[ci] Restore memory impact detail for ESP-IDF builds (#17587)

esphome/analyze_memory/cli.py

@@ -20,6 +20,7 @@ from . import (
     RAM_SECTIONS,
     MemoryAnalyzer,
 )
+from .toolchain import find_elf_path, find_idedata_path, idedata_candidates
 
 if TYPE_CHECKING:
     from . import ComponentMemory
@@ -759,45 +760,25 @@ def main():
         print(f"Error: {build_path} is not a directory", file=sys.stderr)
         sys.exit(1)
 
-    # Find firmware.elf
-    elf_file = None
-    for elf_candidate in [
-        build_path / "firmware.elf",
-        build_path / ".pioenvs" / build_path.name / "firmware.elf",
-    ]:
-        if elf_candidate.exists():
-            elf_file = str(elf_candidate)
-            break
-
-    if not elf_file:
-        print(f"Error: firmware.elf not found in {build_dir}", file=sys.stderr)
+    elf_path = find_elf_path(build_path)
+    if not elf_path:
+        print(f"Error: no firmware ELF found in {build_dir}", file=sys.stderr)
         sys.exit(1)
-
-    # Find idedata.json - check current directory first, then home
-    device_name = build_path.name
-    idedata_candidates = [
-        Path.cwd() / ".esphome" / "idedata" / f"{device_name}.json",
-        Path.home() / ".esphome" / "idedata" / f"{device_name}.json",
-    ]
+    elf_file = str(elf_path)
 
     idedata = None
-    for idedata_path in idedata_candidates:
-        if not idedata_path.exists():
-            continue
+    if idedata_path := find_idedata_path(build_path):
         try:
             with idedata_path.open(encoding="utf-8") as f:
                 raw_data = json.load(f)
             idedata = IDEData(raw_data)
             print(f"Loaded idedata from: {idedata_path}", file=sys.stderr)
-            break
         except (json.JSONDecodeError, OSError) as e:
             print(f"Warning: Failed to load idedata: {e}", file=sys.stderr)
 
     if not idedata:
-        print(
-            f"Warning: idedata not found (searched {idedata_candidates[0]} and {idedata_candidates[1]})",
-            file=sys.stderr,
-        )
+        searched = "\n  ".join(str(p) for p in idedata_candidates(build_path))
+        print(f"Warning: idedata not found, searched:\n  {searched}", file=sys.stderr)
 
     analyzer = MemoryAnalyzerCLI(elf_file, idedata=idedata)
     analyzer.analyze()

esphome/analyze_memory/toolchain.py

@@ -23,6 +23,78 @@ TOOLCHAIN_PREFIXES = [
 ]
 
