aria-code 0.113.0__tar.gz → 0.115.0__tar.gz

{aria_code-0.113.0/src/aria_code.egg-info → aria_code-0.115.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aria-code
-Version: 0.113.0
+Version: 0.115.0
 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
@@ -171,7 +171,7 @@ Numbers you can check, client data kept apart, actions a person approves.</p>
 <p align="center">
   <a href="https://www.npmjs.com/package/@artheras/aria-code"><img src="https://img.shields.io/npm/v/@artheras/aria-code/latest?style=flat-square&logo=npm&label=npm" alt="npm version"></a>
   <!-- Release automation pins PyPI: the default endpoint still selects legacy 4.x. -->
-  <a href="https://pypi.org/project/aria-code/0.113.0/"><img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fpypi.org%2Fpypi%2Faria-code%2F0.113.0%2Fjson&amp;query=%24.info.version&amp;prefix=v&amp;style=flat-square&amp;logo=pypi&amp;label=PyPI&amp;color=blue&amp;cacheSeconds=300" alt="PyPI version"></a>
+  <a href="https://pypi.org/project/aria-code/0.115.0/"><img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fpypi.org%2Fpypi%2Faria-code%2F0.115.0%2Fjson&amp;query=%24.info.version&amp;prefix=v&amp;style=flat-square&amp;logo=pypi&amp;label=PyPI&amp;color=blue&amp;cacheSeconds=300" alt="PyPI version"></a>
   <a href="https://github.com/artheras/aria-code/actions/workflows/ci.yml"><img src="https://img.shields.io/github/actions/workflow/status/artheras/aria-code/ci.yml?branch=main&style=flat-square&logo=githubactions&label=CI" alt="CI status"></a>
   <a href="LICENSE"><img src="https://img.shields.io/badge/license-Apache%202.0-64748b?style=flat-square" alt="Apache License 2.0"></a>
 </p>

{aria_code-0.113.0 → aria_code-0.115.0}/README.md

@@ -15,7 +15,7 @@ Numbers you can check, client data kept apart, actions a person approves.</p>
 <p align="center">
   <a href="https://www.npmjs.com/package/@artheras/aria-code"><img src="https://img.shields.io/npm/v/@artheras/aria-code/latest?style=flat-square&logo=npm&label=npm" alt="npm version"></a>
   <!-- Release automation pins PyPI: the default endpoint still selects legacy 4.x. -->
-  <a href="https://pypi.org/project/aria-code/0.113.0/"><img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fpypi.org%2Fpypi%2Faria-code%2F0.113.0%2Fjson&amp;query=%24.info.version&amp;prefix=v&amp;style=flat-square&amp;logo=pypi&amp;label=PyPI&amp;color=blue&amp;cacheSeconds=300" alt="PyPI version"></a>
+  <a href="https://pypi.org/project/aria-code/0.115.0/"><img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fpypi.org%2Fpypi%2Faria-code%2F0.115.0%2Fjson&amp;query=%24.info.version&amp;prefix=v&amp;style=flat-square&amp;logo=pypi&amp;label=PyPI&amp;color=blue&amp;cacheSeconds=300" alt="PyPI version"></a>
   <a href="https://github.com/artheras/aria-code/actions/workflows/ci.yml"><img src="https://img.shields.io/github/actions/workflow/status/artheras/aria-code/ci.yml?branch=main&style=flat-square&logo=githubactions&label=CI" alt="CI status"></a>
   <a href="LICENSE"><img src="https://img.shields.io/badge/license-Apache%202.0-64748b?style=flat-square" alt="Apache License 2.0"></a>
 </p>

{aria_code-0.113.0 → aria_code-0.115.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "aria-code"
-version = "0.113.0"
+version = "0.115.0"
 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" }
@@ -45,7 +45,7 @@ classifiers = [
 # parsers, the SQL engine and chart renderer are optional extras below — the
 # code already lazy-imports them and degrades gracefully when absent.
 dependencies = [
-    # The default model's client. DEFAULT_MODEL is google/gemini-2.5-pro, and
+    # The default model's client. DEFAULT_MODEL is google/gemini-3.5-flash, and
     # this used to be the optional `google` extra — so a plain
     # `pip install aria-code` failed on the very first prompt with "Gemini/Vertex
     # AI 需要 google-genai，当前未安装". The comment above that extra said as

