@researai/deepscientist 1.5.9 → 1.5.11

package/docs/en/14_PROMPT_SKILLS_AND_MCP_GUIDE.md

@@ -0,0 +1,506 @@
+# 14 Prompt, Skills, and MCP Guide
+
+This guide explains how one DeepScientist turn is actually driven.
+
+Use it when you want to understand:
+
+- how the runtime builds the prompt for each turn
+- what each stage skill is for
+- how the built-in MCP tools are structured
+- which file or tool you should change when behavior feels wrong
+
+If you only want the user-facing product overview, read [13 Core Architecture Guide](./13_CORE_ARCHITECTURE_GUIDE.md) first.
+
+If you only want the built-in memory contract, read [07 Memory and MCP](./07_MEMORY_AND_MCP.md) after this page.
+
+## 1. One-sentence summary
+
+DeepScientist does not run from one static mega-prompt.
+
+For every turn, it rebuilds a prompt from:
+
+- the core system prompt
+- a shared interaction contract
+- runtime state
+- quest files
+- startup contract
+- selected memory
+- connector-specific rules when needed
+- the active skill structure
+
+Then the agent works through three built-in MCP namespaces only:
+
+- `memory`
+- `artifact`
+- `bash_exec`
+
+## 2. What files are the main prompt truth sources
+
+The most important files are:
+
+- `src/prompts/system.md`
+- `src/prompts/contracts/shared_interaction.md`
+- `src/prompts/connectors/qq.md`
+- `src/prompts/connectors/weixin.md`
+- `src/prompts/connectors/lingzhu.md`
+- `src/deepscientist/prompts/builder.py`
+- `src/skills/*/SKILL.md`
+- `src/deepscientist/mcp/server.py`
+
+In practice:
+
+- `system.md` defines the global operating stance
+- `shared_interaction.md` defines user-visible continuity rules
+- `connectors/*.md` inject connector-specific behavior only when that connector is active or bound
+- `builder.py` decides prompt assembly order and runtime context sections
+- `SKILL.md` files define stage-specific execution discipline
+- `mcp/server.py` defines the built-in tool surface
+
+## 3. How one turn prompt is assembled
+
+The current runtime assembles the turn prompt in roughly this order:
+
+1. `system.md`
+2. `contracts/shared_interaction.md`
+3. runtime context block
+4. active communication surface block
+5. optional connector contract block
+6. turn driver and continuation guard
+7. active user requirements
+8. quest context
+9. recent durable state
+10. research delivery policy
+11. paper and evidence snapshot
+12. retry recovery packet when this is a retry turn
+13. interaction style block
+14. priority memory for this turn
+15. recent conversation window
+16. current turn attachments
+17. current user message
+
+That order matters.
+
+The runtime is trying to answer three questions before the model acts:
+
+1. what is the quest trying to do now
+2. what durable state already exists
+3. what behavior rules apply on this surface and in this stage
+
+## 4. What the major prompt blocks actually do
+
+### 4.1 `system.md`
+
+This is the global DeepScientist operating contract.
+
+It defines things like:
+
+- long-horizon evidence-first behavior
+- use `bash_exec` for shell-like execution
+- use durable files, logs, and artifacts as truth
+- do not end a quest early
+- treat web, TUI, and connectors as one quest
+- user-facing reporting style
+
+If the agent starts sounding wrong everywhere, `system.md` is one of the first places to inspect.
+
+### 4.2 `shared_interaction.md`
+
+This file defines the common continuity spine around `artifact.interact(...)`.
+
+It tells the agent:
+
+- `artifact.interact(...)` is the main user-visible thread
+- queued inbound user messages must be acknowledged and handled first
+- blocking replies are for real decisions only
+- progress updates should be concise and human-readable
+
+If the model is bad at staying in the same long-running thread, this file matters a lot.
+
+### 4.3 Active communication surface
+
+The prompt builder adds a surface block each turn.