 
+def find_elf_path(build_path: Path) -> Path | None:
+    """Locate the firmware ELF inside an ESPHome build directory.
+
+    The layout depends on the toolchain that produced the build, so try each
+    known one in turn.
+
+    Args:
+        build_path: Path to an ESPHome build directory
+
+    Returns:
+        Path to the ELF file, or None if no known layout matches
+    """
+    name = build_path.name
+    for candidate in (
+        # Native ESP-IDF: idf.py writes build/<name>.elf, which ESPHome copies
+        # to build/firmware.elf (see espidf.toolchain.create_elf_copy)
+        build_path / "build" / "firmware.elf",
+        # PlatformIO
+        build_path / "firmware.elf",
+        build_path / ".pioenvs" / name / "firmware.elf",
+        # LibreTiny uses raw_firmware.elf
+        build_path / "raw_firmware.elf",
+        build_path / ".pioenvs" / name / "raw_firmware.elf",
+        # Zephyr (nRF52); the SDK nests the artifacts one level deeper from 2.9.2
+        build_path / ".pioenvs" / name / "zephyr" / "zephyr" / "zephyr.elf",
+        build_path / ".pioenvs" / name / "zephyr" / "zephyr.elf",
+    ):
+        if candidate.is_file():
+            return candidate
+    return None
+
+
+def idedata_candidates(build_path: Path) -> list[Path]:
+    """Return the idedata locations searched for a build directory, in order.
+
+    Exposed so a caller reporting "not found" can name the paths it tried
+    without keeping its own copy of the list.
+
+    Args:
+        build_path: Path to an ESPHome build directory
+
+    Returns:
+        The candidate idedata JSON paths, most specific first
+    """
+    name = build_path.name
+    return [
+        # In .pioenvs for test builds
+        build_path / ".pioenvs" / name / "idedata.json",
+        # Both toolchains cache it in the data dir, which holds this build dir:
+        # <data_dir>/idedata/<name>.json next to <data_dir>/build/<name>
+        build_path.parent.parent / "idedata" / f"{name}.json",
+        # Regular builds, invoked from the config dir or from anywhere
+        Path.cwd() / ".esphome" / "idedata" / f"{name}.json",
+        Path.home() / ".esphome" / "idedata" / f"{name}.json",
+    ]
+
+
+def find_idedata_path(build_path: Path) -> Path | None:
+    """Locate the idedata JSON belonging to an ESPHome build directory.
+
+    Args:
+        build_path: Path to an ESPHome build directory
+
+    Returns:
+        Path to the idedata JSON, or None if it was not found
+    """
+    for candidate in idedata_candidates(build_path):
+        if candidate.is_file():
+            return candidate
+    return None
+
+
 def _find_in_platformio_packages(tool_name: str) -> str | None:
     """Search for a tool in PlatformIO package directories.
 

esphome/espidf/idedata.py

@@ -6,7 +6,7 @@ toolchain has no such command, but its CMake build emits
 turns that file into the same fields consumers (IDE integration, clang-tidy)
 expect:
 
-    {cxx_path, cxx_flags, defines, includes: {build, toolchain}}
+    {cc_path, cxx_path, cxx_flags, defines, includes: {build, toolchain}}
 """
 
 from __future__ import annotations
@@ -197,6 +197,28 @@ def _get_toolchain_includes(cxx_path: str) -> list[str]:
     return includes
 
 
+def _cc_path_from_cxx(cxx_path: str) -> str:
+    """Derive the C compiler path from the C++ compiler path.
+
+    compile_commands.json only names the C++ compiler, but consumers reach the
+    rest of the toolchain (objdump, readelf, addr2line) by rewriting the tail of
+    ``cc_path``, so they need the ``gcc``-suffixed name.
+    """
+    stem, suffix = (
+        (cxx_path[: -len(".exe")], ".exe")
+        if cxx_path.endswith(".exe")
+        else (cxx_path, "")
+    )
+    # Rewrite the program name only when it is g++ itself, or a toolchain
+    # prefixed one such as xtensa-esp32-elf-g++ -> xtensa-esp32-elf-gcc.
+    # Requiring a separator before the "g++" keeps names that merely end in
+    # those three characters intact: "clang++" must not become "clangcc".
+    head = stem[: -len("g++")]
+    if stem.endswith("g++") and (not head or head.endswith(("-", "/", "\\"))):
+        stem = f"{head}gcc"
+    return f"{stem}{suffix}"
+
+
 def idedata_from_build(compile_commands: Path) -> dict:
     """Parse compile_commands.json into the idedata fields consumers expect.
 
@@ -218,6 +240,7 @@ def idedata_from_build(compile_commands: Path) -> dict:
             build_includes.setdefault(inc, None)
 
     return {
+        "cc_path": _cc_path_from_cxx(cxx_path),
         "cxx_path": cxx_path,
         "cxx_flags": cxx_flags,
         "defines": defines,

esphome/espidf/toolchain.py

@@ -467,9 +467,16 @@ def get_idedata() -> dict | None:
     cache = CORE.relative_internal_path("idedata", f"{CORE.name}.json")
     if cache.is_file() and cache.stat().st_mtime >= compile_commands.stat().st_mtime:
         try:
-            return json.loads(cache.read_text(encoding="utf-8"))
+            cached = json.loads(cache.read_text(encoding="utf-8"))
         except ValueError:
             pass
+        else:
+            # Caches written before cc_path was emitted stay newer than
+            # compile_commands.json forever, so rebuild them on the field rather
+            # than on the timestamp. Check the type too: a corrupted cache can
+            # still be valid JSON, and "in" would match a substring of a string.
+            if isinstance(cached, dict) and "cc_path" in cached:
+                return cached
 
