sessions are concurrent.

Changed

• Package metadata: Filled in distribution metadata so pip show / PyPI are complete - added the author/maintainer email, project URLs (Homepage, Repository, Issues, Changelog), and classifiers for Windows, console environment, and the Debuggers/QA topics (#36)
• Contributor guide: Migrated AGENTS.md to CLAUDE.md and added .claude/rules/ (Markdown typography and documentation authoring), plus scripts/Format-Docs.ps1 to enforce the typography rules
• Docs tooling: Bumped the docs build dependencies in requirements-docs.txt to their latest patch floors - mkdocs>=1.6.1, mkdocs-material>=9.7.6, pymdown-extensions>=10.21.3 (consolidates #51, #52, #53)
• Test suite: Replaced the ad-hoc tests with a declarative end-to-end harness. Each tests/scenarios/*.yaml runs against a really-hosted python -m mcp_windbg server driven by a real MCP client (only the LLM is faked). Live scenarios carry live/remote markers and skip cleanly without CDB, so -m "not live" runs anywhere. Adds pyyaml as a dev dependency.
• Coverage: The harness hosts the server under coverage run --parallel-mode (set MCP_WINDBG_COVERAGE), so coverage reflects the subprocess where tool dispatch runs rather than the test process. The suite reaches 90%+; CI enforces a floor with coverage report --fail-under=88 (margin below the ~91% actual for cross-version and cdb-output jitter). Code that cannot be line-measured end-to-end (the streamable-http transport on Windows, atexit cleanup, debug-only and defensive branches) is excluded with documented # pragma: no cover and exclude_also rules. Adds coverage as a dev dependency.
• LFS dumps are mandatory: a scenario whose requires.dump file is missing now hard-fails (run git lfs pull) instead of skipping, so a half-set-up checkout is a loud error rather than silent green.

Removed

• Dead code: Removed three unused functions surfaced while raising test coverage - server.execute_common_analysis_commands, CDBSession.get_session_id, and prompts.get_available_prompts. None had callers.

Fixed

• Stdio server resilience: The stdio transport no longer crashes on a malformed input line. serve() now uses the SDK default raise_exceptions=False, so an unparseable line (e.g. when the server is run directly in a terminal) is logged instead of tearing down the whole process (#45)

[0.14.0] - 2026-03-21

Added

• Tool Content Filter Script Hooks: Added --filter-script so trusted Python helpers can rewrite string-valued tool arguments and tool text output for use cases like PII redaction without exposing full MCP protocol messages

[0.13.0] - 2026-03-18

Added

• Live Debugger Break-In: Added the send_ctrl_break tool to interrupt an active CDB/WinDbg session with CTRL+BREAK for dump and remote debugging workflows (#40)

Changed

• Dependency Refresh: Updated runtime dependency floors for mcp, pydantic, starlette, and uvicorn, and refreshed test and validation tooling versions in pyproject.toml
• CI Dependency Maintenance: Updated GitHub Actions dependencies for Python setup and artifact handling in release workflows (#39, #42)

Fixed

• Registry Compatibility: Restored MCP registry compatibility by reverting server.json to the supported 2025-10-17 schema version
• Publishing Workflow: Adjusted MCP publishing workflows to match current registry publisher behavior

[0.12.2] - 2025-12-15

Fixed

• Registry Schema Migration: Updated MCP server schema from deprecated 2025-10-17 to current 2025-12-11 version for mcp-publisher compatibility

[0.12.1] - 2025-12-15

Added

• HTTP Transport in Registry: Added streamable-http transport configuration to server.json for MCP registry discovery
• Schema Validation in CI: New validate-server-schema.py script validates server.json against the official MCP schema

Fixed

• Registry Schema Update: Updated MCP server schema version from 2025-09-29 to 2025-10-17 for compatibility with registry.modelcontextprotocol.io
• CI Cache Warning: Disabled unnecessary dependency caching in PyPI publish job to eliminate spurious warnings

[0.12.0] - 2025-12-15

Added

• HTTP Transport Support: New --transport streamable-http option enables HTTP-based communication alongside the default stdio transport (#31)
• MCP Prompt API: Implemented prompt templates for AI-assisted crash dump triage and analysis (#25)

Changed

• Updated Dependencies: Bumped mcp to 1.17.0, pydantic to 2.12.0, and other dependencies (#26)
• Improved Prompt Templates: Removed hard-coded model references from prompt templates for better flexibility (#29)
• Updated Dependabot Configuration: Improved automated dependency update settings

Fixed

• Session Cleanup: Prevent stale debugging sessions if .shutdown() fails (#28)

[0.10.0] - 2025-10-10

What's New in This Release

This release focuses on making mcp-windbg more reliable, faster, and easier to use for everyone - from beginners to advanced users.

New Features

Core

• Live debugging session support via open_windbg_remote and close_windbg_remote
• Extended dump file support for .mdmp and .hdmp formats
• Microsoft Store WinDbg CDB compatibility

Devops

• Set up continuous integration that automatically tests the code with Python versions 3.10 through 3.14
• Added automatic dependency updates to keep everything secure and up-to-date
• Streamlined the release process so new versions reach users faster

Development

• Switched to uv - a lightning-fast Python package manager that's 10-100x faster than pip
• Development setup is now much quicker with commands like uv sync and uv run
• More reliable builds with locked dependency versions

Documentation

• Added comprehensive debugging instructions for AI assistants (AGENTS.md)
• Created structured templates to help analyze crash dumps more effectively (.github/prompts/dump-triage.prompt.md)
• All documentation is now available in the repository Wiki for easy access
• Simplified the main README.md to focus on getting started quickly
• Added this structured CHANGELOG.md to track all project changes

Improvements

Performance Boost: Build times are significantly faster thanks to the new tooling Enhanced Security: Automatic scanning and updates keep dependencies secure

Community Contributions

Special thanks to @sooknarine for these valuable contributions:

• Find local dumps with other common extensions #6 - Now finds more crash dump files automatically
• Add support for remote debugging #10 - Connect to live debugging sessions

[0.1.0] - 2025-05-03

• Initial version as blogged about.