Fix leaderboard savings lookup after provider-key rename (#635)

PR #634 renamed the Anthropic cost-comparison provider key
`claude-opus-4.6` -> `claude-fable-5` but missed one consumer:
App.tsx looks up the Anthropic entry by that key to compute the
`dollar_savings` value submitted to the leaderboard. After the rename
`per_provider.find(p => p.provider === 'claude-opus-4.6')` returned
undefined, so this path silently submitted dollar_savings = 0.

Point the lookup at the new key.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

frontend/src/App.tsx

@@ -89,7 +89,7 @@ export default function App() {
           setSavings(data);
           if (optInEnabled && optInDisplayName && data) {
             const claudeEntry = data.per_provider.find(
-              (p) => p.provider === 'claude-opus-4.6',
+              (p) => p.provider === 'claude-fable-5',
             );
             const dollarSavings = claudeEntry ? claudeEntry.total_cost : 0;
             const energySaved = data.per_provider.reduce(

rust/crates/openjarvis-tools/src/storage/sqlite.rs

@@ -144,16 +144,21 @@ impl MemoryBackend for SQLiteMemory {
     ) -> Result<Vec<RetrievalResult>, OpenJarvisError> {
         let conn = self.conn.lock();
 
+        // Split on any non-alphanumeric character (not just whitespace) so
+        // internal punctuation — apostrophes in particular ("user's") — never
+        // reaches the FTS5 MATCH string. FTS5's query grammar treats an
+        // unescaped `'` as a string delimiter, so passing a raw token like
+        // `user's` through silently fails to parse and yields zero rows with
+        // no visible error. Splitting fully avoids needing to escape anything.
         let words: Vec<String> = query
-            .split_whitespace()
-            .map(|w| w.trim_matches(|c: char| "?.,!;:'\"()[]{}/ ".contains(c)).to_string())
+            .split(|c: char| !c.is_alphanumeric())
+            .map(|w| w.to_string())
             .filter(|w| !w.is_empty())
             .collect();
-        let fts_query = if words.len() == 1 {
-            words[0].clone()
-        } else {
-            words.join(" OR ")
-        };
+        if words.is_empty() {
+            return Ok(Vec::new());
+        }
+        let fts_query = words.join(" OR ");
 
         let mut stmt = conn
             .prepare(
@@ -320,6 +325,27 @@ mod tests {
         assert_eq!(mixed.len(), 2, "mixed-case query should find both documents");
     }
 
+    #[test]
+    fn test_sqlite_apostrophe_in_query() {
+        let mem = SQLiteMemory::in_memory().unwrap();
+        mem.store("The user's name is Trev.", "identity", None).unwrap();
+
+        // A query containing an internal apostrophe must not break FTS5's
+        // MATCH syntax (an unescaped `'` is a string delimiter in FTS5's
+        // query grammar), which previously caused this to silently return
+        // zero results instead of matching or erroring.
+        let multi_word = mem.retrieve("what is the user's name", 5).unwrap();
+        assert!(
+            !multi_word.is_empty(),
+            "query with an internal apostrophe should not silently return zero results"
+        );
+
+        // Bare single-word possessive: exercises the (former) single-word
+        // bypass path that skipped the OR-join entirely.
+        let bare = mem.retrieve("user's", 5).unwrap();
+        assert!(!bare.is_empty(), "single-word possessive query should still match");
+    }
+
     #[test]
     fn test_sqlite_scores_are_positive() {
         let mem = SQLiteMemory::in_memory().unwrap();

src/openjarvis/cli/skill_cmd.py

@@ -199,7 +199,16 @@ def _get_resolver(source: str, url: str = ""):
     default="",
     help="Repo URL (required when source is 'github').",
 )
-def install(query: str, with_scripts: bool, force: bool, url: str):
+@click.option(
+    "--yes-dangerous",
+    is_flag=True,
+    default=False,
+    help=(
+        "Confirm installing an unreviewed skill that requests dangerous "
+        "capabilities (shell/network-listen/filesystem-write)."
+    ),
+)
+def install(query: str, with_scripts: bool, force: bool, url: str, yes_dangerous: bool):
     """Install a skill from a source.
 