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/_version.py

@@ -17,4 +17,4 @@ from __future__ import annotations
 
 __all__ = ["__version__"]
 
-__version__ = "0.113.0"
+__version__ = "0.115.0"

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/apps/cli/bootstrap.py

@@ -146,7 +146,10 @@ def initialize_cli_environment() -> None:
 # Provider configuration is the user's to make (that is what `/model` and
 # `.ariarc` are for); the code's job is to have a sane default and then get
 # out of the way.
-DEFAULT_MODEL = os.getenv("ARIA_DEFAULT_MODEL", "google/gemini-2.5-pro")
+DEFAULT_MODEL = os.getenv("ARIA_DEFAULT_MODEL", "google/gemini-3.5-flash")
+# A stable model (Google: available until at least May 2027), not a short-term
+# one: an install that is rarely updated should not lose its default within
+# months. gemini-2.5-pro, the previous default, retires in October 2026.
 
 
 def default_config() -> dict:
@@ -159,7 +162,7 @@ def default_config() -> dict:
         "ollama_url": os.getenv("OLLAMA_URL", "http://localhost:11434"),
         # Which backend serves a bare, unprefixed model name. Only consulted
         # for names that do not name their own provider ("qwen2.5:7b"); a
-        # prefixed id like "google/gemini-2.5-pro" always wins.
+        # prefixed id like "google/gemini-3.5-flash" always wins.
         "local_provider": os.getenv("ARIA_LLM_PROVIDER", "ollama"),
         "provider_fallback": "configured",
         "model": DEFAULT_MODEL,

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/apps/cli/commands/model_cmds.py

@@ -185,6 +185,17 @@ class ModelCommandsMixin:
     async def cmd_model(self, args: str):
         name = args.strip()
 
+        # A model Google has retired would fail on the next request; switch
+        # to its replacement and say so (model_retirement.py).
+        from aria_code.apps.cli.model_retirement import retired_replacement
+
+        _replacement = retired_replacement(name)
+        if _replacement:
+            msg = ui_text(self, f"ℹ {name} 已被 Google 停用，改用 {_replacement}。",
+                          f"ℹ {name} has been retired by Google; using {_replacement}.")
+            self.context.console.print(f"[yellow]{msg}[/yellow]") if self.context.has_rich else print(msg)
+            name = _replacement
+
         # ── "provider/model" format (Open Interpreter style) ─────────────────
         # Examples: /model deepseek/deepseek-chat  /model ollama/qwen2.5:7b
         #           /model openai/gpt-4.5          /model openai/o3  /model openai/o4-mini
@@ -1421,7 +1432,7 @@ class ModelCommandsMixin:
             ("deepseek",    "DeepSeek",      "deepseek/deepseek-chat"),
             ("anthropic",   "Anthropic",     "anthropic/claude-sonnet-4-6"),
             ("openai",      "OpenAI",        "openai/gpt-4.5"),
-            ("google",      "Google Gemini", "google/gemini-2.0-flash-exp"),
+            ("google",      "Google Gemini", "google/gemini-3.5-flash"),
             ("xai",         "xAI Grok",      "xai/grok-3-fast"),
             ("groq",        "Groq",          "groq/llama-3.3-70b-versatile"),
             ("mistral",     "Mistral",       "mistral/mistral-large-latest"),

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/apps/cli/commands/orchestrator_cmds.py

@@ -78,7 +78,7 @@ class OrchestratorCommandsMixin:
             return
             
         if self.context.has_rich:
-            provider = ConfiguredProvider(self.context.config, "gemini-2.5-flash")
+            provider = ConfiguredProvider(self.context.config, "gemini-3.5-flash")
         
         with self.context.console.status(f"[dim]{ui_text(self, 'Orchestrator 正在编排智能体网络', 'Orchestrator is planning the agent network')}...[/dim]"): 
             result = await dynamic_agent_orchestration(request, provider)