     data = idedata_from_build(compile_commands)
     data["prog_path"] = str(get_elf_path())

script/ci_memory_impact_extract.py

@@ -33,6 +33,11 @@ sys.path.insert(0, str(Path(__file__).parent.parent))
 
 # pylint: disable=wrong-import-position
 from esphome.analyze_memory import MemoryAnalyzer
+from esphome.analyze_memory.toolchain import (
+    find_elf_path,
+    find_idedata_path,
+    idedata_candidates,
+)
 from esphome.platformio.toolchain import IDEData
 from script.ci_helpers import write_github_output
 
@@ -130,53 +135,31 @@ def run_detailed_analysis(build_dir: str) -> dict | None:
         print(f"Build directory not found: {build_dir}", file=sys.stderr)
         return None
 
-    # Find firmware.elf (or raw_firmware.elf for LibreTiny)
-    elf_path = None
-    for elf_candidate in [
-        build_path / "firmware.elf",
-        build_path / ".pioenvs" / build_path.name / "firmware.elf",
-        # LibreTiny uses raw_firmware.elf
-        build_path / "raw_firmware.elf",
-        build_path / ".pioenvs" / build_path.name / "raw_firmware.elf",
-    ]:
-        if elf_candidate.exists():
-            elf_path = str(elf_candidate)
-            break
-
+    elf_path = find_elf_path(build_path)
     if not elf_path:
-        print(
-            f"firmware.elf/raw_firmware.elf not found in {build_dir}", file=sys.stderr
-        )
+        print(f"No firmware ELF found in {build_dir}", file=sys.stderr)
         return None
 
-    # Find idedata.json - check multiple locations
-    device_name = build_path.name
-    idedata_candidates = [
-        # In .pioenvs for test builds
-        build_path / ".pioenvs" / device_name / "idedata.json",
-        # In .esphome/idedata for regular builds
-        Path.home() / ".esphome" / "idedata" / f"{device_name}.json",
-        # Check parent directories for .esphome/idedata (for test_build_components)
-        build_path.parent.parent.parent / "idedata" / f"{device_name}.json",
-    ]
-
     idedata = None
-    for idedata_path in idedata_candidates:
-        if not idedata_path.exists():
-            continue
+    if idedata_path := find_idedata_path(build_path):
         try:
             with idedata_path.open(encoding="utf-8") as f:
                 raw_data = json.load(f)
             idedata = IDEData(raw_data)
             print(f"Loaded idedata from: {idedata_path}", file=sys.stderr)
-            break
         except (json.JSONDecodeError, OSError) as e:
             print(
                 f"Warning: Failed to load idedata from {idedata_path}: {e}",
                 file=sys.stderr,
             )
+    else:
+        # Without idedata the analyzer falls back to whatever binutils are on
+        # PATH, which are the wrong architecture for a cross build, so say where
+        # we looked rather than let the results quietly get worse.
+        searched = "\n  ".join(str(p) for p in idedata_candidates(build_path))
+        print(f"Warning: idedata not found, searched:\n  {searched}", file=sys.stderr)
 
-    analyzer = MemoryAnalyzer(elf_path, idedata=idedata)
+    analyzer = MemoryAnalyzer(str(elf_path), idedata=idedata)
     components = analyzer.analyze()
 
     # Convert to JSON-serializable format
@@ -320,6 +303,19 @@ def main() -> int:
     else:
         print(f"{ram_bytes},{flash_bytes}")
 
