fix(topic): 深话题 analyzer 失败退避 + 富化失败原因可见化（不泄漏对话原文） #11822

Workflow file for this run

.github/workflows/analyze.yml at eb6b883

	name: Analyze

	on:
	push:
	branches: [main]
	pull_request:
	branches: [main]
	workflow_dispatch:

	# Static-check jobs only need to read the checked-out source.
	permissions:
	contents: read

	# ─────────────────────────────────────────────────────────────────────────────
	# Five parallel jobs, split by concern so a failing check name points at a
	# category (expand it to see which step is red) instead of one giant title
	# enumerating every sub-check:
	#
	# Ruff → the ruff linter (its own deps: uv + ruff)
	# Code safety → runtime hazards in the backend (blocking-in-async,
	# banned imports, startup import cost)
	# Prompts & i18n → LLM-call discipline + localization/docstring rules
	# API & layering → HTTP/route + module-layer architecture rules
	# Core package contracts → the main_logic/core mixin/facade structural gate
	#
	# All run in parallel, all block merge, each is independently re-runnable. Only
	# the diff-based steps (i18n-sync, docstring-cjk) need full history; the rest
	# run on a shallow checkout.
	# ─────────────────────────────────────────────────────────────────────────────

	jobs:
	ruff:
	name: Ruff
	runs-on: ubuntu-latest
	timeout-minutes: 10
	steps:
	- uses: actions/checkout@v5

	- name: Set up Python 3.11
	uses: actions/setup-python@v5
	with:
	python-version: '3.11'

	- name: Install uv
	uses: astral-sh/setup-uv@v4

	- name: Install ruff
	# ruff is the only dev dep we need; a full uv sync pulls heavy native
	# extensions (playwright, numpy, etc.) unnecessary for static checks.
	run: uv tool install ruff==0.15.4

	- name: ruff check (incl. ASYNC210/220/221/222/251)
	run: ruff check .

	code-safety:
	name: Code safety
	runs-on: ubuntu-latest
	timeout-minutes: 10
	steps:
	- uses: actions/checkout@v5

	- name: Set up Python 3.11
	uses: actions/setup-python@v5
	with:
	python-version: '3.11'

	- name: Forbid blocking calls in async def bodies
	# Custom AST checker, broader than ruff's ASYNC* rules:
	# * gaps flake8-async doesn't catch: Thread/Process.join,
	# queue.Queue.get, raw socket recv/accept/connect
	# * extra blocking stdlib/3p calls: PIL.Image.open,
	# pyautogui.screenshot, Fernet encrypt/decrypt, shutil.*,
	# json.load, plus bare-import forms (sleep, rmtree, urlopen)
	# * depth-1 transitive: a sync helper whose body hits any of the
	# above, called directly from async, is also flagged.
	run: python scripts/check_async_blocking.py

	- name: Forbid loguru / structlog / logbook imports
	# Logging is unified through utils.logger_config (RobustLoggerConfig).
	# Re-introducing a third-party logging frontend fragments the surface
	# (formatter, sinks, file naming, multi-process behaviour) and breaks
	# plugin/main parity. This check fails the build on any such import.
	run: python scripts/check_no_loguru.py

	- name: Forbid `tkinter` imports
	# Architecture rule: frontend/backend are split, ALL GUI operations
	# belong on the Electron frontend (Node.js renderer/main), not the
	# Python backend. The backend is a headless HTTP/async service —
	# it must stay relocatable (remote deployment, headless CI) and
	# not own a Tcl/Tk event loop. Past incident: PR #1014's Windows
	# tk screenshot overlay crashed the whole app under Nuitka builds
	# without `--enable-plugin=tk-inter` (SystemExit from tk.Tk()
	# escaped the asyncio worker, killed uvicorn). For dialogs use
	# Electron's `dialog` (or platform-native shell bridges in
	# storage_location_router.py); for screenshot framing use
	# Electron's desktopCapturer region path.
	run: python scripts/check_no_tkinter.py

	- name: Forbid `temperature=` kwargs on LLM client calls (memory + utils)
	# Project policy: do NOT pass `temperature=` to create_chat_llm /
	# ChatOpenAI / wrappers in memory_server.py + memory/ + utils/. The
	# default (None) omits the field — required for o1/o3/gpt-5-thinking/
	# Claude extended-thinking, and avoids per-call-site temperature drift
	# across memory tasks. See memory/__init__.py and .agent/rules/
	# neko-guide.md for the rationale.
	run: python scripts/check_no_temperature.py