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/apps/cli/config_store.py

@@ -64,15 +64,57 @@ def load_cli_config(
     starting, because the user would then have no way to run the tool that
     would fix it.
     """
-    config = build_settings_service(paths, defaults, sync_policy=sync_policy).load()
+    service = build_settings_service(paths, defaults, sync_policy=sync_policy)
+    config = service.load()
+    # A model Google has retired fails on every request. The user's own saved
+    # choice is replaced once and written back; see model_retirement.
+    from aria_code.apps.cli.model_retirement import migrate_config
+
+    saved = migrate_config(config)
+    if saved:
+        try:
+            service.save(config)
+        except Exception:
+            pass
+        _announce(saved, config, pinned=False)
     try:
         from aria_code.ariarc import apply_to_config
 
         apply_to_config(config)
     except Exception:
         pass
+    # A project's .ariarc is the user's file: replace in memory, say so, never rewrite it.
+    pinned = migrate_config(config)
+    if pinned:
+        _announce(pinned, config, pinned=True)
     return config
 
 
+_ANNOUNCED: set[tuple[str, str]] = set()
+
+
+def _announce(changes, config: dict, *, pinned: bool) -> None:
+    """One line on stderr per retired model, once per process."""
+    import sys
+
+    zh = str(config.get("ui_lang", "")).lower().startswith("zh")
+    for _key, old, new in changes:
+        if (old, new) in _ANNOUNCED:
+            continue
+        _ANNOUNCED.add((old, new))
+        if zh:
+            where = "项目 .ariarc 里固定的" if pinned else "你设置的"
+            tail = "，请修改 .ariarc 里的 model" if pinned else "。可以用 /model 换成其他模型"
+            msg = f"ℹ {where}模型 {old} 已被 Google 停用，本次改用 {new}{tail}。"
+        else:
+            where = "The model pinned in this project's .ariarc" if pinned else "Your model"
+            tail = "; update the model in .ariarc" if pinned else "; use /model to pick another"
+            msg = f"ℹ {where}, {old}, has been retired by Google. Using {new}{tail}."
+        try:
+            print(msg, file=sys.stderr)
+        except Exception:
+            pass
+
+
 def save_cli_config(paths: AriaConfigPaths, cfg: dict) -> None:
     build_settings_service(paths, {}).save(cfg)