+    # The build produced usable totals, so a missing detailed analysis means the
+    # build layout moved out from under this script rather than a broken build.
+    # Fail loudly: the comment would otherwise silently drop the component
+    # breakdown and the symbol tables, which is easy to miss for a long time.
+    if detailed_analysis is None:
+        print(
+            "::error::Detailed memory analysis unavailable even though the build "
+            f"succeeded (build directory: {build_dir or 'not detected'}). The PR "
+            "comment would be missing its component breakdown and symbol changes.",
+            file=sys.stderr,
+        )
+        return 1
+
     return 0
 
 

tests/script/test_determine_jobs.py

@@ -2993,3 +2993,53 @@ def test_main_force_all_off_uses_detection(
     assert output["component_test_count"] == 0
     mock_determine_integration_tests.assert_called_once()
     mock_should_run_clang_tidy.assert_called_once()
+
+
+# Every platform the memory impact analysis can select must produce an ELF that
+# find_elf_path knows how to locate. The analysis fails the job when it cannot
+# find one, so a platform with an unknown layout would turn a clean build red.
+_MEMORY_IMPACT_ELF_LAYOUTS = {
+    # Native ESP-IDF toolchain (the esp32 default): <build>/build/firmware.elf
+    "esp32-c6-idf": "build/firmware.elf",
+    "esp32-idf": "build/firmware.elf",
+    "esp32-c3-idf": "build/firmware.elf",
+    "esp32-s2-idf": "build/firmware.elf",
+    "esp32-s3-idf": "build/firmware.elf",
+    # PlatformIO: <build>/.pioenvs/<name>/firmware.elf
+    "esp8266-ard": ".pioenvs/{name}/firmware.elf",
+    "rp2040-ard": ".pioenvs/{name}/firmware.elf",
+    "rp2350-ard": ".pioenvs/{name}/firmware.elf",
+    # LibreTiny: <build>/.pioenvs/<name>/raw_firmware.elf
+    "bk72xx-ard": ".pioenvs/{name}/raw_firmware.elf",
+    "rtl87xx-ard": ".pioenvs/{name}/raw_firmware.elf",
+    "ln882x-ard": ".pioenvs/{name}/raw_firmware.elf",
+    # Zephyr: <build>/.pioenvs/<name>/zephyr/[zephyr/]zephyr.elf
+    "nrf52-adafruit": ".pioenvs/{name}/zephyr/zephyr/zephyr.elf",
+}
+
+
+def test_memory_impact_platforms_have_known_elf_layout() -> None:
+    """Every selectable memory impact platform has a documented ELF layout.
+
+    Adding a platform to the preference list without teaching find_elf_path
+    where its ELF lands would fail the memory impact job on a clean build.
+    """
+    selectable = {
+        platform.value for platform in determine_jobs.MEMORY_IMPACT_PLATFORM_PREFERENCE
+    }
+    selectable.add(determine_jobs.MEMORY_IMPACT_FALLBACK_PLATFORM.value)
+
+    assert selectable == set(_MEMORY_IMPACT_ELF_LAYOUTS)
+
+
+def test_memory_impact_elf_layouts_are_found(tmp_path: Path) -> None:
+    """find_elf_path locates the ELF each memory impact platform produces."""
+    from esphome.analyze_memory.toolchain import find_elf_path
+
+    for platform, layout in _MEMORY_IMPACT_ELF_LAYOUTS.items():
+        build_path = tmp_path / platform / ".esphome" / "build" / "mydevice"
+        elf = build_path / layout.format(name=build_path.name)
+        elf.parent.mkdir(parents=True)
+        elf.write_text("")
+
+        assert find_elf_path(build_path) == elf, f"{platform} ELF not found"