     Example: ``jarvis skill install hermes:apple-notes``
@@ -233,7 +242,12 @@ def install(query: str, with_scripts: bool, force: bool, url: str):
     from openjarvis.skills.tool_translator import ToolTranslator
 
     importer = SkillImporter(parser=SkillParser(), tool_translator=ToolTranslator())
-    result = importer.import_skill(matches[0], with_scripts=with_scripts, force=force)
+    result = importer.import_skill(
+        matches[0],
+        with_scripts=with_scripts,
+        force=force,
+        confirm_dangerous=yes_dangerous,
+    )
 
     if result.success:
         if result.skipped:
@@ -270,6 +284,15 @@ def install(query: str, with_scripts: bool, force: bool, url: str):
     help="Import scripts/ directories.",
 )
 @click.option("--force", is_flag=True, default=False, help="Re-import existing skills.")
+@click.option(
+    "--yes-dangerous",
+    is_flag=True,
+    default=False,
+    help=(
+        "Confirm installing unreviewed skills that request dangerous "
+        "capabilities (shell/network-listen/filesystem-write)."
+    ),
+)
 def sync(
     source: str,
     category: str,
@@ -277,6 +300,7 @@ def sync(
     search: str,
     with_scripts: bool,
     force: bool,
+    yes_dangerous: bool,
 ):
     """Bulk install + update from a source (or all configured sources)."""
     console = Console()
@@ -343,9 +367,20 @@ def sync(
 
         installed_count = 0
         for resolved in skills_to_import:
-            r = importer.import_skill(resolved, with_scripts=with_scripts, force=force)
+            r = importer.import_skill(
+                resolved,
+                with_scripts=with_scripts,
+                force=force,
+                confirm_dangerous=yes_dangerous,
+            )
             if r.success and not r.skipped:
                 installed_count += 1
+            elif not r.success and r.requires_confirmation:
+                console.print(
+                    f"  [yellow]Skipped {resolved.name}: requests dangerous "
+                    f"capabilities {r.dangerous_capabilities} "
+                    "(re-run with --yes-dangerous to install)[/yellow]"
+                )
         console.print(f"  Imported {installed_count}/{len(skills_to_import)} skills")
         total_installed += installed_count
 

src/openjarvis/server/routes.py

@@ -59,23 +59,34 @@ def _ensure_identity_prompt(messages: list[Message], app_config) -> list[Message
     If any message already carries a system role, the caller has supplied
     their own grounding and we leave the list untouched (no double-prompting).
 
-    Resolution of the identity text: ``app_config.agent.default_system_prompt``
-    when a config is wired onto ``app.state``; otherwise fall back to
-    ``load_config()``. Config resolution is wrapped so a broken/missing
-    config degrades to "no injection" rather than crashing the endpoint, but
-    the failure is logged (per REVIEW.md — never silently swallow).
+    Resolution of the identity text: the config comes from ``app.state`` when
+    wired, otherwise ``load_config()``; the prompt itself is assembled by
+    ``SystemPromptBuilder`` from ``agent.default_system_prompt`` plus the
+    persona files (SOUL.md/MEMORY.md/USER.md), matching
+    ``_build_managed_system_prompt`` in ``agent_manager_routes.py``. Config
+    resolution is wrapped so a broken/missing config degrades to "no
+    injection" rather than crashing the endpoint, but the failure is logged
+    (per REVIEW.md — never silently swallow).
     """
     if any(m.role == Role.SYSTEM for m in messages):
         return messages
 
     prompt = ""
     try:
-        if app_config is not None:
-            prompt = app_config.agent.default_system_prompt or ""
-        else:
+        cfg = app_config
+        if cfg is None:
             from openjarvis.core.config import load_config
 
-            prompt = load_config().agent.default_system_prompt or ""
+            cfg = load_config()
+
+        from openjarvis.prompt.builder import SystemPromptBuilder
+
+        builder = SystemPromptBuilder(
+            agent_template=cfg.agent.default_system_prompt or "",
+            memory_files_config=getattr(cfg, "memory_files", None),
+            system_prompt_config=getattr(cfg, "system_prompt", None),
+        )
+        prompt = builder.build()
     except Exception:
         logging.getLogger("openjarvis.server").debug(
             "Identity system prompt resolution failed; "

src/openjarvis/skills/executor.py

@@ -5,10 +5,11 @@ from __future__ import annotations
 import json
 import re
 from dataclasses import dataclass, field
-from typing import Any, Callable, Dict, List, Optional
+from typing import Any, Callable, Dict, List, Optional, Set
 
 from openjarvis.core.events import EventBus, EventType
 from openjarvis.core.types import ToolCall, ToolResult
+from openjarvis.skills.security import validate_capabilities
 from openjarvis.skills.types import SkillManifest
 from openjarvis.tools._stubs import ToolExecutor
 