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/apps/cli/model_catalog.py

@@ -16,27 +16,43 @@ MODELS = {
     # ════════════════════════════════════════════════════════════════════
     # ── Google Gemini 平台核心默认模型（Google Cloud）───────────────────────
     # ════════════════════════════════════════════════════════════════════
-    "gemini-pro": {
-        "id": "google/gemini-2.5-pro",
-        "name": "Gemini Pro",
-        "version": "2.5",
+    # Gemini 2.5 retires in October 2026 (model_retirement.py). The default is
+    # a stable model; the newest Flash is offered beside it, but its line is
+    # replaced every few months.
+    "gemini-flash": {
+        "id": "google/gemini-3.5-flash",
+        "name": "Gemini Flash",
+        "version": "3.5",
         "tag": "Platform",
         "speed": "★★★★★",
         "intelligence": "★★★★★",
-        "description": "平台内置：Google Cloud 旗舰多模态、百万级上下文、量化核心",
+        "description": "平台默认：稳定版（至少到 2027 年 5 月）、百万级上下文、代码与金融分析",
         "capabilities": ["multimodal", "1M context", "tool calls", "quant", "code"],
         "thinking": True, "tools": True,
         "max_tokens": 8192, "num_ctx": 1048576, "temperature": 0.2,
         "badge": "Default",
     },
-    "gemini-flash": {
-        "id": "google/gemini-2.5-flash",
-        "name": "Gemini Flash",
-        "version": "2.5",
+    "gemini-pro": {
+        "id": "google/gemini-3.8-flash",
+        "name": "Gemini Flash (latest)",
+        "version": "3.8",
+        "tag": "Platform",
+        "speed": "★★★★★",
+        "intelligence": "★★★★★",
+        "description": "最新最强：推理与工具调用最好；迭代快，Google 几个月内会推出替代版",
+        "capabilities": ["multimodal", "1M context", "tool calls", "quant", "code"],
+        "thinking": True, "tools": True,
+        "max_tokens": 8192, "num_ctx": 1048576, "temperature": 0.2,
+        "badge": "Latest",
+    },
+    "gemini-lite": {
+        "id": "google/gemini-3.5-flash-lite",
+        "name": "Gemini Flash-Lite",
+        "version": "3.5",
         "tag": "Platform",
         "speed": "★★★★★",
         "intelligence": "★★★★",
-        "description": "极速默认：超低延迟、毫秒级响应、代码与实时行情",
+        "description": "极速低价：低延迟、适合批量与实时行情",
         "capabilities": ["fast inference", "1M context", "tool calls", "low latency"],
         "thinking": False, "tools": True,
         "max_tokens": 8192, "num_ctx": 1048576, "temperature": 0.2,
@@ -341,8 +357,10 @@ MODELS = {
 # Model aliases: short names / Ollama IDs → MODELS key
 MODEL_ALIASES = {
     # ── Google Gemini ─────────────────────────────────────────────────
-    "gemini": "gemini-pro", "gemini-pro": "gemini-pro", "gemini2.5": "gemini-pro", "gpro": "gemini-pro",
+    "gemini": "gemini-flash", "gemini-pro": "gemini-pro", "gemini2.5": "gemini-flash", "gpro": "gemini-pro",
+    "gemini-latest": "gemini-pro", "glatest": "gemini-pro",
     "gemini-flash": "gemini-flash", "gflash": "gemini-flash", "flash": "gemini-flash",
+    "gemini-lite": "gemini-lite", "glite": "gemini-lite",
     # ── Qwen 2.5 ──────────────────────────────────────────────────────
     "qwen7b": "qwen7b",   "q7": "qwen7b",   "sonata": "qwen7b",   "s": "qwen7b",
     "qwen14b": "qwen14b", "q14": "qwen14b",
@@ -464,7 +482,7 @@ _MODEL_ID_INDEX: dict[str, str] | None = None
 def _model_id_index() -> dict[str, str]:
     """Lazily build a lowercase {model id -> MODELS key} index.
 
-    The config stores the provider-qualified *id* ("google/gemini-2.5-pro")
+    The config stores the provider-qualified *id* ("google/gemini-3.5-flash")
     while MODELS is keyed by short name ("gemini-pro"), so without this lookup
     a registered model resolves to "_community_" and its own registry entry is
     never used.

aria_code-0.115.0/src/aria_code/apps/cli/model_retirement.py

@@ -0,0 +1,63 @@
+"""Models Google has retired, and what Aria moves their users to.
+
+Google's Vertex AI lifecycle page retires gemini-2.5-pro, -flash and
+-flash-lite on 20 October 2026 (the release notes give 16 October, so plan
+for the earlier date); the Gemini API already admits only users who have
+used 2.5 before. gemini-2.0-flash and -flash-lite shut down on 1 June 2026,
+and every 1.5 model in 2025. A saved config that still names one of them
+would fail on every request with an error that does not say why.
+
+So a retired id is replaced when the config is loaded, and the replacement
+is written back so it happens once: a stable model, not a short-term one,
+so an install that is rarely updated does not break again within months.
+The replacements follow Google's recommendations; 2.5 Flash moves to 3.5
+Flash rather than a Lite model so the user loses nothing in capability.
+"""
+
+from __future__ import annotations
+
+import re
+
+# bare id -> bare replacement
+RETIRED: dict[str, str] = {
+    "gemini-2.5-pro": "gemini-3.5-flash",
+    "gemini-2.5-flash": "gemini-3.5-flash",
+    "gemini-2.5-flash-lite": "gemini-3.1-flash-lite",
+    "gemini-2.0-flash": "gemini-3.5-flash",
+    "gemini-2.0-flash-lite": "gemini-3.1-flash-lite",
+    "gemini-1.5-pro": "gemini-3.5-flash",
+    "gemini-1.5-flash": "gemini-3.5-flash",
+    "gemini-1.0-pro": "gemini-3.5-flash",
+}
+
+# Dated, versioned and preview spellings of the same models:
+# gemini-2.5-pro-preview-06-05, gemini-2.0-flash-001, gemini-1.5-pro-latest.
+_SUFFIX = re.compile(r"(-preview(-[\w-]+)?|-exp(-[\w-]+)?|-latest|-\d{3}|-\d{2}-\d{2}|-\d{2}-\d{4})+$")
+
+
+def retired_replacement(model: str) -> str | None:
+    """The id to use instead of *model*, keeping its provider prefix; None if it is not retired."""
+    if not model:
+        return None
+    text = str(model).strip()
+    prefix, _, bare = text.rpartition("/")
+    key = _SUFFIX.sub("", bare.lower())
+    replacement = RETIRED.get(key)
+    if replacement is None:
+        return None
+    return f"{prefix}/{replacement}" if prefix else replacement
+
+
+def migrate_config(config: dict, keys: tuple[str, ...] = ("model",)) -> list[tuple[str, str, str]]:
+    """Replace retired model ids in *config* in place; return (key, old, new) for each."""
+    changes = []
+    for key in keys:
+        old = config.get(key)
+        new = retired_replacement(old) if isinstance(old, str) else None
+        if new:
+            config[key] = new
+            changes.append((key, old, new))
+    return changes
+
+
+__all__ = ["RETIRED", "migrate_config", "retired_replacement"]

