aria-code 0.100.0__tar.gz → 0.101.0__tar.gz

{aria_code-0.100.0/src/aria_code.egg-info → aria_code-0.101.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aria-code
-Version: 0.100.0
+Version: 0.101.0
 Summary: Open-source AI agent for coding, financial analysis and logistics operations: numbers you can check, client data kept apart, actions a person approves
 License: Apache-2.0
 Project-URL: Homepage, https://github.com/artheras/aria-code

{aria_code-0.100.0 → aria_code-0.101.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "aria-code"
-version = "0.100.0"
+version = "0.101.0"
 description = "Open-source AI agent for coding, financial analysis and logistics operations: numbers you can check, client data kept apart, actions a person approves"
 readme = "README.md"
 license = { text = "Apache-2.0" }

{aria_code-0.100.0 → aria_code-0.101.0}/src/aria_code/_version.py

@@ -17,4 +17,4 @@ from __future__ import annotations
 
 __all__ = ["__version__"]
 
-__version__ = "0.100.0"
+__version__ = "0.101.0"

{aria_code-0.100.0 → aria_code-0.101.0}/src/aria_code/apps/cli/commands/portfolio_cmds.py

@@ -431,6 +431,7 @@ class PortfolioCommandsMixin:
                 data_quality=data_quality,
                 data_bundle=data_bundle,
                 now=_dt.now(),
+                lang=ui_text(self, "zh", "en"),
             )
 
             await self.terminal.send_message(ai_prompt)

{aria_code-0.100.0 → aria_code-0.101.0}/src/aria_code/apps/cli/commands/report.py

@@ -273,24 +273,29 @@ def metric_line(label: str, value: Any, digits: int = 2, suffix: str = "") -> st
     return f"- {label}: {rendered}" if rendered != "-" else ""
 
 
-def markdown_data_block(market_data: dict[str, Any]) -> str:
+def markdown_data_block(market_data: dict[str, Any], lang: str = "zh") -> str:
+    en = lang == "en"
     data_lines = [
-        metric_line("当前价", market_data.get("price")),
-        metric_line("涨跌", market_data.get("change_pct"), suffix="%"),
+        metric_line("Price" if en else "当前价", market_data.get("price")),
+        metric_line("Change" if en else "涨跌", market_data.get("change_pct"), suffix="%"),
         metric_line("RSI(14)", market_data.get("rsi")),
         metric_line("MACD", market_data.get("macd"), digits=4),
         metric_line("MA20", market_data.get("ma20")),
         metric_line("MA60", market_data.get("ma60")),
-        metric_line("布林上轨", market_data.get("bb_upper")),
-        metric_line("布林下轨", market_data.get("bb_lower")),
+        metric_line("Bollinger upper" if en else "布林上轨", market_data.get("bb_upper")),
+        metric_line("Bollinger lower" if en else "布林下轨", market_data.get("bb_lower")),
     ]
     data_block = "\n".join(line for line in data_lines if line)
     if data_block:
         return data_block
+    if en:
+        return ("- Live market data is unavailable; the report must state this limitation and must not "
+                "invent prices or indicators.")
     return "- 实时行情数据暂不可用；报告必须明确说明数据限制，不得编造价格或指标。"
 
 
-def markdown_provenance_block(data_quality: dict[str, Any], data_bundle: Any = None) -> str:
+def markdown_provenance_block(data_quality: dict[str, Any], data_bundle: Any = None, lang: str = "zh") -> str:
+    en = lang == "en"
     provider_chain = (
         data_quality.get("providers")
         or getattr(data_bundle, "provider_chain", [])
@@ -301,16 +306,26 @@ def markdown_provenance_block(data_quality: dict[str, Any], data_bundle: Any = N
         or data_quality.get("missing_fields")
         or []
     )
+    status = data_quality.get('status', getattr(data_bundle, 'status', 'unknown') if data_bundle else 'unknown')
+    labels = (("Data status", "Stale", "Provider chain", "Missing fields") if en
+              else ("数据状态", "是否过期", "数据源链", "缺失字段"))
     lines = [
-        f"- 数据状态: {data_quality.get('status', getattr(data_bundle, 'status', 'unknown') if data_bundle else 'unknown')}",
-        f"- 是否过期: {'yes' if data_quality.get('stale') else 'no'}",
-        f"- 数据源链: {', '.join(provider_chain) if provider_chain else 'unknown'}",
-        f"- 缺失字段: {', '.join(missing_fields) if missing_fields else 'none'}",
+        f"- {labels[0]}: {status}",
+        f"- {labels[1]}: {'yes' if data_quality.get('stale') else 'no'}",
+        f"- {labels[2]}: {', '.join(provider_chain) if provider_chain else 'unknown'}",
+        f"- {labels[3]}: {', '.join(missing_fields) if missing_fields else 'none'}",
     ]
     return "\n".join(lines)
 