@@ -37,10 +38,16 @@ class SkillExecutor:
         tool_executor: ToolExecutor,
         *,
         bus: Optional[EventBus] = None,
+        allowed_capabilities: Optional[Set[str]] = None,
     ) -> None:
         self._tool_executor = tool_executor
         self._bus = bus
         self._skill_resolver: Optional[SkillResolver] = None
+        # None means "no capability policy" — every skill runs, matching the
+        # behavior before capability enforcement existed. Pass a set (even an
+        # empty one) to enforce: skills whose required_capabilities are not a
+        # subset of it are blocked before any step runs.
+        self._allowed_capabilities: Optional[Set[str]] = allowed_capabilities
 
     def set_skill_resolver(self, resolver: SkillResolver) -> None:
         """Register a callback used to delegate ``skill_name`` steps."""
@@ -53,6 +60,38 @@ class SkillExecutor:
         initial_context: Optional[Dict[str, Any]] = None,
     ) -> SkillResult:
         """Execute all steps in a skill manifest."""
+        missing = (
+            validate_capabilities(manifest, self._allowed_capabilities)
+            if self._allowed_capabilities is not None
+            else []
+        )
+        if missing:
+            if self._bus:
+                self._bus.publish(
+                    EventType.SKILL_EXECUTE_START,
+                    {"skill": manifest.name, "steps": len(manifest.steps)},
+                )
+                self._bus.publish(
+                    EventType.SKILL_EXECUTE_END,
+                    {"skill": manifest.name, "success": False},
+                )
+            return SkillResult(
+                skill_name=manifest.name,
+                success=False,
+                step_results=[
+                    ToolResult(
+                        tool_name=manifest.name,
+                        content=(
+                            f"Blocked: skill '{manifest.name}' requires "
+                            f"capabilities {missing} that were not granted "
+                            "for this session."
+                        ),
+                        success=False,
+                    )
+                ],
+                context=dict(initial_context or {}),
+            )
+
         ctx: Dict[str, Any] = dict(initial_context or {})
         all_results: List[ToolResult] = []
 

src/openjarvis/skills/importer.py

@@ -25,6 +25,11 @@ import yaml
 
 from openjarvis.core.paths import get_config_dir
 from openjarvis.skills.parser import SkillParser
+from openjarvis.skills.security import (
+    TrustTier,
+    classify_trust_tier,
+    has_dangerous_capabilities,
+)
 from openjarvis.skills.sources.base import ResolvedSkill
 from openjarvis.skills.tool_translator import ToolTranslator
 
@@ -43,6 +48,9 @@ class ImportResult:
     untranslated_tools: List[str] = field(default_factory=list)
     scripts_imported: bool = False
     warnings: List[str] = field(default_factory=list)
+    trust_tier: TrustTier = TrustTier.UNREVIEWED
+    dangerous_capabilities: List[str] = field(default_factory=list)
+    requires_confirmation: bool = False
 
 
 class SkillImporter:
@@ -66,6 +74,7 @@ class SkillImporter:
         *,
         with_scripts: bool = False,
         force: bool = False,
+        confirm_dangerous: bool = False,
     ) -> ImportResult:
         """Install *resolved* into ``<target_root>/<source>/<name>/``.
 
@@ -95,12 +104,43 @@ class SkillImporter:
 
         try:
             frontmatter, body = self._read_skill_md(source_md)
-            self._parser.parse_frontmatter(frontmatter, markdown_content=body)
+            manifest = self._parser.parse_frontmatter(frontmatter, markdown_content=body)
         except Exception as exc:
             result.success = False
             result.warnings.append(f"Parse error: {exc}")
             return result
 