tests/unit_tests/analyze_memory/test_build_artifacts.py

@@ -0,0 +1,157 @@
+"""Tests for locating build artifacts across the supported toolchain layouts."""
+
+from pathlib import Path
+
+import pytest
+
+from esphome.analyze_memory.toolchain import (
+    find_elf_path,
+    find_idedata_path,
+    idedata_candidates,
+)
+from esphome.espidf.idedata import _cc_path_from_cxx
+from esphome.platformio.toolchain import IDEData
+
+
+def _make_build_dir(tmp_path: Path, name: str = "mydevice") -> Path:
+    """Create <tmp_path>/.esphome/build/<name>, mirroring a real data dir."""
+    build_path = tmp_path / ".esphome" / "build" / name
+    build_path.mkdir(parents=True)
+    return build_path
+
+
+def _touch(path: Path) -> Path:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text("")
+    return path
+
+
+def test_find_elf_path_native_esp_idf(tmp_path: Path) -> None:
+    """The native ESP-IDF toolchain writes the ELF under build/."""
+    build_path = _make_build_dir(tmp_path)
+    elf = _touch(build_path / "build" / "firmware.elf")
+
+    assert find_elf_path(build_path) == elf
+
+
+def test_find_elf_path_platformio(tmp_path: Path) -> None:
+    """The PlatformIO toolchain writes the ELF under .pioenvs/<name>/."""
+    build_path = _make_build_dir(tmp_path)
+    elf = _touch(build_path / ".pioenvs" / build_path.name / "firmware.elf")
+
+    assert find_elf_path(build_path) == elf
+
+
+def test_find_elf_path_libretiny(tmp_path: Path) -> None:
+    """The LibreTiny toolchain names the unwrapped ELF raw_firmware.elf."""
+    build_path = _make_build_dir(tmp_path)
+    elf = _touch(build_path / ".pioenvs" / build_path.name / "raw_firmware.elf")
+
+    assert find_elf_path(build_path) == elf
+
+
+@pytest.mark.parametrize(
+    "relative_elf",
+    [
+        # SDK < 2.9.2
+        "zephyr/zephyr.elf",
+        # SDK >= 2.9.2 nests the artifacts one level deeper
+        "zephyr/zephyr/zephyr.elf",
+    ],
+)
+def test_find_elf_path_zephyr(tmp_path: Path, relative_elf: str) -> None:
+    """Zephyr (nRF52) keeps the ELF under .pioenvs/<name>/zephyr/."""
+    build_path = _make_build_dir(tmp_path)
+    elf = _touch(build_path / ".pioenvs" / build_path.name / relative_elf)
+
+    assert find_elf_path(build_path) == elf
+
+
+def test_find_elf_path_missing(tmp_path: Path) -> None:
+    """An unknown layout resolves to None rather than a bogus path."""
+    assert find_elf_path(_make_build_dir(tmp_path)) is None
+
+
+def test_find_idedata_path_in_data_dir(tmp_path: Path) -> None:
+    """The idedata cache sits in the data dir that holds the build dir."""
+    build_path = _make_build_dir(tmp_path)
+    idedata = _touch(tmp_path / ".esphome" / "idedata" / f"{build_path.name}.json")
+
+    assert find_idedata_path(build_path) == idedata
+
+
+def test_find_idedata_path_in_pioenvs(tmp_path: Path) -> None:
+    """Test builds may keep idedata alongside the PlatformIO env."""
+    build_path = _make_build_dir(tmp_path)
+    idedata = _touch(build_path / ".pioenvs" / build_path.name / "idedata.json")
+
+    assert find_idedata_path(build_path) == idedata
+
+
+def test_find_idedata_path_missing(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """A missing idedata resolves to None."""
+    # Keep the cwd/home fallbacks from finding an unrelated file on this machine
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setattr(Path, "home", classmethod(lambda cls: tmp_path))
+
+    assert find_idedata_path(_make_build_dir(tmp_path)) is None
+
+
+def test_idedata_candidates_are_what_find_probes(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """Every advertised candidate is one find_idedata_path actually accepts.
+
+    The candidates are reported to the user when idedata is missing, so a list
+    that drifts from the lookup would send someone hunting in the wrong place.
+    """
+    # Two candidates are relative to the cwd and to home; keep the test from
+    # writing into the real ones.
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setattr(Path, "home", classmethod(lambda cls: tmp_path))
+
+    build_path = _make_build_dir(tmp_path)
+    candidates = idedata_candidates(build_path)
+
+    assert candidates, "no candidates advertised"
+    for candidate in candidates:
+        _touch(candidate)
+        assert find_idedata_path(build_path) == candidate
+        candidate.unlink()
+
+
+@pytest.mark.parametrize(
+    ("cxx_path", "expected"),
+    [
+        ("/tools/bin/xtensa-esp32-elf-g++", "/tools/bin/xtensa-esp32-elf-gcc"),
+        ("/tools/bin/riscv32-esp-elf-g++", "/tools/bin/riscv32-esp-elf-gcc"),
+        (
+            r"C:\tools\bin\xtensa-esp32-elf-g++.exe",
+            r"C:\tools\bin\xtensa-esp32-elf-gcc.exe",
+        ),
+        # Nothing to rewrite; leave the path alone
+        ("/tools/bin/clang++", "/tools/bin/clang++"),
+    ],
+)
+def test_cc_path_from_cxx(cxx_path: str, expected: str) -> None:
+    """cc_path is derived from the C++ compiler that compile_commands.json names."""
+    assert _cc_path_from_cxx(cxx_path) == expected
+
+
+def test_native_idedata_resolves_toolchain_tools() -> None:
+    """The binutils paths are derived from the native ESP-IDF cc_path.
+
+    Without cc_path, IDEData.objdump_path raises KeyError and the memory
+    analysis silently degrades to no component or symbol detail.
+    """
+    idedata = IDEData(
+        {
+            "cc_path": _cc_path_from_cxx("/tools/bin/xtensa-esp32-elf-g++"),
+            "cxx_path": "/tools/bin/xtensa-esp32-elf-g++",
+        }
+    )
+
+    assert idedata.objdump_path == "/tools/bin/xtensa-esp32-elf-objdump"
+    assert idedata.readelf_path == "/tools/bin/xtensa-esp32-elf-readelf"