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/aria_cli.py

@@ -508,7 +508,7 @@ _PROVIDER_DESC: Dict[str, str] = {
     "anthropic":   "Claude Sonnet 4 / Opus 4  长文档·代码·分析",
     "groq":        "Llama3/Qwen  超高速推理 (免费额度)",
     "together":    "开源模型聚合  Llama/Qwen/DeepSeek 等 100+ 模型",
-    "google":      "Gemini 2.5 Pro / 2.0 Flash  多模态·超长上下文",
+    "google":      "Gemini 3.8 Flash / 3.5 Flash / 3.5 Flash-Lite  多模态·超长上下文",
     "xai":         "Grok-3 / Grok-3-Fast  实时网络数据·Twitter整合",
     "mistral":     "Mistral Large / Codestral  欧洲顶级·代码生成强",
     "cohere":      "Command R+  RAG检索增强·企业文档处理",
@@ -660,17 +660,17 @@ _PROVIDER_GUIDE: Dict[str, str] = {
         "1. 打开 aistudio.google.com/app/apikey\n"
         "2. 用 Google 账号登录 → 「Create API key」\n"
         "3. 复制 AIzaSy... 格式的密钥\n\n"
-        "解锁: Gemini 2.5 Pro · Gemini 2.0 Flash · 多模态视觉\n"
-        "用法: /model google/gemini-2.0-flash-exp\n"
-        "免费额度: Flash 每分钟 15次，每天 1500次"
+        "解锁: Gemini 3.8 Flash · Gemini 3.5 Flash · 多模态视觉\n"
+        "用法: /model google/gemini-3.5-flash\n"
+        "免费额度以 Google AI Studio 当前公布为准"
     ),
     "gemini": (
         "同 google provider，填入 Google AI Studio 的 API Key\n\n"
         "推荐模型:\n"
-        "  gemini-2.5-pro        — 最强推理，128K 上下文\n"
-        "  gemini-2.0-flash-exp  — 超快，每分钟 15 次免费\n"
-        "  gemini-1.5-flash      — 稳定版，适合生产\n\n"
-        "用法: /model gemini/gemini-2.5-pro"
+        "  gemini-3.5-flash       — 稳定版，适合生产（默认）\n"
+        "  gemini-3.8-flash       — 最新最强，迭代较快\n"
+        "  gemini-3.5-flash-lite  — 更快更省\n\n"
+        "用法: /model gemini/gemini-3.5-flash"
     ),
     "xai": (
         "1. 打开 console.x.ai → 注册/登录\n"