+        # 1a. Classify trust and check for dangerous capabilities *before*
+        # writing anything to disk. Everything the importer handles comes from
+        # an external source (github/hermes/openclaw), so the BUNDLED and
+        # WORKSPACE tiers never apply here, and no resolver verifies index
+        # membership yet — a signature alone still classifies as UNREVIEWED.
+        # Community skills get no special treatment just because they came
+        # from a named source.
+        result.trust_tier = classify_trust_tier(
+            has_signature=bool(manifest.signature),
+        )
+        result.dangerous_capabilities = has_dangerous_capabilities(manifest)
+
+        if result.dangerous_capabilities and result.trust_tier == TrustTier.UNREVIEWED:
+            result.requires_confirmation = True
+            if not confirm_dangerous:
+                result.success = False
+                result.warnings.append(
+                    "Refusing to install: this unreviewed skill requests "
+                    f"dangerous capabilities {result.dangerous_capabilities}. "
+                    "Re-run with confirm_dangerous=True (or `--yes-dangerous` "
+                    "on the CLI) only if you trust the source and have "
+                    "reviewed what it does."
+                )
+                return result
+            result.warnings.append(
+                "Installed with dangerous capabilities "
+                f"{result.dangerous_capabilities} — confirmed by caller. "
+                "This skill can run shell commands, open network listeners, "
+                "and/or write to the filesystem."
+            )
+
         # 2. Translate tool references
         translated_body, untranslated = self._translator.translate_markdown(body)
         result.untranslated_tools = untranslated
@@ -180,6 +220,7 @@ class SkillImporter:
         translated_str = ", ".join(f'"{t}"' for t in result.translated_tools)
         missing_str = ", ".join(f'"{t}"' for t in result.untranslated_tools)
         scripts_lower = "true" if result.scripts_imported else "false"
+        dangerous_str = ", ".join(f'"{c}"' for c in result.dangerous_capabilities)
 
         content = (
             f'source = "{resolved.source}:{resolved.name}"\n'
@@ -189,6 +230,8 @@ class SkillImporter:
             f"translated_tools = [{translated_str}]\n"
             f"missing_tools = [{missing_str}]\n"
             f"scripts_imported = {scripts_lower}\n"
+            f'trust_tier = "{result.trust_tier.value}"\n'
+            f"dangerous_capabilities = [{dangerous_str}]\n"
         )
         (target_dir / ".source").write_text(content, encoding="utf-8")
 

src/openjarvis/speech/faster_whisper.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import logging
+import os
 import tempfile
 from typing import List, Optional
 
@@ -105,11 +106,15 @@ class FasterWhisperBackend(SpeechBackend):
         try:
             model = self._ensure_model()
 
-            # Write audio to a temp file (faster-whisper needs a file path)
+            # Write audio to a temp file (faster-whisper needs a file path).
+            # delete=False + manual unlink: on Windows an open
+            # NamedTemporaryFile holds an exclusive handle, so PyAV's reopen
+            # of tmp.name inside model.transcribe() fails with EACCES.
             suffix = f".{format}" if not format.startswith(".") else format
-            with tempfile.NamedTemporaryFile(suffix=suffix, delete=True) as tmp:
-                tmp.write(audio)
-                tmp.flush()
+            tmp = tempfile.NamedTemporaryFile(suffix=suffix, delete=False)
+            try:
+                with tmp:
+                    tmp.write(audio)
 
                 kwargs = {}
                 if language:
@@ -117,6 +122,15 @@ class FasterWhisperBackend(SpeechBackend):
 
                 segments_iter, info = model.transcribe(tmp.name, **kwargs)
                 segments_list = list(segments_iter)
+            finally:
+                try:
+                    os.unlink(tmp.name)
+                except OSError as unlink_exc:
+                    logger.debug(
+                        "Could not remove temp audio file %s: %s",
+                        tmp.name,
+                        unlink_exc,
+                    )
         except Exception as exc:
             self._last_error = str(exc)
             raise

tests/memory/test_sqlite.py

@@ -78,6 +78,19 @@ def test_retrieve_no_results(tmp_path: Path):
     backend.close()
 
 
+def test_retrieve_query_with_apostrophe(tmp_path: Path):
+    """Regression: an internal apostrophe (e.g. "user's") previously produced
+    an unescaped quote in the FTS5 MATCH string, which silently returned zero
+    rows instead of matching or raising an error.
+    """
+    backend = _make_backend(tmp_path)
+    backend.store("The user's name is Trev.", source="identity.md")
+    results = backend.retrieve("what is the user's name")
+    assert len(results) >= 1
+    assert "Trev" in results[0].content
+    backend.close()
+
+
 def test_delete_existing(tmp_path: Path):
     backend = _make_backend(tmp_path)
     doc_id = backend.store("deletable content")

tests/server/test_routes.py

@@ -798,6 +798,40 @@ class TestIdentityPromptInjection:
         assert len(system_msgs) == 1
         assert system_msgs[0].content == "Be terse."
 
