Zephyr Debug

Zephyr Debug

Audit workflow

For every non-trivial task that writes to the workspace, choose a stable task slug and keep agent-only records under:

.zephyr-skills/<agent-task>/
├── audit.md        # Baseline, decisions, evidence, verification, and gaps
├── commands.md     # Redacted commands, working directories, and results
├── scripts/        # Intermediate debug documents, scripts, and harnesses
├── output/         # Non-Zephyr dependencies and temporary binaries
└── logs/           # Decisive build, test, runtime, or diagnostic logs

Put every generated intermediate debugging document, debug script, and scaffolding or harness script, including Bash and Python scripts, under scripts/, never in a build directory. Put non-Zephyr third-party source or libraries acquired solely for the agent's task, and temporary binaries produced by agent-only probes or tools, under output/. Do not copy normal Zephyr build artifacts there. Keep other agent plans and temporary evidence in the same task directory. In a Git worktree, add /.zephyr-skills/ to the repository-local exclude file returned by git rev-parse --git-path info/exclude before writing audit artifacts. Preserve existing entries, avoid duplicates, and verify before handoff that git status --short contains no .zephyr-skills/ paths. Keep Zephyr-native build trees, binaries, and test outputs separate: honor a user-specified path, otherwise use defaults such as configured build.dir-fmt, build/, or twister-out/. Record the effective paths, do not stage .zephyr-skills/ unless requested, and report the task directory and unresolved gaps at handoff.

Diagnose

1. Capture expected and observed behavior, the narrowest reproducer, and the first decisive log.
2. Exclude wrong revisions, boards, configuration, stale output, and generated artifact or toolchain mismatches before changing source.
3. Locate the first broken invariant across Kconfig, CMake, devicetree, generated files, runner commands, boot flow, and runtime state.
4. Test each hypothesis with the smallest reversible probe and distinguish observations from inferences.
5. For diagnosis-only requests, stop at the supported cause and remaining uncertainty. Implement a fix and regression gate only when requested.

Choose a debugging path

• Use Zephyr logging and targeted instrumentation when the failure remains observable without stopping the system.
• Use a QEMU GDB stub for reproducible emulated failures, early boot, exception state, memory inspection, breakpoints, and instruction-level stepping.
• Use the board's configured debug runner with OpenOCD or J-Link for hardware failures. Select the runner from the board, probe, and user environment; do not assume that one backend is universally available.
• Combine logs with GDB when console history or timing explains state that a stopped target cannot.

Run a remote GDB session

1. Preserve the exact ELF and build configuration that reproduce the failure. Use the matching toolchain GDB and keep symbols in the Zephyr build tree.
2. Start the remote endpoint with the configured Zephyr runner: a QEMU GDB stub for emulation, or a debug server backed by OpenOCD or J-Link for hardware. Prefer runner-generated arguments over guessed probe, transport, reset, or device settings.
3. If installed, use Tmux for concurrent panes; otherwise use Zellij, then separate terminals. Keep the debug server or QEMU in one pane, GDB in a second, and optional console or log monitoring in a third.
4. Generate a task-specific GDB command file under .zephyr-skills/<agent-task>/scripts/, including the symbol file, remote endpoint, breakpoints or watchpoints, thread inspection, and evidence-producing commands. Avoid changing global GDB configuration.
5. Connect with target remote or target extended-remote as required by the server. Capture the server command, GDB command file, decisive backtraces, register or memory state, and target logs in the audit directory.
6. For a live hardware target, prefer attach semantics when flashing or reset is not requested. Record any operation that changes target state.

Use Zephyr logging and instrumentation

• Enable the Zephyr logging subsystem and the narrowest useful module or subsystem log level through task-specific configuration or overlays.
• Add LOG_DBG, LOG_INF, LOG_WRN, LOG_ERR, or hexdump probes at state transitions, error paths, ownership changes, and relevant interrupt or scheduling boundaries.
• Include stable identifiers and state in messages, but never log secrets or unbounded data. Keep probes minimal and easy to remove or gate.
• Account for timing changes from added logging, synchronous output, transport buffering, and log overflow. Recheck the failure with reduced instrumentation before treating the result as causal.
• Save decisive console output under .zephyr-skills/<agent-task>/logs/ and state which logging configuration and probes produced it.

Guardrails

• Do not hide a failure by weakening assertions or tests.
• Stop only QEMU processes started for the task; never kill broad process groups.
• Stop only debug servers and multiplexer sessions started for the task.