 
-def report_depth_description(report_type: str) -> str:
+def report_depth_description(report_type: str, lang: str = "zh") -> str:
+    if lang == "en":
+        if report_type == "deep":
+            return ("deep edition (about 8 pages): valuation (DCF and relative), three years of P&L, "
+                    "management, and the competitive landscape")
+        if report_type == "brief":
+            return "brief edition: one page, the core view, key numbers and a one-sentence conclusion"
+        return "standard edition: cover, technical analysis, fundamentals overview, risk factors"
     if report_type == "deep":
         return "深度（8页）版本：包含估值模型（DCF + 相对估值）、财务分析（3年P&L）、管理层分析、行业竞争格局"
     if report_type == "brief":
@@ -326,11 +341,35 @@ def build_markdown_report_prompt(
     data_quality: dict[str, Any],
     data_bundle: Any = None,
     now: datetime | None = None,
+    lang: str = "zh",
 ) -> str:
+    """The prompt for a Markdown research report, in the session's language.
+
+    It was Chinese only, ending "用中文输出": an English session got a Chinese
+    report, and when the model was unavailable the fallback analysis followed
+    the prompt into Chinese too.
+    """
     report_date = (now or datetime.now()).strftime("%Y-%m-%d")
-    data_block = markdown_data_block(market_data)
-    provenance_block = markdown_provenance_block(data_quality, data_bundle)
-    depth = report_depth_description(report_type)
+    data_block = markdown_data_block(market_data, lang)
+    provenance_block = markdown_provenance_block(data_quality, data_bundle, lang)
+    depth = report_depth_description(report_type, lang)
+    if lang == "en":
+        return (
+            f"Write a professional Markdown research report on {symbol} ({depth}).\n\n"
+            f"**Live data (use only the fields below; do not fill in missing ones)**:\n"
+            f"{data_block}\n\n"
+            f"**Data quality (state it plainly in the report)**:\n"
+            f"{provenance_block}\n\n"
+            f"Report structure (Markdown):\n"
+            f"# {symbol} Research Report\n"
+            f"**Rating**: Buy/Neutral/Reduce  **Target price**: X.XX  **Date**: {report_date}\n\n"
+            f"## Key points\n"
+            f"## Technical analysis\n"
+            f"## Fundamentals\n"
+            f"## Risk factors\n"
+            f"## Recommendation\n\n"
+            f"Use the real data, no placeholders; where data is missing, say what is missing. Write in English."
+        )
     return (
         f"为 {symbol} 生成一份专业 Markdown 投研报告（{depth}）。\n\n"
         f"**实时数据（仅使用下列已返回字段；缺失字段不要补写）**：\n"

{aria_code-0.100.0 → aria_code-0.101.0}/src/aria_code/apps/cli/providers/vertexai_stream.py

@@ -27,6 +27,36 @@ _MISSING_SDK_MESSAGE = (
 )
 
 
+import logging
+
+logger = logging.getLogger(__name__)
+
+_EMPTY_ROUND_RETRIES = 2
+
+
+def _chunk_text(chunk) -> str:
+    """The text parts of a chunk, read without chunk.text.
+
+    chunk.text logs "there are non-text parts in the response: ['function_call']"
+    for every chunk that carries a call, which filled the eval logs.
+    """
+    out = []
+    for candidate in (getattr(chunk, "candidates", None) or [])[:1]:
+        content = getattr(candidate, "content", None)
+        for part in (getattr(content, "parts", None) or []):
+            text = getattr(part, "text", None)
+            if text and not getattr(part, "thought", False):
+                out.append(text)
+    if out:
+        return "".join(out)
+    if not getattr(chunk, "candidates", None):
+        try:
+            return chunk.text or ""
+        except Exception:
+            return ""
+    return ""
+
+
 class VertexAIProvider(LLMProvider):
     """Native Vertex AI provider using google-genai."""
     