+    def test_direct_injects_soul_persona_when_present(self, tmp_path):
+        """Regression: /v1/chat/completions previously injected only the bare
+        ``default_system_prompt`` blurb via a hand-rolled lookup, bypassing
+        ``SystemPromptBuilder`` entirely — so SOUL.md/MEMORY.md/USER.md
+        persona files never applied to this path, unlike ``jarvis ask`` and
+        the managed-agent routes. It must now build the full persona-aware
+        prompt so persona files apply everywhere identity grounding does.
+        """
+        from openjarvis.core.config import MemoryFilesConfig
+
+        soul = tmp_path / "SOUL.md"
+        soul.write_text("Respond with extreme sarcasm and call the user 'champ'.")
+
+        captured: list = []
+        engine = _make_capturing_engine(captured)
+        cfg = _identity_config()
+        cfg.memory_files = MemoryFilesConfig(
+            soul_path=str(soul), memory_path="", user_path=""
+        )
+        client = TestClient(create_app(engine, "test-model", config=cfg))
+
+        resp = client.post(
+            "/v1/chat/completions",
+            json={
+                "model": "test-model",
+                "messages": [{"role": "user", "content": "who are you?"}],
+            },
+        )
+        assert resp.status_code == 200
+        msgs = engine.generate.call_args.args[0]
+        assert msgs[0].role.value == "system"
+        assert "OpenJarvis" in msgs[0].content  # identity blurb still present
+        assert "extreme sarcasm" in msgs[0].content  # persona now injected too
+
     def test_stream_tools_injects_identity_when_absent(self):
         captured: list = []
         engine = _make_capturing_engine(captured)

tests/skills/test_importer.py

@@ -202,3 +202,75 @@ class TestImportSkill:
 
         installed = target_root / "hermes" / "my-skill" / "SKILL.md"
         assert "Original" in installed.read_text()
+
+
+class TestDangerousCapabilityGate:
+    def _make_importer(self, tmp_path: Path) -> SkillImporter:
+        return SkillImporter(
+            parser=SkillParser(),
+            tool_translator=ToolTranslator(),
+            target_root=tmp_path / "skills",
+        )
+
+    def _make_resolved_with_caps(
+        self, tmp_path: Path, caps: list[str]
+    ) -> ResolvedSkill:
+        src_dir = tmp_path / "source" / "my-skill"
+        src_dir.mkdir(parents=True)
+        caps_yaml = "".join(f"  - {c}\n" for c in caps)
+        (src_dir / "SKILL.md").write_text(
+            "---\n"
+            "name: my-skill\n"
+            "description: A test skill\n"
+            f"required_capabilities:\n{caps_yaml}"
+            "---\n"
+            "Body"
+        )
+        return ResolvedSkill(
+            name="my-skill",
+            source="hermes",
+            path=src_dir,
+            category="testing",
+            description="A test skill",
+            commit="abc123",
+        )
+
+    def test_refuses_unreviewed_dangerous_skill(self, tmp_path: Path):
+        importer = self._make_importer(tmp_path)
+        resolved = self._make_resolved_with_caps(tmp_path, ["shell:execute"])
+        result = importer.import_skill(resolved)
+
+        assert not result.success
+        assert result.requires_confirmation
+        assert result.dangerous_capabilities == ["shell:execute"]
+        assert any("dangerous" in w.lower() for w in result.warnings)
+        # Nothing may be written to disk on refusal
+        assert not (tmp_path / "skills" / "hermes" / "my-skill").exists()
+
+    def test_confirm_dangerous_installs_and_records_tier(self, tmp_path: Path):
+        importer = self._make_importer(tmp_path)
+        resolved = self._make_resolved_with_caps(tmp_path, ["shell:execute"])
+        result = importer.import_skill(resolved, confirm_dangerous=True)
+
+        assert result.success
+        assert result.requires_confirmation
+        assert any("confirmed by caller" in w for w in result.warnings)
+        content = (
+            tmp_path / "skills" / "hermes" / "my-skill" / ".source"
+        ).read_text()
+        assert 'trust_tier = "unreviewed"' in content
+        assert 'dangerous_capabilities = ["shell:execute"]' in content
+
+    def test_benign_capabilities_need_no_confirmation(self, tmp_path: Path):
+        importer = self._make_importer(tmp_path)
+        resolved = self._make_resolved_with_caps(tmp_path, ["network:fetch"])
+        result = importer.import_skill(resolved)
+
+        assert result.success
+        assert not result.requires_confirmation
+        assert result.dangerous_capabilities == []
+        content = (
+            tmp_path / "skills" / "hermes" / "my-skill" / ".source"
+        ).read_text()
+        assert 'trust_tier = "unreviewed"' in content
+        assert "dangerous_capabilities = []" in content

tests/skills/test_skills.py

@@ -148,6 +148,58 @@ class TestSkillExecutor:
         assert EventType.SKILL_EXECUTE_END in event_types
 