	- name: Keep heavy SDKs off the startup import chain (lazy-import contract)
	# Merged production mode imports the app tree serially before any port
	# binds, so a module-scope `import openai` (etc.) slows EVERY launch,
	# silently — reference incident: #1496 cut main_server import to ~0.6s,
	# openai 2.x types growth quietly ate it back to ~2.1s within 6 weeks.
	# Banned-at-module-scope list (openai/anthropic/bs4/bilibili_api/
	# google.genai/translatepy/dashscope/...) lives in the script; the
	# sanctioned pattern is in-function import + utils/module_warmup.py
	# background pre-import after ready. `# noqa: STARTUP_LAZY_IMPORT`
	# opts out with justification.
	run: python scripts/check_startup_import_lazy.py

	prompts-i18n:
	name: Prompts & i18n
	runs-on: ubuntu-latest
	timeout-minutes: 10
	steps:
	- uses: actions/checkout@v5
	with:
	# i18n-sync / docstring-cjk are diff-based and need full history to
	# diff against the merge-base of origin/main. The other steps don't
	# care, but fetch-depth: 0 has negligible cost on this repo.
	#
	# The checkout@v5 bump also fixes a credential-handling regression the
	# old @v4 tripped against the runner's git 2.54 on the full-history
	# fetch ("could not read Username", git exit 128).
	fetch-depth: 0

	- name: Set up Python 3.11
	uses: actions/setup-python@v5
	with:
	python-version: '3.11'

	- name: Enforce prompt-i18n conventions (inline-EN-only + multilang-in-config)
	# Two rules in one walker:
	# - INLINE_PROMPT_NON_EN: any string at an LLM call site or in a
	# module-level PROMPT/INSTRUCTION/*SYSTEM constant must be
	# English-bodied (CJK ratio <30%). Embedded short examples in
	# CJK are allowed under the threshold.
	# - I18N_NOT_IN_CONFIG: multi-language dicts (≥2 lang keys, must
	# include 'en') belong in config/prompts_*.py, not regular code.
	# Per-line `# noqa: <CODE>` suppression supported. See PR #974 for
	# the reference incident that motivated this lint.
	run: python scripts/check_prompt_hygiene.py

	- name: Enforce LLM budget/timeout discipline (output budget + timeout + input budget)
	# Project policy: every LLM call must be bounded.
	# - LLM_OUTPUT_BUDGET: every create_chat_llm() / ChatOpenAI()
	# construction must set a token budget (max_completion_tokens= /
	# max_tokens=) AND a timeout=. Without a budget the reply can run
	# away (cost + latency + context blow-up); without a timeout a hung
	# upstream wedges the async pipeline — neither has a safe default in
	# utils/llm_client.py. Sites that set budget/timeout per-call (via
	# invoke/ainvoke/..(**overrides)) opt out with a justified
	# `# noqa: LLM_OUTPUT_BUDGET`.
	# - LLM_INPUT_BUDGET (heuristic): every dynamic LLM call must be
	# input-budget-aware (truncate_to_tokens / *_MAX_TOKENS in the
	# enclosing function). Intentionally-uncapped sites (user-config /
	# OS window title etc., cf. docs/design/llm-prompt-budget.md §6)
	# opt out with `# noqa: LLM_INPUT_BUDGET`.
	# See docs/design/llm-prompt-budget.md for the full contract.
	run: python scripts/check_llm_budget.py

	- name: Verify i18n locale files move in lockstep (PR-only)
	# Diff-based: when ANY static/locales/*.json or
	# frontend/plugin-manager/src/i18n/locales/*.ts changes, ALL files
	# in the group must change AND every hunk must occupy the same line
	# ranges across all languages. Skipped on direct push to main —
	# there's nothing to diff against.
	if: github.event_name == 'pull_request'
	env:
	# Indirection instead of inlining ${{ github.base_ref }} into the
	# shell line: with the `pull_request: branches: [main]` trigger the
	# ref is always `main`, but zizmor flags the inline form as
	# template-injection, and env-var expansion is inert either way.
	BASE_REF: ${{ github.base_ref }}
	run: python scripts/check_i18n_sync.py --base "origin/${BASE_REF}"

	- name: Forbid CJK in new/modified Python docstrings (PR-only)
	# Convention: docstrings in the main program are written in English.
	# The repo carries a large legacy stock of CJK docstrings, so this is
	# a diff-ratchet: only docstrings whose line span overlaps lines
	# added/modified in this PR (vs the merge-base) are checked — touch
	# it, translate it. plugin/ and local_server/ are policy-exempt
	# (plugin internals follow their own conventions; local_server is a
	# semi-independent unit). `--full` exists for local migration sweeps.
	# Suppress a single docstring with `# noqa: DOCSTRING_CJK` (e.g.
	# fixtures whose CJK content is itself under test). Skipped on direct
	# push to main — nothing to diff against.
	if: github.event_name == 'pull_request'
	env:
	# Same zizmor template-injection indirection as the i18n-sync step.
	BASE_REF: ${{ github.base_ref }}
	run: python scripts/check_docstring_no_cjk.py --base "origin/${BASE_REF}"

