- Merge the separate custom-providers section into the unified provider
grid; "Custom" in the add-provider picker now acts as an add-new action
(trailing + mark) and opens the dedicated modal, supporting multiple
OpenAI-compatible endpoints
- Simplify the custom provider modal: drop the default-model field, add
an inline delete button, align colors with the theme
- Keep the legacy single "custom" card visible (models page, chat
dropdown and legacy config page) while custom_api_key/custom_api_base
is still in use, so existing single-provider setups don't disappear
- Unify user-facing wording from "vendor" to "provider" in UI and docs
- Restructure custom provider docs (zh/en/ja) around Web console and
config file usage
_set_chat rejected the expanded "custom:<id>" ids with "unknown
provider", so switching to a custom provider was only possible from
the custom providers section. Now the chat card and the custom section
behave consistently: _set_chat validates the id against
custom_providers, falls back to the provider's default model when none
is picked, and _chat_capability expands the dropdown with the
"custom:<id>" entries (legacy single-custom mode unchanged).
Co-authored-by: Cursor <cursoragent@cursor.com>
Two independent robustness fixes:
1. ChatChannel.cancel_session / cancel_all_session raised KeyError when a
session existed in self.sessions but no future had been dispatched yet.
self.sessions[sid] is created in produce(), but self.futures[sid] is only
created later in consume() on first dispatch. Cancelling in that window
(e.g. user sends a message then immediately cancels) crashed the cancel
path. Use self.futures.get(sid, []) so an absent entry is a no-op.
2. compress_imgfile decremented JPEG quality by 5 with no lower bound. For an
image that cannot be compressed below max_size, quality went 0, negative,
... — the loop never terminated and passed invalid quality values to PIL.
Add a min_quality floor (10) and return the best effort once reached.
Adds tests/test_robustness_fixes.py covering both paths (5 tests).
1. Creating a provider no longer auto-switches bot_type. Only an
explicit make_active=true changes the active model — prevents
silently hijacking users on Claude/OpenAI/etc.
2. When a provider IS activated, its 'model' is now written into the
global 'model' field. This ensures all three paths (regular chat,
agent_bridge, vision) use the correct model without per-path patches.
3. Removed unused i18n key 'models_custom_name_exists' (no longer
referenced after the id-based rework removed name-collision checks).
4. Updated tests: 39 passing (added model-sync tests, fixed tests that
relied on the removed auto-activation behavior).
Rework the multi custom-provider design per maintainer review:
1. Data model: use server-generated uuid4 short id as primary key;
'name' is now a pure display label that can be freely renamed.
2. Routing: drop 'custom_active_provider'; activate a provider by
setting bot_type to 'custom:<id>'. Single source of truth — no
pointer drift between bot_type and a separate active selector.
3. Security: drag_sensitive() now recursively masks api_key/secret in
nested structures (custom_providers list); previously only top-level
string fields were masked.
4. Per-provider model: the provider's 'model' field now takes effect on
the main chat path and agent path (was silently ignored before).
5. XSS fix: replace all inline onclick handlers in custom-provider UI
with data-* attributes + event delegation. Provider names never
appear in executable HTML contexts.
Legacy compatibility: bot_type='custom' (no colon) still reads the flat
custom_api_key/custom_api_base fields byte-for-byte identically.
Closes: consolidates #2876 into this PR as requested.
Ref: #2838
Adds first-class support for configuring more than one custom
OpenAI-compatible provider (e.g. SiliconFlow, DeepSeek, local vLLM)
and switching the active one from the web console, addressing #2838.
Backend:
- config: new `custom_providers` (list) and `custom_active_provider`
fields, fully backward compatible with the legacy single
`open_ai_api_base`/`model` fields (used as fallback).
- models/custom_provider.py: centralized resolver
`resolve_custom_credentials()` returning (api_key, api_base, model),
with active-provider selection and graceful fallback.
- chat_gpt_bot.py wired to use the resolver.
- web_channel.py: `_provider_overview` expands `custom_providers` into
one card per provider (id `custom:<name>`, active flag, masked key);
new POST actions `set_custom_provider`, `delete_custom_provider`,
`set_active_custom_provider` with hermetic persistence + bridge reset.
Frontend:
- console.js: dedicated "Custom providers" section with add / edit /
delete / set-active actions, masked-key keep-existing handling, and
~20 new zh/en i18n strings.
- chat.html: custom provider modal.
Tests:
- tests/test_custom_provider.py (11) - resolver/config behavior.
- tests/test_custom_provider_handlers.py (18) - write-side handlers and
overview expansion, including duplicate-name rejection.
All 29 unit tests pass.
Add a self-evolution subsystem that reviews idle conversations in an
isolated agent and durably learns from them — patching/creating skills,
finishing unfinished tasks, and backfilling missed memory.
- Trigger: background idle scan, fires when a session is idle >= N min AND
(>= N turns OR context usage > 80%). In-memory cursor reviews only new
messages so a session never re-learns old content.