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/ariarc.py

@@ -51,7 +51,7 @@ parent directories (walks up to filesystem root, stops at $HOME).
       // Which model this project runs on. Pins the backend for everyone who
       // opens the repo, so it does not depend on each developer's global
       // config. An explicit /model during a session still wins.
-      "model":      "google/gemini-2.5-pro",
+      "model":      "google/gemini-3.5-flash",
       "provider":   "google",
       "max_rounds": 40,
 

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/doctor.py

@@ -680,7 +680,7 @@ def _iter_required_modules() -> Iterable[tuple[str, str]]:
     yield "pandas", "dataframes"
     yield "numpy", "numeric processing"
     yield "yfinance", "US/HK/global market data"
-    # The default model is google/gemini-2.5-pro, so without this a fresh
+    # The default model is google/gemini-3.5-flash, so without this a fresh
     # install cannot answer a single prompt.
     yield "google.genai", "Gemini / Vertex AI (the default model)"
 

aria_code-0.115.0/src/aria_code/evals/gap.py

@@ -0,0 +1,93 @@
+"""In-sample minus out-of-sample: public tasks against their private twins.
+
+A public task can end up in a model's training data, and a tuned model can
+learn the public bank instead of the skill. Either way it scores well on the
+public task and worse on a twin it has never seen: the same skill, with
+different data, numbers and wording, kept in a private repository. SWE-bench
+Pro reads leakage the same way, from the gap between its public and held-out
+splits; a quant reads overfitting from the gap between in-sample and
+out-of-sample returns.
+
+A holdout suite names each twin's public task with ``twin_of``. This pairs
+them, takes each side's pass rate over its repeats, and reports the mean gap
+(public minus twin) with a bootstrap 95% interval. An interval above zero is
+a warning that the public score overstates the skill.
+
+    python -m aria_code.evals.gap --public evals-*.json --holdout holdout.json --suites holdout/suites/*.yaml
+
+``--markdown`` prints counts and rates only, never a private task id, so the
+section can go into a public job summary.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Iterable
+
+from aria_code.training.compare import bootstrap_ci, rates
+
+
+def twins(suite_files: Iterable[Path]) -> dict[str, str]:
+    """Holdout task id -> the public task it is a twin of."""
+    import yaml
+
+    pairs = {}
+    for suite_file in suite_files:
+        doc = yaml.safe_load(Path(suite_file).read_text(encoding="utf-8")) or {}
+        pairs.update({str(t["id"]): str(t["twin_of"]) for t in doc.get("tasks") or [] if t.get("twin_of")})
+    return pairs
+
+
+def gap(public_reports: Iterable[Path], holdout_reports: Iterable[Path], pairs: dict[str, str]) -> dict:
+    public, _, _ = rates(public_reports)
+    holdout, _, _ = rates(holdout_reports)
+    rows = {h: (public[p], holdout[h]) for h, p in pairs.items() if p in public and h in holdout}
+    diffs = [pub - twin for pub, twin in rows.values()]
+    lo, hi = bootstrap_ci(diffs)
+    mean = sum(diffs) / len(diffs) if diffs else 0.0
+    return {
+        "pairs": len(rows),
+        "public_pass_rate": round(sum(p for p, _ in rows.values()) / len(rows), 4) if rows else None,
+        "twin_pass_rate": round(sum(t for _, t in rows.values()) / len(rows), 4) if rows else None,
+        "mean_gap": round(mean, 4),
+        "ci95": [round(lo, 4), round(hi, 4)],
+        "public_overstates": lo > 0,
+        "unpaired": sorted(h for h in pairs if h not in rows),
+    }
+
+
+def markdown(result: dict) -> str:
+    """The gap for a public summary: rates and the interval, no task ids."""
+    if not result["pairs"]:
+        return "### Public vs holdout\nNo twin pairs were scored on both sides.\n"
+    lo, hi = result["ci95"]
+    verdict = ("The public score overstates the skill: the interval is above 0." if result["public_overstates"]
+               else "No significant gap: the interval includes 0 or is below it.")
+    return "\n".join([
+        "### Public vs holdout",
+        f"{result['pairs']} twin pairs. Public {result['public_pass_rate']:.0%}, "
+        f"holdout {result['twin_pass_rate']:.0%}; gap {result['mean_gap']:+.0%} "
+        f"(95% CI {lo:+.0%} to {hi:+.0%}).",
+        verdict,
+        *([f"{len(result['unpaired'])} twins had no scored pair."] if result["unpaired"] else []),
+        "",
+    ])
+
+
+def main(argv: list[str] | None = None) -> int:
+    ap = argparse.ArgumentParser(description="Gap between public tasks and their private twins.")
+    ap.add_argument("--public", nargs="+", type=Path, required=True)
+    ap.add_argument("--holdout", nargs="+", type=Path, required=True)
+    ap.add_argument("--suites", nargs="+", type=Path, required=True,
+                    help="the holdout suite YAML files (their tasks carry twin_of)")
+    ap.add_argument("--markdown", action="store_true", help="counts and rates only, safe for a public summary")
+    args = ap.parse_args(argv)
+    result = gap(args.public, args.holdout, twins(args.suites))
+    print(markdown(result) if args.markdown else json.dumps(result, ensure_ascii=False, indent=2))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/evals/summary.py

