Add CONTEXT_RUNTIME_ACTIVE_ENV bundle and zero-LLM active probe covering
MCP compact, zvec, and headroom proxy readiness while keeping memory
injection on shadow.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add active-mode probe script and document mixed shadow/active promotion;
first active switch is MEMIND_MCP_COMPACT_MODE only.
Co-authored-by: Cursor <cursoragent@cursor.com>
Allow custom shadow portal probe messages and document admin-db steps to
enable recall_fusion_resolved without turning on active injection.
Co-authored-by: Cursor <cursoragent@cursor.com>
Pin @deepseek-ai/dsh@0.1.1-rc.2 probing in check-dsh-executor-local.mjs,
keep live headless smoke opt-in only, and record the 10k-token approval gate.
Co-authored-by: Cursor <cursoragent@cursor.com>
Document the ~/.local/bin vs /opt/homebrew/bin mismatch that breaks MCP
connection when install targets a path the global npm binary does not use.
Co-authored-by: Cursor <cursoragent@cursor.com>
Document MIT dev-tool setup for Cursor code intelligence and add an
offline probe for CLI, Cursor MCP config, and Memind index readiness.
Co-authored-by: Cursor <cursoragent@cursor.com>
Provide context-runtime-profile simulation for gateway shadow events,
run-context-runtime-offline.mjs for zero-LLM verification, and a local
dev guide for user-led shadow self-tests later.
Co-authored-by: Cursor <cursoragent@cursor.com>
Block unattended DashScope smokes by default and provide a zero-token
merge verification entry via run-goosed-v149-phase3-offline.mjs.
Co-authored-by: Cursor <cursoragent@cursor.com>
Prevent phase2/phase3/all reruns from burning DashScope tokens unless
GOOSE_V149_ALLOW_REAL_LLM=1 is explicitly set with human approval.
Co-authored-by: Cursor <cursoragent@cursor.com>
Introduce off/shadow/active compaction for large tkmind read/research and
excel report payloads with optional Redis-backed ctx_fetch handles, keeping
fail-open behavior when storage is unavailable.
Co-authored-by: Cursor <cursoragent@cursor.com>
Record the merged Context Runtime + Goose v1.49 canary work as closed
so the branch cannot be reused as a development baseline.
Co-authored-by: Cursor <cursoragent@cursor.com>
Use the default headroom health probe instead of the goosed fetch adapter
so active mode routes deepseek-no-think through the proxy, and add an E2E
script that verifies goosed pong via headroom with fail-open portal checks.
Co-authored-by: Cursor <cursoragent@cursor.com>
Wire DeepSeek Harness as an optional fourth code executor (default off)
and add a loopback observation script that compares prompt tokens through
headroom proxy with OPENAI_TARGET_API_URL pointed at the compat proxy.
Co-authored-by: Cursor <cursoragent@cursor.com>
Introduce RRF-based recall fusion for personal/episodic/temporal paths,
zvec-grep workspace shadow probing for code executors, and a local
headroom proxy starter aligned with the fusion plan defaults.
Co-authored-by: Cursor <cursoragent@cursor.com>
Portal resume smoke can exceed 300s when agent-run worker is busy; raise
phase3 default to 600s and refresh memory/sanitize baselines from smoke.
Co-authored-by: Cursor <cursoragent@cursor.com>
Lock health recording behind confirm-only state machines, reject public publishes for the health category, and ship the 14-day experiment Page Data templates with tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
Wire Portal and TKMind proxy to loopback Goose v1.49 via canary env blocks,
with verification scripts, Phase 2/3 evidence baselines, and rollback runbooks
so local upgrade stays isolated from stable 1.41 and production.
Co-authored-by: Cursor <cursoragent@cursor.com>
Auto-enable device/location prefixes for cross_device scenarios, thread question_device into keyword and embedding paths, and boost lexical scores from source-tag tokens.
Co-authored-by: Cursor <cursoragent@cursor.com>
Map Mi-Memory Structure/Expansion into h5_experience V1 fields and record JKRiver Sleep design lessons under AGPL constraints, plus post-fix MemFuseBench numbers.
Co-authored-by: Cursor <cursoragent@cursor.com>
Wire an offline benchmark over the production pgvector ranking path so Memory V2 recall can be measured with candidate/ranking loss split before lifecycle work lands.
Co-authored-by: Cursor <cursoragent@cursor.com>
Record feature/wechat-intent-channel-separation as merged and provide a reproducible admin-db toggle for Tang canary LLM router shadow on production.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep Context Stability First as the active plan, document reuse vs conflict, and freeze the 15-question answers so the next session does not redesign from scratch.
Co-authored-by: Cursor <cursoragent@cursor.com>
Document R1–R3 root causes in the H5 direct/agent routing path, reject the Task Runtime approach, and define a phased rollout with Memory V2 compatibility constraints.
Co-authored-by: Cursor <cursoragent@cursor.com>
Introduce Draft → Confirm → Commit flow for WeChat schedule intents behind
feature flags, plus h5_tasks dual-write/read aggregation and rollout scripts
so reminders and automations get explicit user confirmation before persisting.
Co-authored-by: Cursor <cursoragent@cursor.com>
The read_image poisoning fix had no end-to-end guard because scenarios could
only send text. Add an upload_images step that renders and uploads visually
distinct images, let chat steps attach them the same way the WeChat channel
does, and assert the generated page actually embeds every upload.
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>
Register branch disposition after 103 deploy/repair, and add a lightweight
103 repair script that avoids heavy runtime module imports.
Co-authored-by: Cursor <cursoragent@cursor.com>
Formalize fast release as --mode fast so verified changes can skip canary promotion and full Gate reruns while keeping backups and health checks.
Co-authored-by: Cursor <cursoragent@cursor.com>
Shared paths like db.mjs were pulling PAGE/DATA live agent suites into every
hotfix. Keep those cases for actual page-data changes, resume passed suites on
the same artifact, and fail fast on Docker/port issues instead of rerunning 100+
scenarios.
Co-authored-by: Cursor <cursoragent@cursor.com>
DATA-01 told the agent not to stop at 方案确认, then failed because the first
reply no longer contains those words. Assert 问卷/后台 delivery instead.
Co-authored-by: Cursor <cursoragent@cursor.com>
Expired paid users were rebuilt as free with period_images_limit=0, which the quota system treats as unlimited. Write the catalog quota on rebuild and repair existing finite plans on schema ensure.
Co-authored-by: Cursor <cursoragent@cursor.com>
Drop the user_confirmed_at gate so public, online, unexpired pages can enter sitemap/llms and receive SEO/GEO tags. Defaults now enable all discovery switches; stored all-off config is still preserved until admin saves.
Co-authored-by: Cursor <cursoragent@cursor.com>
john