@@ -265,41 +295,66 @@ class VertexAIProvider(LLMProvider):
         )
         
         try:
-            # We must use asyncio.to_thread because the google-genai async client might not be used here
-            # Or we can just use the sync stream generator wrapped in an async generator.
-            # google-genai provides AsyncClient as well:
-            # async_client = genai.Client(vertexai=True).aio
-            response_stream = await client.aio.models.generate_content_stream(
-                model=self.model,
-                contents=contents,
-                config=config,
-            )
-            
+            # Gemini sometimes ends a round with no text and no function call:
+            # finish_reason MALFORMED_FUNCTION_CALL (it wrote a call it could
+            # not emit), or an empty STOP after reading tool results. The round
+            # then came back empty, the turn ended as "empty_response", and
+            # `aria-code -p` exited 1 a few seconds in: three of nine eval tasks
+            # on the first Vertex runs, different ones each time. The same
+            # request usually succeeds when sent again, so it is, up to twice,
+            # and if it is still empty the finish reason goes into the error.
             full_response = ""
             usage = {}
             tool_calls = []
-            
-            async for chunk in response_stream:
-                if cancel_event and cancel_event.is_set():
-                    yield LLMDone(response=full_response, provider="vertexai", success=True, cancelled=True)
-                    return
-                    
-                if chunk.text:
-                    full_response += chunk.text
-                    yield LLMToken(text=chunk.text)
-                    
-                if chunk.function_calls:
-                    for fc in chunk.function_calls:
-                        args = {k: v for k, v in fc.args.items()} if fc.args else {}
-                        yield LLMToolCall(tool=fc.name, params=args)
-                        tool_calls.append({"tool": fc.name, "params": args})
-                        
-                if chunk.usage_metadata:
-                    usage = {
-                        "prompt_tokens": chunk.usage_metadata.prompt_token_count,
-                        "completion_tokens": chunk.usage_metadata.candidates_token_count,
-                    }
-                    
+            finish_reasons: list = []
+            for attempt in range(1 + _EMPTY_ROUND_RETRIES):
+                response_stream = await client.aio.models.generate_content_stream(
+                    model=self.model,
+                    contents=contents,
+                    config=config,
+                )
+                async for chunk in response_stream:
+                    if cancel_event and cancel_event.is_set():
+                        yield LLMDone(response=full_response, provider="vertexai", success=True, cancelled=True)
+                        return
+
+                    text = _chunk_text(chunk)
+                    if text:
+                        full_response += text
+                        yield LLMToken(text=text)
+
+                    if chunk.function_calls:
+                        for fc in chunk.function_calls:
+                            args = {k: v for k, v in fc.args.items()} if fc.args else {}
+                            yield LLMToolCall(tool=fc.name, params=args)
+                            tool_calls.append({"tool": fc.name, "params": args})
+
+                    for candidate in (getattr(chunk, "candidates", None) or []):
+                        reason = getattr(candidate, "finish_reason", None)
+                        if reason:
+                            finish_reasons.append(str(getattr(reason, "name", reason)))
+
+                    if chunk.usage_metadata:
+                        usage = {
+                            "prompt_tokens": chunk.usage_metadata.prompt_token_count,
+                            "completion_tokens": chunk.usage_metadata.candidates_token_count,
+                        }
+                if full_response.strip() or tool_calls:
+                    break
+                logger.info("Vertex returned an empty round (finish_reason %s), attempt %d",
+                            finish_reasons[-1:] or "none", attempt + 1)
+                if attempt < _EMPTY_ROUND_RETRIES:
+                    finish_reasons = []
+
+            if not full_response.strip() and not tool_calls:
+                reason = finish_reasons[-1] if finish_reasons else "none"
+                yield LLMDone(
+                    response="", provider="vertexai", success=False,
+                    error=f"empty_response (Vertex finish_reason {reason} after "
+                          f"{1 + _EMPTY_ROUND_RETRIES} attempts)",
+                )
+                return
+
             yield LLMDone(
                 response=full_response,
                 tool_calls_pending=tool_calls,
@@ -308,7 +363,7 @@ class VertexAIProvider(LLMProvider):
                 success=True,
                 cancelled=False,
             )
-            
+
         except APIError as e:
             yield LLMDone(response="", provider="vertexai", success=False, error=f"Vertex AI API Error: {e.message}")
         except Exception as e:

{aria_code-0.100.0 → aria_code-0.101.0}/src/aria_code/evals/harness.py

@@ -55,6 +55,7 @@ from __future__ import annotations
 
 import importlib.util
 import json
+import re
 import os
 import shlex
 import shutil
@@ -155,6 +156,26 @@ class TaskSpec:
         )
 
 
+_SECRET = re.compile(
+    r"(?i)\b(api[_-]?key|token|password|secret)(\s*[=:]\s*)\S+"
+    r"|\b(?:bearer)\s+\S+"
+    r"|\b(?:sk|pk|rk)-[A-Za-z0-9_-]{8,}|\bgh[pousr]_[A-Za-z0-9]{12,}|\bAIza[0-9A-Za-z_-]{20,}"
+    r"|\bya29\.[0-9A-Za-z_-]+"
+)
+
+
+def _mask(line: str) -> str:
+    def repl(match: re.Match) -> str:
+        return f"{match.group(1)}{match.group(2)}***" if match.group(1) else "***"
+    return _SECRET.sub(repl, line)
+
+
+def _log_tail(log: str, lines: int = 23) -> list:
+    """The last lines of a task's log, secret-shaped values masked."""
+    kept = [line.rstrip() for line in str(log or "").splitlines() if line.strip()]
+    return [_mask(line)[:300] for line in kept[-lines:]]
+
+
 @dataclass(frozen=True)
 class TaskResult:
     task_id: str
@@ -185,6 +206,9 @@ class TaskResult:
             "detail": self.detail,
             "tags": list(self.tags),
             "changed": list(self.changed),
+            # Why it did not pass. The first CI run's report said only "the
+            # agent did not complete (exit 1)"; the reason was in the job log.
+            "log_tail": _log_tail(self.log) if self.outcome != PASS else [],
         }
 
 
@@ -579,9 +603,21 @@ def run_task(
         detail = f"`{task.verify}` exited {after_code}"
         if not changed:
             detail += " — the agent changed nothing"
+        # The check's output says what is wrong; the agent's last words say
+        # what it thought it did. inventory-reorder failed twice on "reorder.json
+        # was not written" with nothing to show whether the agent computed the
+        # policy and only reported it, or never got that far.
+        agent_said = "\n".join(
+            part for part in (str(getattr(outcome, "stdout", "") or ""), str(getattr(outcome, "stderr", "") or ""))
+            if part.strip()
+        )
+        agent_tail = "\n".join([line for line in agent_said.splitlines() if line.strip()][-8:])
+        log = _trim(after_log)
+        if agent_tail:
+            log = f"{log}\n--- agent, last lines ---\n{agent_tail}"
         return _result(
             FAIL, exit_code=after_code, detail=detail,
-            log=_trim(after_log), changed=changed,
+            log=log, changed=changed,
         )
     finally:
         if not keep_workspace and scratch_root is None:

aria_code-0.101.0/src/aria_code/evals/summary.py

@@ -0,0 +1,75 @@
+"""Turn eval reports into a Markdown scoreboard, and decide whether the run is healthy.
+
+    python3 -m aria_code.evals.summary evals-core.json evals-operations.json
+
+The suites ran in CI only with ``--check``: every fixture still starts red,
+but no model ever attempted them, so nothing measured whether Aria can fix
+code. The weekly workflow (.github/workflows/evals.yml) runs them with a real
+model and writes this summary.
+
+A task the model gets wrong is a result, not a broken build: the exit code is
+1 only when a task could not run (ERROR) or no longer measures anything
+(INVALID), the two outcomes that mean the harness, not the model, needs work.
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+
+def summarise(reports: list[dict], model: str = "") -> tuple[str, int]:
+    """Markdown for the reports, and the exit code the run should have."""
+    lines = [f"## Aria eval scoreboard{f' · {model}' if model else ''}", ""]
+    unhealthy = 0
+    for report in reports:
+        suite = report.get("suite", "?")
+        lines.append(f"### {suite}: {report.get('passed', 0)}/{report.get('scored', 0)} "
+                     f"({report.get('pass_rate', 0):.0%})")
+        extras = [f"{report[k]} {k}" for k in ("errored", "invalid") if report.get(k)]
+        if extras:
+            lines.append(f"Not scored: {', '.join(extras)}.")
+        unhealthy += int(report.get("errored", 0)) + int(report.get("invalid", 0))
+        lines += ["", "| Task | Outcome | Seconds | Detail |", "|---|---|---:|---|"]
+        for result in report.get("results", []):
+            detail = str(result.get("detail") or "").replace("|", "\\|")[:120]
+            lines.append(f"| {result.get('task_id')} | {result.get('outcome')} | "
+                         f"{result.get('seconds', 0):.0f} | {detail} |")
+        for result in report.get("results", []):
+            tail = result.get("log_tail") or []
+            if tail and str(result.get("outcome", "")).lower() != "pass":
+                body = "\n".join(tail).replace("```", "` ` `")
+                lines += ["", f"<details><summary>{result.get('task_id')}: last lines of the log</summary>",
+                          "", "```", body, "```", "</details>"]
+        by_tag = report.get("by_tag") or {}
+        if by_tag:
+            tags = ", ".join(f"{tag} {s['passed']}/{s['scored']}" for tag, s in by_tag.items() if s.get("scored"))
+            if tags:
+                lines += ["", f"By tag: {tags}"]
+        lines.append("")
+    return "\n".join(lines), (1 if unhealthy else 0)
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = list(sys.argv[1:] if argv is None else argv)
+    model = ""
+    if args[:1] == ["--model"] and len(args) > 1:
+        model, args = args[1], args[2:]
+    reports = []
+    for name in args:
+        path = Path(name)
+        if path.exists():
+            reports.append(json.loads(path.read_text(encoding="utf-8")))
+        else:
+            print(f"(no report at {name})", file=sys.stderr)
+    if not reports:
+        print("no eval reports to summarise", file=sys.stderr)
+        return 1
+    text, code = summarise(reports, model)
+    print(text)
+    return code
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

{aria_code-0.100.0 → aria_code-0.101.0/src/aria_code.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aria-code
-Version: 0.100.0
+Version: 0.101.0
 Summary: Open-source AI agent for coding, financial analysis and logistics operations: numbers you can check, client data kept apart, actions a person approves
 License: Apache-2.0
 Project-URL: Homepage, https://github.com/artheras/aria-code

{aria_code-0.100.0 → aria_code-0.101.0}/src/aria_code.egg-info/SOURCES.txt

@@ -336,6 +336,7 @@ src/aria_code/domain/strategy_vault.py
 src/aria_code/evals/__init__.py
 src/aria_code/evals/harness.py
 src/aria_code/evals/runner.py
+src/aria_code/evals/summary.py
 src/aria_code/packages/__init__.py
 src/aria_code/packages/adk_bridge/__init__.py
 src/aria_code/packages/adk_bridge/code_review_tools.py
@@ -583,6 +584,7 @@ tests/test_engine_boundary.py
 tests/test_english_output_group.py
 tests/test_enterprise_tool_honesty.py
 tests/test_env_switch_docs.py
+tests/test_eval_summary.py
 tests/test_evals_harness.py
 tests/test_event_import_roots.py
 tests/test_exec_events.py
@@ -690,6 +692,7 @@ tests/test_release_verification_waits.py
 tests/test_release_workflow_permissions.py
 tests/test_report_charts.py
 tests/test_report_generator.py
+tests/test_report_prompt_language.py
 tests/test_research_protocol.py
 tests/test_research_quality.py
 tests/test_responsive_rendering.py
@@ -758,6 +761,7 @@ tests/test_verification_hint.py
 tests/test_verification_policy.py
 tests/test_version_consistency.py
 tests/test_version_counter.py
+tests/test_vertex_empty_rounds.py
 tests/test_video_analysis.py
 tests/test_video_editor.py
 tests/test_vscode_mcp.py