tests/unit_tests/analyze_memory/test_ci_memory_impact_extract.py

@@ -0,0 +1,57 @@
+"""Tests for script/ci_memory_impact_extract.py."""
+
+import io
+from pathlib import Path
+import sys
+
+import pytest
+
+# Add script directory to path so we can import the module
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "script"))
+
+from ci_memory_impact_extract import main  # noqa: E402
+
+_COMPILE_OUTPUT = (
+    "RAM:   [====      ]  36.1% (used 29548 bytes from 81920 bytes)\n"
+    "Flash: [===       ]  34.0% (used 348511 bytes from 1023984 bytes)\n"
+)
+
+
+@pytest.fixture(autouse=True)
+def _no_github_output(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.delenv("GITHUB_OUTPUT", raising=False)
+
+
+def _run(monkeypatch: pytest.MonkeyPatch, compile_output: str, argv: list[str]) -> int:
+    monkeypatch.setattr(sys, "stdin", io.StringIO(compile_output))
+    monkeypatch.setattr(sys, "argv", ["ci_memory_impact_extract.py", *argv])
+    return main()
+
+
+def test_missing_detailed_analysis_fails(
+    monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+) -> None:
+    """A build with no usable ELF fails instead of posting a comment without detail."""
+    build_dir = tmp_path / ".esphome" / "build" / "mydevice"
+    build_dir.mkdir(parents=True)
+    out_json = tmp_path / "analysis.json"
+
+    rc = _run(
+        monkeypatch,
+        _COMPILE_OUTPUT,
+        ["--build-dir", str(build_dir), "--output-json", str(out_json)],
+    )
+
+    assert rc == 1
+    # The totals are still written so the failure can be diagnosed from the artifact
+    assert out_json.is_file()
+
+
+def test_undetected_build_dir_fails(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Compile output without a build path cannot be analyzed, so it fails."""
+    assert _run(monkeypatch, _COMPILE_OUTPUT, []) == 1
+
+
+def test_unparseable_output_fails(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Output with no memory totals at all is still a failure."""
+    assert _run(monkeypatch, "nothing useful here\n", []) == 1

tests/unit_tests/test_espidf_toolchain.py

@@ -7,6 +7,8 @@ import os
 from pathlib import Path
 from unittest.mock import patch
 