@@ -1,6 +1,11 @@
 """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
+    python3 -m aria_code.evals.summary evals-core.json evals-operations.json [--private evals-holdout.json]
+
+Each suite's score comes with the mean per-task pass rate and its standard
+error, so two runs can be told apart from noise. A report passed after
+``--private`` (the holdout suites, kept in a private repository) is shown as
+counts only: no task ids, details or logs, because the summary is public.
 
 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
@@ -15,22 +20,50 @@ A task the model gets wrong is a result, not a broken build: the exit code is
 from __future__ import annotations
 
 import json
+import math
 import sys
+from collections import defaultdict
 from pathlib import Path
 
 
-def summarise(reports: list[dict], model: str = "") -> tuple[str, int]:
+def mean_sem(report: dict) -> tuple[float, float, int] | None:
+    """Mean per-task pass rate over scored attempts, its standard error, and the task count."""
+    passed: dict[str, int] = defaultdict(int)
+    scored: dict[str, int] = defaultdict(int)
+    for r in report.get("results") or []:
+        if str(r.get("outcome", "")).lower() in ("pass", "fail"):
+            scored[str(r.get("task_id"))] += 1
+            passed[str(r.get("task_id"))] += str(r.get("outcome")).lower() == "pass"
+    rates = [passed[t] / scored[t] for t in scored]
+    if not rates:
+        return None
+    mean = sum(rates) / len(rates)
+    if len(rates) < 2:
+        return mean, 0.0, 1
+    var = sum((x - mean) ** 2 for x in rates) / (len(rates) - 1)
+    return mean, math.sqrt(var / len(rates)), len(rates)
+
+
+def summarise(reports: list[dict], model: str = "", private: list[dict] | None = None) -> 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:
+    for report in [*reports, *(private or [])]:
+        is_private = any(report is p for p in private or [])
         suite = report.get("suite", "?")
-        lines.append(f"### {suite}: {report.get('passed', 0)}/{report.get('scored', 0)} "
-                     f"({report.get('pass_rate', 0):.0%})")
+        lines.append(f"### {suite}{' (holdout)' if is_private else ''}: "
+                     f"{report.get('passed', 0)}/{report.get('scored', 0)} ({report.get('pass_rate', 0):.0%})")
+        stats = mean_sem(report)
+        if stats:
+            mean, sem, n = stats
+            lines.append(f"Mean task pass rate {mean:.0%} ± {sem:.0%} (SEM over {n} task{'s' if n != 1 else ''}).")
         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))
+        if is_private:
+            lines += ["Task ids, details and logs are not published for holdout suites.", ""]
+            continue
         lines += ["", "| Task | Outcome | Seconds | Detail |", "|---|---|---:|---|"]
         for result in report.get("results", []):
             detail = str(result.get("detail") or "").replace("|", "\\|")[:120]
@@ -56,17 +89,26 @@ def main(argv: list[str] | None = None) -> int:
     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:
+    public_names, private_names = args, []
+    if "--private" in args:
+        i = args.index("--private")
+        public_names, private_names = args[:i], args[i + 1:]
+
+    def load(names: list[str]) -> list[dict]:
+        out = []
+        for name in names:
+            path = Path(name)
+            if path.exists():
+                out.append(json.loads(path.read_text(encoding="utf-8")))
+            else:
+                print(f"(no report at {name})", file=sys.stderr)
+        return out
+
+    reports, private = load(public_names), load(private_names)
+    if not reports and not private:
         print("no eval reports to summarise", file=sys.stderr)
         return 1
-    text, code = summarise(reports, model)
+    text, code = summarise(reports, model, private)
     print(text)
     return code
 