aria_code-0.101.0/tests/test_eval_summary.py

@@ -0,0 +1,75 @@
+"""The weekly eval scoreboard, and when a run counts as broken.
+
+A task the model gets wrong is a measurement; the run is unhealthy only when a
+task could not run (ERROR) or no longer measures anything (INVALID).
+"""
+
+from __future__ import annotations
+
+import json
+
+import yaml
+
+from aria_code.evals.summary import main, summarise
+
+REPORT = {
+    "suite": "core", "pass_rate": 0.6667, "passed": 2, "failed": 1, "invalid": 0, "errored": 0, "scored": 3,
+    "by_tag": {"python": {"passed": 2, "scored": 3, "pass_rate": 0.6667}},
+    "results": [
+        {"task_id": "fix-failing-test", "outcome": "PASS", "seconds": 81.2, "detail": ""},
+        {"task_id": "off-by-one", "outcome": "PASS", "seconds": 64.0, "detail": ""},
+        {"task_id": "reconcile-waybills", "outcome": "FAIL", "seconds": 140.5, "detail": "edited a | protected file"},
+    ],
+}
+
+
+def test_a_wrong_answer_is_a_result_not_a_broken_run():
+    text, code = summarise([REPORT], "google/gemini-2.5-flash")
+    assert code == 0
+    assert "## Aria eval scoreboard · google/gemini-2.5-flash" in text
+    assert "### core: 2/3 (67%)" in text
+    assert "| reconcile-waybills | FAIL | 140 | edited a \\| protected file |" in text
+    assert "By tag: python 2/3" in text
+
+
+def test_errored_or_invalid_tasks_fail_the_run():
+    for key in ("errored", "invalid"):
+        text, code = summarise([dict(REPORT, **{key: 1})])
+        assert code == 1 and f"Not scored: 1 {key}." in text
+
+
+def test_the_cli_reads_report_files(tmp_path, capsys):
+    path = tmp_path / "evals-core.json"
+    path.write_text(json.dumps(REPORT))
+    assert main(["--model", "m", str(path), str(tmp_path / "missing.json")]) == 0
+    assert "### core: 2/3" in capsys.readouterr().out
+    assert main([str(tmp_path / "missing.json")]) == 1
+
+
+def test_the_workflow_signs_in_without_a_key():
+    workflow = yaml.safe_load(open(".github/workflows/evals.yml"))
+    job = workflow["jobs"]["evals"]
+    assert workflow["permissions"]["id-token"] == "write"
+    uses = [step.get("uses", "") for step in job["steps"]]
+    assert any(u.startswith("google-github-actions/auth@") for u in uses)
+    assert "GEMINI_API_KEY" not in open(".github/workflows/evals.yml").read()
+    assert job["env"]["GOOGLE_GENAI_USE_VERTEXAI"] == "true"
+
+
+def test_the_reason_a_task_did_not_pass_reaches_the_summary():
+    """The first CI run said only "the agent did not complete (exit 1)"; the reason was in the job log."""
+    from aria_code.evals.harness import ERROR, PASS, TaskResult
+
+    errored = TaskResult(task_id="stripe-refund-validation", outcome=ERROR, exit_code=1,
+                         detail="the agent did not complete (exit 1); check still red",
+                         log="starting\nAuthorization: Bearer abc.def\nVertex 400: INVALID_ARGUMENT")
+    passed = TaskResult(task_id="off-by-one", outcome=PASS, log="12 passed")
+    assert errored.to_dict()["log_tail"] == ["starting", "Authorization: ***", "Vertex 400: INVALID_ARGUMENT"]
+    assert passed.to_dict()["log_tail"] == []
+
+    report = dict(REPORT, results=[errored.to_dict(), passed.to_dict()], errored=1)
+    text, code = summarise([report])
+    assert code == 1
+    assert "<details><summary>stripe-refund-validation: last lines of the log</summary>" in text
+    assert "Vertex 400: INVALID_ARGUMENT" in text and "abc.def" not in text
+    assert "off-by-one: last lines" not in text