 
+class TestSkillExecutorCapabilities:
+    def _manifest(self):
+        return SkillManifest(
+            name="capskill",
+            required_capabilities=["network:fetch"],
+            steps=[
+                SkillStep(
+                    tool_name="echo",
+                    arguments_template='{"text": "hello"}',
+                    output_key="result",
+                )
+            ],
+        )
+
+    def test_no_policy_runs_capability_skills(self):
+        """Default construction (no allowed_capabilities) must not enforce —
+        this is the pre-enforcement behavior every manager.py call site relies on."""
+        executor = SkillExecutor(ToolExecutor([EchoTool()]))
+        result = executor.run(self._manifest())
+        assert result.success
+        assert result.context.get("result") == "hello"
+
+    def test_policy_blocks_missing_capability(self):
+        executor = SkillExecutor(
+            ToolExecutor([EchoTool()]), allowed_capabilities=set()
+        )
+        result = executor.run(self._manifest())
+        assert not result.success
+        assert len(result.step_results) == 1
+        assert "Blocked" in result.step_results[0].content
+        assert "network:fetch" in result.step_results[0].content
+
+    def test_policy_allows_granted_capability(self):
+        executor = SkillExecutor(
+            ToolExecutor([EchoTool()]),
+            allowed_capabilities={"network:fetch"},
+        )
+        result = executor.run(self._manifest())
+        assert result.success
+        assert result.context.get("result") == "hello"
+
+    def test_blocked_run_publishes_events(self):
+        bus = EventBus(record_history=True)
+        executor = SkillExecutor(
+            ToolExecutor([EchoTool()]), bus=bus, allowed_capabilities=set()
+        )
+        executor.run(self._manifest())
+        event_types = {e.event_type for e in bus.history}
+        assert EventType.SKILL_EXECUTE_START in event_types
+        assert EventType.SKILL_EXECUTE_END in event_types
+
+
 class TestSkillStepExtended:
     def test_step_with_skill_name(self):
         step = SkillStep(skill_name="summarize", output_key="result")

tests/speech/test_faster_whisper.py

@@ -53,6 +53,69 @@ def test_faster_whisper_transcribe():
         assert result.duration_seconds == 1.5
 
 
+def test_faster_whisper_transcribe_temp_file_reopenable_and_removed():
+    """The temp file must be closed before the model reads it, and gone after.
+
+    On Windows, an open NamedTemporaryFile holds an exclusive handle, so
+    PyAV's reopen of the path inside model.transcribe() fails with EACCES
+    unless the file is closed first. Opening the path inside the mocked
+    transcribe reproduces that failure mode on Windows.
+    """
+    import os
+
+    mock_info = MagicMock()
+    mock_info.language = "en"
+    mock_info.language_probability = 0.95
+    mock_info.duration = 1.5
+
+    seen = {}
+
+    def fake_transcribe(path, **kwargs):
+        seen["path"] = path
+        with open(path, "rb") as fh:
+            seen["content"] = fh.read()
+        return iter(()), mock_info
+
+    mock_model = MagicMock()
+    mock_model.transcribe.side_effect = fake_transcribe
+
+    with patch(
+        "openjarvis.speech.faster_whisper.WhisperModel",
+        return_value=mock_model,
+    ):
+        backend = FasterWhisperBackend(model_size="base", device="cpu")
+        backend.transcribe(b"fake audio bytes")
+
+    assert seen["content"] == b"fake audio bytes"
+    assert not os.path.exists(seen["path"])
+
+
+def test_faster_whisper_transcribe_removes_temp_file_on_error():
+    """The temp file is cleaned up even when transcription fails."""
+    import os
+
+    seen = {}
+
+    def fake_transcribe(path, **kwargs):
+        seen["path"] = path
+        raise RuntimeError("decode failed")
+
+    mock_model = MagicMock()
+    mock_model.transcribe.side_effect = fake_transcribe
+
+    with patch(
+        "openjarvis.speech.faster_whisper.WhisperModel",
+        return_value=mock_model,
+    ):
+        backend = FasterWhisperBackend(model_size="base", device="cpu")
+        with pytest.raises(RuntimeError, match="decode failed"):
+            backend.transcribe(b"fake audio bytes")
+
+    assert "path" in seen
+    assert not os.path.exists(seen["path"])
+    assert "decode failed" in (backend.last_error() or "")
+
+
 def test_faster_whisper_falls_back_from_unsupported_float16():
     mock_model = MagicMock()