	- name: Require zh-TW on new localized prompt dicts (PR-only)
	# A dict under config/prompts/ with an 'en' key plus 'zh'/'zh-CN' is a
	# localized prompt table, and `_loc` falls back to 'en' rather than
	# 'zh' on a missing key — so a table without 'zh-TW' serves Traditional
	# Chinese users an English prompt. 339 existing tables are still short
	# one (the batched backfill in issue #2500), so this is a ratchet: it
	# compares how many such tables exist at the merge-base vs at HEAD and
	# fails only when the total grew. A plain total is what keeps renames,
	# copy edits, added locales, and the eventual 'zh'->'zh-CN' migration
	# from tripping it — see the script docstring for the three narrower
	# keys that were tried and why each broke. `--full` lists the whole
	# backlog, `--count` just sizes it. Suppress a single dict with
	# a `noqa` comment naming PROMPT_ZH_TW on its opening or closing line
	# (comma-separated lists and a bare `noqa` work too, as in the sibling
	# gates). Skipped on direct push
	# to main — nothing to diff against.
	if: github.event_name == 'pull_request'
	env:
	# Same zizmor template-injection indirection as the i18n-sync step.
	BASE_REF: ${{ github.base_ref }}
	run: python scripts/check_prompt_zh_tw.py --base "origin/${BASE_REF}"

	api-layering:
	name: API & layering
	runs-on: ubuntu-latest
	timeout-minutes: 10
	steps:
	- uses: actions/checkout@v5

	- name: Set up Python 3.11
	uses: actions/setup-python@v5
	with:
	python-version: '3.11'

	- name: Forbid trailing-slash route paths on FastAPI decorators (backend)
	# Project convention: every backend HTTP/WebSocket endpoint is
	# declared WITHOUT a trailing slash. Avoids Starlette's absolute-URL
	# 307 redirect under reverse proxies — root cause of the PR #938
	# chara_manager regression: nginx/etc that don't transparently
	# forward Host send the redirect Location to 127.0.0.1:<internal>,
	# which the LAN browser can't reach (ERR_CONNECTION_REFUSED).
	# The lint exempts the literal '/' root page and explicit alias pairs
	# (same function carrying both '/foo' and '/foo/'). See
	# .agent/rules/neko-guide.md (§"API URL 末尾不带斜杠") and
	# main_routers/characters_router.py docstring.
	run: python scripts/check_api_trailing_slash.py

	- name: Forbid trailing-slash /api/... URL literals (frontend)
	# Counterpart to check_api_trailing_slash.py: the backend may declare
	# /api/foo without trailing slash, but if frontend code calls
	# fetch('/api/foo/') the same 307 → ERR_CONNECTION_REFUSED happens.
	# Regex sniffer over static/, frontend/, templates/. Recognises
	# prefix builders ('/api/foo/' + id, `/api/foo/${id}`) and exempts
	# them; flags only standalone literals ending in '/'. Suppress with
	# // noqa: API_TRAILING_SLASH if calling a third-party API that
	# genuinely requires the slash.
	run: python scripts/check_frontend_api_trailing_slash.py

	- name: Enforce top-level module layering (no inversions, no cycles)
	# Top-level packages have a strict ordering — only higher layers may
	# depend on lower ones (config/steamworks → utils → memory/main_logic
	# → main_routers → plugin → brain → app). The walker descends into
	# function bodies and string-form dynamic imports
	# (importlib.import_module / __import__), so deferred / conditional
	# imports cannot smuggle in a layer inversion or cycle. Run
	# `python scripts/check_module_layering.py --show-layers` to print
	# the hierarchy. Companion unit test: tests/unit/test_module_layering.py.
	run: python scripts/check_module_layering.py

	- name: Forbid relative-up markdown links inside docs/
	# docs/ ships through VitePress with itself as the deploy root;
	# any markdown link target starting with '..' resolves outside the
	# site and breaks deploy. We've shipped this regression more than
	# once — this check fails the build before the next attempt lands.
	# Fix is to inline the path as code (`foo/bar.js`) instead of a
	# link, or move the referenced content into docs/.
	run: python scripts/check_docs_no_relative_paths.py

	core-contracts:
	name: Core package contracts
	runs-on: ubuntu-latest
	timeout-minutes: 10
	steps:
	- uses: actions/checkout@v5

	- name: Set up Python 3.11
	uses: actions/setup-python@v5
	with:
	python-version: '3.11'

	- name: Enforce main_logic/core structural contracts
	# The LLMSessionManager mixin split (#2272) rests on contracts a
	# comment cannot enforce: symbols that tests rebind on the
	# main_logic.core facade must be read by manager/mixin code through
	# the _core_facade late-binding object (a from-import snapshot lets
	# isolation-style stubs go silently green while the real function
	# runs), mixins hold methods only, method sets stay disjoint, and
	# the facade keeps its layout. The patched-symbol set is harvested
	# from tests/ by AST on every run, so new patch targets tighten the
	# gate automatically. Not diff-based — cheap full check, runs on
	# push and PR alike.
	run: python scripts/check_core_contracts.py

Navigation Menu

Uh oh!

fix(topic): 深话题 analyzer 失败退避 + 富化失败原因可见化（不泄漏对话原文） #11822

Workflow file

fix(topic): 深话题 analyzer 失败退避 + 富化失败原因可见化（不泄漏对话原文） #11822

Uh oh!

Workflow file for this run