- Isolated review agent: same model, restricted toolset, hard write-guard
confining edits to the workspace (built-in skills are protected).
- Safety: file-level backup before edits + evolution_undo tool; notify the
user ONLY when a workspace file actually changed (no-nag rule); capped
concurrency.
- Records to memory/evolution/<date>.md, surfaced in the memory UI's
renamed "Self-Evolution" tab (merged with dream diaries).
- Hide internal [SCHEDULED]/[EVOLUTION]/backup_id markers from chat history
display (also fixes scheduler marker leakage) while keeping them in stored
content for undo.
- Flat config: self_evolution_enabled (default off until release),
self_evolution_idle_minutes (15), self_evolution_min_turns (6).
- Tests: tests/test_evolution.py (stub + real model modes, 7 scenarios).
- Add language labels and copy buttons to code blocks
- Enhance drag-and-drop to full chat view
- Fix data consistency bugs in message operations
- Use RLock to prevent deadlock in conversation store"
- Add MINIMAX_M3 = "MiniMax-M3" constant and put it first in MODEL_LIST
- Default MinimaxBot model: MiniMax-M2.7 -> MiniMax-M3
- Keep MiniMax-M2.7 and MiniMax-M2.7-highspeed as legacy options
- Drop MINIMAX_M2_5 / MINIMAX_M2_1 / MINIMAX_M2_1_LIGHTNING / MINIMAX_M2
- Update web console recommended/provider model lists
- Update README capability table and docs/models index (en/zh/ja)
- Update docs/models/minimax.mdx and coding-plan.mdx MiniMax section
- Update run.sh / run.ps1 installer default and menu hint
- Update zh CLI status sample output
- Update unit tests to assert new M3 default and constant
TTS (speech-2.*) and API base URL remain unchanged.
Agent-generated images are sent as IMAGE_URL with a file:// path, but the wechatmp channel always used requests.get, which fails on file:// with InvalidSchema. Now read local files directly (file:// or local path) and fall back to HTTP download for remote URLs, in both passive and active reply modes.
Co-authored-by: Cursor <cursoragent@cursor.com>
Previously the passive reply only drained the cache after the agent task fully finished, so for long multi-turn tasks the user could not retrieve already-cached intermediate segments. Now return cached segments as soon as they are available, even while the task is still running; the next user message fetches the rest.
Co-authored-by: Cursor <cursoragent@cursor.com>
In subscription account passive reply mode, WeChat allows only one reply per request. Multi-turn agent output was cached as separate entries, forcing the user to send an extra message to fetch each one. Now drain and merge all consecutive cached text segments into a single reply; media still returns one at a time.
Co-authored-by: Cursor <cursoragent@cursor.com>
Adopt the same channel-level pattern as weixin/wecom_bot/feishu so
the agent actually sees attachments the user sent:
- IMAGE: agent mode never reads memory.USER_IMAGE_CACHE, so a photo
sent before a question (e.g. "image" then 30s later "what's this?")
used to be lost. Now lone images go into channel.file_cache and
the next TEXT turn appends "[图片: <path>]" to the query before
producing the context. Cross-batch image+text combinations now
work as users expect.
- FILE: previously dropped at the sync_msg filter and unsupported
by WechatKfMessage. Add msgtype="file" parsing, download via the
WeCom media API, preserve the original filename from
Content-Disposition (RFC 5987 + plain forms), and route through
the same file_cache pipeline as images, surfacing as
"[文件: <path>]" in the next text turn.
Move the sync_msg cursor file from the project-local tmp/ dir to ~/.wechat_kf_cursors.json so it survives tmp/ cleanups and cwd changes across restarts. Aligns with the weixin channel's credentials file convention.
- add wechat_kf_cursor_path config (default ~/.wechat_kf_cursors.json)
- expand ~ via os.path.expanduser in the channel init
- chmod the cursor file to 0o600 after each flush (no-op on Windows)
_dedup_image_text_pair previously fell back to returning only the last message whenever the batch was not exactly an image+text pair, which silently dropped multiple texts/images sent in quick succession.
Cursor freshness is already guaranteed by sync_msg, so no extra stale-history protection is needed. Now we return all messages by default and only collapse a batch when it is exactly a 2-message image+text pair within a 5s window (order-insensitive, normalized to [image, text]).
WeCom requires the callback HTTP response within ~5s, otherwise it retries the same notification. The previous code ran sync_msg pulling synchronously inside Query.POST, so a backlog could exceed the deadline and trigger retries that race on the same cursor and end up replying to the same user multiple times.
- Dispatch consume_callback to a background ThreadPoolExecutor and return 'success' immediately from the HTTP handler.
- Serialize work per open_kfid with a lock so retried/concurrent callbacks queue up instead of racing the cursor window.
- Shutdown the executor on channel stop().