+import pytest
+
 from esphome.const import CONF_FRAMEWORK, CONF_SOURCE
 from esphome.core import CORE
 from esphome.espidf import toolchain
@@ -100,7 +102,7 @@ def test_get_idedata_uses_cache_when_valid(setup_core: Path) -> None:
     compile_commands.parent.mkdir(parents=True, exist_ok=True)
     compile_commands.write_text("[]")
     cache.parent.mkdir(parents=True, exist_ok=True)
-    cache.write_text('{"cxx_path": "cached"}')
+    cache.write_text('{"cc_path": "cached-gcc", "cxx_path": "cached"}')
     cc_mtime = compile_commands.stat().st_mtime
     os.utime(cache, (cc_mtime + 1, cc_mtime + 1))
 
@@ -108,7 +110,31 @@ def test_get_idedata_uses_cache_when_valid(setup_core: Path) -> None:
         result = toolchain.get_idedata()
 
     mock_transform.assert_not_called()
-    assert result == {"cxx_path": "cached"}
+    assert result == {"cc_path": "cached-gcc", "cxx_path": "cached"}
+
+
+def test_get_idedata_regenerates_cache_without_cc_path(setup_core: Path) -> None:
+    """A cache predating cc_path is rebuilt even though it is newer.
+
+    Such a cache stays newer than the compile DB forever, so consumers that
+    derive the binutils paths from cc_path would keep failing on it.
+    """
+    compile_commands, cache = _setup_build(setup_core)
+    compile_commands.parent.mkdir(parents=True, exist_ok=True)
+    compile_commands.write_text("[]")
+    cache.parent.mkdir(parents=True, exist_ok=True)
+    cache.write_text('{"cxx_path": "cached"}')
+    cc_mtime = compile_commands.stat().st_mtime
+    os.utime(cache, (cc_mtime + 1, cc_mtime + 1))
+
+    with patch(
+        "esphome.espidf.idedata.idedata_from_build",
+        return_value={"cc_path": "gcc", "cxx_path": "g++"},
+    ) as mock_transform:
+        result = toolchain.get_idedata()
+
+    mock_transform.assert_called_once()
+    assert result["cc_path"] == "gcc"
 
 
 def test_get_idedata_regenerates_when_compile_commands_newer(setup_core: Path) -> None:
@@ -131,6 +157,33 @@ def test_get_idedata_regenerates_when_compile_commands_newer(setup_core: Path) -
     assert result == {"cxx_path": "fresh", "prog_path": str(toolchain.get_elf_path())}
 
 
+@pytest.mark.parametrize("cached", ['"cc_path is a string"', "[]", "42"])
+def test_get_idedata_regenerates_on_non_dict_cache(
+    setup_core: Path, cached: str
+) -> None:
+    """A newer cache holding valid JSON that is not an object is regenerated.
+
+    A bare string would otherwise pass the cc_path check by substring and be
+    handed to consumers expecting a dict.
+    """
+    compile_commands, cache = _setup_build(setup_core)
+    compile_commands.parent.mkdir(parents=True, exist_ok=True)
+    compile_commands.write_text("[]")
+    cache.parent.mkdir(parents=True, exist_ok=True)
+    cache.write_text(cached)
+    cc_mtime = compile_commands.stat().st_mtime
+    os.utime(cache, (cc_mtime + 1, cc_mtime + 1))
+
+    with patch(
+        "esphome.espidf.idedata.idedata_from_build",
+        return_value={"cc_path": "gcc", "cxx_path": "g++"},
+    ) as mock_transform:
+        result = toolchain.get_idedata()
+
+    mock_transform.assert_called_once()
+    assert isinstance(result, dict)
+
+
 def test_get_idedata_regenerates_on_corrupted_cache(setup_core: Path) -> None:
     """An unparseable (but newer) cache falls back to regeneration."""
     compile_commands, cache = _setup_build(setup_core)