Skill v1.0.0
currentTrusted Publisher100/100version: "1.0.0" name: windbg-user-ttd-reverse-debugging-triage description: 'Use when an app, service, or user-mode driver host TTD recording is available and earlier calls, writes, or lifetimes matter. Not for kernel replay or history from a normal dump.'
TTD Reverse Debugging Triage
Load `windbg-diagnostic-method` first if it is not already loaded in this conversation, and apply it throughout for evidence ranking, hypothesis testing, confidence calibration, independent review, and report validation. This skill adds the bug-family-specific commands and evidence requirements.
Requirements and scope
Time Travel Debugging records a user-mode process for replay, including an authorized application, service, or user-mode driver host process. A dump alone contains no execution timeline. Confirm a valid recording and matching module symbols before attempting timeline queries. It does not record kernel execution or automatically include another process's server-side activity.
Recording can change timing, require substantial storage, and capture sensitive memory. Obtain authorization before capture. Check the installed recorder's help for supported target, architecture, and options. WinDbg's Launch executable (advanced) / Record with Time Travel Debugging flow can record a named test application without guessing a recorder command.
Workflow
1. Open and index the recording
Open the recorded .run file in WinDbg and load public Windows symbols and matching PDBs for your own binaries.
!tt.indexdx @$cursession.TTD
Confirm the timeline range. Indexing and data-model availability depend on WinDbg/TTD versions; consult the installed help if the command is unavailable.
2. Ask a bounded history question
dx @$cursession.TTD.Calls("MyModule!MyFunction")dx @$cursession.TTD.Memory(<start-address>, <end-address>, "w")
Query the known function or exact memory range, then inspect relevant results and their timeline positions. An empty result can mean unmatched symbols, uninstrumented execution, or an out-of-range query, not proof of absence.
Use the call query to find allocation/free or module lifetime transitions and the memory query to locate candidate writes. Account for allocator reuse: the same virtual address may represent different objects over the recording.
3. Seek and inspect state
!tt <position>krg-
Seek to an actual position returned by the query. g- continues backward; use the installed reverse-step controls to refine the search. Capture the last-good and first-bad states, writer/caller stack, and ownership transition.
Do not equate the last write with the bug until you establish the object's identity, valid lifetime, intended invariant, and relevant cross-thread order.
4. Continue with a shipped investigation
| Evidence | Skill | |
|---|---|---|
| Bad free/write or allocation-boundary violation | windbg-user-heap-corruption-investigation | |
| Growing allocation/reservation usage | windbg-user-virtual-memory-exhaustion | |
| Lock survives coroutine suspension | windbg-user-mutex-held-across-co-await | |
| Blocker or cross-process dependency at the selected moment | windbg-user-wait-chain-analysis |
For other user-mode families continue reasoning from the timeline rather than dispatching to absent skills. For a kernel crash use windbg-kernel-bugcheck-triage; TTD is not a kernel-history substitute.
Validation
Record the question, trace identity, timeline positions, relevant object lifetime, and source/stack evidence. Test alternatives and state recording boundaries. A reproducible timeline proves what occurred in that recording; it does not establish that instrumentation preserved all production timing.
References
Feedback
Follow FEEDBACK.md and report a reviewed summary to WinDbg-Feedback. Include windbg-user-ttd-reverse-debugging-triage and the package version from plugin.json; never automatically upload a recording or its memory/query contents.