Add channel prefixes, pin edit targets in Cursor prompts, and force fresh sessions for page refine URLs so DeepSeek/Goose chat history does not pollute page tasks.
Co-authored-by: Cursor <cursoragent@cursor.com>
Image turns already get a vision-model description injected into the prompt,
but the agent kept calling read_image to "confirm" the pictures. Those tool
results carry base64 image parts that Goose persists, so every later turn
against the text-only chat provider failed with `unknown variant image_url`
before the agent could write the page. WeChat page requests therefore fell
through to the fail-closed delivery message.
Drop read_image for the turn whenever a vision model handles the images, say
so explicitly in the injected prompt, and teach the poison scan to recognise
tool image parts so already-polluted sessions rotate instead of failing again.
Co-authored-by: Cursor <cursoragent@cursor.com>
Wait for HTML materialization before releasing delivery contracts, pass
verified MindSpace URLs through proactive WeChat sends, resend on reconcile
when pages land late, and report deferred customer-service delivery as unsent.
Co-authored-by: Cursor <cursoragent@cursor.com>
After a successful image report flow, Goose may keep image_url parts that
DeepSeek rejects on later turns while PUT scrub returns 405. Detect polluted
sessions up front, rotate to a fresh agent route with carried assistant
context, and recover image URLs from recent media on historical retries.
Co-authored-by: Cursor <cursoragent@cursor.com>
Phrases like "解读详细报告，做成页面" now bind to the in-session report
context instead of behaving like a brand-new page topic.
Co-authored-by: Cursor <cursoragent@cursor.com>
Goose cannot persist historical image scrub (PUT 405), so DeepSeek rejects
image_url on the next turn. Reattach the recent report image and rotate to a
fresh session instead of treating the follow-up as a missing page source.
Co-authored-by: Cursor <cursoragent@cursor.com>
Reconcile now re-adds missing sandbox-fs before reply; update the dedicated
session test to assert add_extension and restart are invoked.
Co-authored-by: Cursor <cursoragent@cursor.com>
When passive Recognition is empty, convert AMR to mp3 and call WeChat
addvoicetorecofortext before the existing asr.tkmind.cn fallback.
Co-authored-by: Cursor <cursoragent@cursor.com>
Prevent chat-to-page flows like writing a poem then saying "做成页面吧" from opening
a fresh session without carried content, failing with a vague error, or letting Memory
page-template preferences replace the immediate topic.
Co-authored-by: Cursor <cursoragent@cursor.com>
Intent routing and reply content now both trigger HTML delivery: public links,
public/*.html paths, or html writes in the agent reply enable artifact recovery,
link attachment, and fail-closed when the claimed page is missing.
Co-authored-by: Cursor <cursoragent@cursor.com>
Persist formal replies when WeChat customer-service quota is exhausted,
claim them once on the next passive response, skip progress notices for
page generation, and suppress failure notices while a formal reply is pending.
Co-authored-by: Cursor <cursoragent@cursor.com>
Skip hot-adding stdio extensions goosed omits from /extensions, stabilize MCP
scope during session start, and wire chat intent routing into the agent-run worker.
Co-authored-by: Cursor <cursoragent@cursor.com>
Prevent text-only planning replies from marking WeChat delivery done when
static-page-publish never confirmed a public HTML file.
Co-authored-by: Cursor <cursoragent@cursor.com>
Fall back to workspace share thumbnails so already-published HTML can be
sent instead of blocking on REQUIRED_FRESH. Also allow an explicit
ALLOW_DIRECT_STABLE_RELEASE path matching the pre-2026-07-23 direct
stable packaging flow.
Co-authored-by: Cursor <cursoragent@cursor.com>
Recognize retry, HTML page, page edits, and poem content edits as session continuations so delivery skips fresh-thumbnail forcing and retains agent context for multi-turn page work.
Co-authored-by: Cursor <cursoragent@cursor.com>
Match phrases like「生成一个精美的 HTML 页面」so poem follow-ups skip fresh-thumbnail forcing and reuse the same session poem context.
Co-authored-by: Cursor <cursoragent@cursor.com>
Skip fresh thumbnail forcing and retry in-session when load_skill completes without HTML落盘, so「把刚才的诗做成页面」can deliver the prior poem as a static page.
Co-authored-by: Cursor <cursoragent@cursor.com>
Short follow-ups like "把刚才的诗做成页面" must keep the existing agent
session so Goose still sees the adjacent poem in snapshot history. Rotating
orphan routes before these requests dropped the poem and revived stale memory.
Co-authored-by: Cursor <cursoragent@cursor.com>
Backfill missing platform-brand footer at serve time and block service-
account links when the HTML file omits data-mindspace-page-tag.
Co-authored-by: Cursor <cursoragent@cursor.com>
Mock HTML in wechat-mp tests now includes mindspace-cover metadata so
page link delivery assertions match production guard behavior.
Co-authored-by: Cursor <cursoragent@cursor.com>
Notify users when HTML publish guards reject stub pages instead of silently
failing, send real page links when non-stub files exist, and split wechat-mp
into wechat-mp.bundle.mjs for fast production updates without full portal rebuilds.
Co-authored-by: Cursor <cursoragent@cursor.com>
When Goose surfaces session poison errors as assistant text instead of Error events,
assertWechatAgentReplyIsSendable now throws to trigger session recreation. Outbound
sanitization prevents raw API errors from reaching WeChat users.
Co-authored-by: Cursor <cursoragent@cursor.com>
Use session snapshot message counts alongside WeChat route counts so long-lived dedicated sessions rotate before Goose history bloats, and refresh snapshots asynchronously so customer replies are not blocked.
Co-authored-by: Cursor <cursoragent@cursor.com>
When a dedicated WeChat session history leaves unmatched tool_calls, the next reply fails with a 400 LLM validation error; detect that pattern and recreate the session before retrying.
Co-authored-by: Cursor <cursoragent@cursor.com>
john
tkmind