{aria_code-0.100.0 → aria_code-0.101.0}/tests/test_evals_harness.py

@@ -618,3 +618,28 @@ class RunResidueTests(HarnessBase):
             (workspace / "fixed.txt").write_text("x", encoding="utf-8")
 
         self.assertEqual(self._run(self._task(), solver).changed, ("fixed.txt",))
+
+
+class AgentTailTests(HarnessBase):
+    """A failed task's log ends with what the agent said, not only the check's output.
+
+    inventory-reorder failed twice on "reorder.json was not written", with
+    nothing showing whether the agent computed the policy and only reported it.
+    """
+
+    def test_a_failed_task_keeps_the_agents_last_lines(self):
+        from types import SimpleNamespace
+
+        _fixture(self.fixtures, "broken", {"check.py": _GUARD})
+        said = SimpleNamespace(returncode=0, stdout="Computed reorder points for 6 SKUs.\nA100: order 335",
+                               stderr="")
+        result = self._run(self._task(), lambda p, w: said)
+        self.assertEqual(result.outcome, FAIL)
+        self.assertIn("--- agent, last lines ---", result.log)
+        self.assertIn("A100: order 335", result.log)
+        self.assertIn("A100: order 335", result.to_dict()["log_tail"][-1])
+
+    def test_a_silent_agent_adds_nothing(self):
+        _fixture(self.fixtures, "broken", {"check.py": _GUARD})
+        result = self._run(self._task(), lambda p, w: None)
+        self.assertNotIn("--- agent", result.log)

{aria_code-0.100.0 → aria_code-0.101.0}/tests/test_output_rendering.py

@@ -385,7 +385,8 @@ def test_report_markdown_prompt_omits_na_placeholders(monkeypatch, tmp_path):
 
     class FakeTerminal:
         conversation = [{"role": "assistant", "content": "# Report\nok"}]
-        config = {}
+        # The report prompt follows ui_lang; this test checks the Chinese labels.
+        config = {"ui_lang": "zh"}
 
         async def send_message(self, prompt):
             prompts.append(prompt)

aria_code-0.101.0/tests/test_report_prompt_language.py

@@ -0,0 +1,43 @@
+"""/report asks for the report in the session's language.
+
+The prompt was Chinese only and ended "用中文输出", so an English session got a
+Chinese report; with the model unavailable, the Chinese prompt matched the
+chart-analysis route and a Chinese chart summary was saved as the report.
+"""
+
+from __future__ import annotations
+
+import re
+from datetime import datetime
+
+from aria_code.apps.cli.commands.report import build_markdown_report_prompt
+
+CJK = re.compile(r"[一-鿿]")
+DATA = {"price": 529.3, "change_pct": 0.4, "rsi": 76.3, "macd": 9.51, "ma20": 504.7, "ma60": 493.4}
+QUALITY = {"status": "complete", "stale": False, "providers": ["yfinance"]}
+
+
+def _prompt(lang, **extra):
+    return build_markdown_report_prompt(symbol="MSFT", report_type="standard", market_data=DATA,
+                                        data_quality=QUALITY, now=datetime(2026, 10, 7), lang=lang, **extra)
+
+
+def test_english_prompt():
+    text = _prompt("en")
+    assert not CJK.findall(text), CJK.findall(text)
+    assert "# MSFT Research Report" in text and "**Date**: 2026-10-07" in text
+    assert "Price" in text and "529.3" in text
+    assert "use only the fields below" in text and "Write in English." in text
+    assert "- Provider chain: yfinance" in text
+
+
+def test_missing_data_is_named_in_english():
+    text = build_markdown_report_prompt(symbol="MSFT", report_type="brief", market_data={},
+                                        data_quality={}, now=datetime(2026, 10, 7), lang="en")
+    assert "must not invent prices or indicators" in text and "brief edition" in text
+
+
+def test_chinese_is_unchanged_by_default():
+    text = build_markdown_report_prompt(symbol="MSFT", report_type="standard", market_data=DATA,
+                                        data_quality=QUALITY, now=datetime(2026, 10, 7))
+    assert text.startswith("为 MSFT 生成一份专业 Markdown 投研报告") and text.endswith("用中文输出。")