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/model_capability.py

@@ -196,6 +196,11 @@ _CAPABILITY_TABLE: Dict[str, ModelCapability] = {
     # Registered here as well as in the MODELS catalogue: callers such as the
     # tool-routing check in aria_cli use get_model_capability() directly, and a
     # missing entry silently degrades the model to text_only / 4K context.
+    "gemini-3.8-flash":         ModelCapability(tool_calls=True, format="openai_native", context_window=1048576, temperature=0.2, size_class="large",  vision=True, thinking=True, notes="Gemini 3.8 Flash — 1M context, short-term line"),
+    "gemini-3.5-flash-lite":    ModelCapability(tool_calls=True, format="openai_native", context_window=1048576, temperature=0.2, size_class="medium", vision=True,                notes="Gemini 3.5 Flash-Lite — 1M context"),
+    "gemini-3.5-flash":         ModelCapability(tool_calls=True, format="openai_native", context_window=1048576, temperature=0.2, size_class="large",  vision=True, thinking=True, notes="Gemini 3.5 Flash — 1M context, stable (default)"),
+    "gemini-3.1-flash-lite":    ModelCapability(tool_calls=True, format="openai_native", context_window=1048576, temperature=0.2, size_class="medium", vision=True,                notes="Gemini 3.1 Flash-Lite — 1M context"),
+    # Retired by Google (model_retirement.py migrates configs that name them).
     "gemini-2.5-pro":           ModelCapability(tool_calls=True, format="openai_native", context_window=1048576, temperature=0.2, size_class="large",  vision=True, thinking=True, notes="Gemini 2.5 Pro — 1M context"),
     "gemini-2.5-flash":         ModelCapability(tool_calls=True, format="openai_native", context_window=1048576, temperature=0.2, size_class="medium", vision=True,                notes="Gemini 2.5 Flash — 1M context"),
     "gemini-2.0-flash":         ModelCapability(tool_calls=True, format="openai_native", context_window=1048576, temperature=0.2, size_class="medium", vision=True),

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/providers/llm/openai_compat.py

@@ -225,7 +225,7 @@ class GoogleProvider(OpenAICompatProvider):
 
     provider_name    = "google"
     DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta/openai"
-    DEFAULT_MODEL    = "gemini-2.5-flash"
+    DEFAULT_MODEL    = "gemini-3.5-flash"
 
     def __init__(self, config: ProviderConfig):
         if not config.api_key:

{aria_code-0.113.0 → aria_code-0.115.0}/src/aria_code/providers/llm/registry.py

@@ -73,7 +73,7 @@ _DEFAULT_FALLBACK_CHAIN = [
     ("dashscope",    "DASHSCOPE_API_KEY",     "qwen-plus"),
     ("moonshot",     "MOONSHOT_API_KEY",      "moonshot-v1-8k"),
     ("zhipu",        "ZHIPUAI_API_KEY",       "glm-4-flash"),
-    ("google",       "GEMINI_API_KEY",        "gemini-2.5-flash"),
+    ("google",       "GEMINI_API_KEY",        "gemini-3.5-flash"),
     ("openai",       "OPENAI_API_KEY",        "gpt-4o-mini"),
     ("anthropic",    "ANTHROPIC_API_KEY",     "claude-3-5-haiku-latest"),
     ("groq",         "GROQ_API_KEY",          "llama-3.3-70b-versatile"),

aria_code-0.115.0/src/aria_code/training/__init__.py

@@ -0,0 +1 @@
+"""Training data and evaluation for Aria's own model (docs/training-protocol.md)."""