+
+This tells the model:
+
+- whether the current turn is local, QQ, Weixin, or another connector
+- how many external connectors are bound
+- which surface is active right now
+- how much detail is appropriate on that surface
+
+This is why connector behavior should not be hard-coded globally.
+
+The same quest may be viewed from the web UI, TUI, or a connector, but the reply shape should adapt.
+
+### 4.4 Connector contract
+
+Connector prompt fragments are loaded only when needed.
+
+Current connector prompt files are:
+
+- `src/prompts/connectors/qq.md`
+- `src/prompts/connectors/weixin.md`
+- `src/prompts/connectors/lingzhu.md`
+
+These files define surface-specific rules such as:
+
+- reply length
+- text-first versus media-enabled behavior
+- how attachments should be sent
+- what not to expose in chat
+
+For example:
+
+- QQ is treated as a milestone operator surface
+- Weixin is treated as a concise phone-side operator surface with `context_token` continuity
+- Lingzhu is treated as an even shorter, more constrained surface
+
+If one connector needs behavior changes, change its connector prompt first before bloating the global system prompt.
+
+### 4.5 Runtime context and durable quest state
+
+The builder injects runtime facts such as:
+
+- `quest_id`
+- `quest_root`
+- current workspace branch
+- active idea id
+- active analysis campaign id
+- bound conversations
+- startup contract
+- baseline gate
+- active interactions
+- recent artifacts
+- recent runs
+
+This is what makes the prompt quest-aware instead of generic.
+
+### 4.6 Quest context
+
+The builder reads these quest files directly into the prompt:
+
+- `brief.md`
+- `plan.md`
+- `status.md`
+- `SUMMARY.md`
+
+This is important:
+
+- the live prompt is not only based on chat history
+- durable quest docs are treated as first-class truth surfaces
+
+### 4.7 Research delivery policy
+
+This block converts startup choices into concrete execution rules.
+
+It includes logic around:
+
+- whether paper delivery is required
+- launch mode
+- custom profile
+- baseline routing
+- idea routing
+- paper branch behavior
+- review gate behavior
+
+If `Start Research` behavior feels wrong, you usually need to inspect:
+
+- the `startup_contract`
+- `src/deepscientist/prompts/builder.py`
+- the stage skill the quest is currently using
+
+### 4.8 Interaction style
+
+This block tells the model how to speak on this turn.
+
+It includes:
+
+- locale bias
+- blocking versus threaded behavior
+- long-run update cadence
+- how to acknowledge mailbox messages
+- how to compress progress into human-readable updates
+
+This is why DeepScientist can keep the same runtime but behave differently across:
+
+- long experiments
+- connector replies
+- writing stages
+- waiting-for-decision stages
+
+### 4.9 Priority memory
+
+DeepScientist does not inject memory randomly.
+
+`PromptBuilder` uses a stage-specific memory plan.
+
+Examples:
+
+- `scout` prefers `papers`, `knowledge`, `decisions`
+- `baseline` prefers `papers`, `decisions`, `episodes`, `knowledge`
+- `idea` prefers `papers`, `ideas`, `decisions`, `knowledge`
+- `experiment` prefers `ideas`, `decisions`, `episodes`, `knowledge`
+
+This means the prompt is stage-biased on purpose.
+
+The agent should not see the same memory bundle on every turn.
+
+## 5. What can be overridden locally
+
+Prompt fragments can be overridden per quest under:
+
+- `.codex/prompts/system.md`
+- `.codex/prompts/contracts/shared_interaction.md`
+- `.codex/prompts/connectors/<connector>.md`
+
+This means:
+
+- repo defaults live under `src/prompts/`
+- quest-local prompt overrides live under `.codex/prompts/`
+
+Use quest-local prompt overrides only when the quest truly needs a different local contract.
+
+Do not fork the repo-level prompt unless the change should affect the product globally.
+
+## 6. How skills are structured
+
+DeepScientist currently has two skill layers:
+
+1. standard stage skills
+2. companion skills
+
+### 6.1 Standard stage skills
+
+These are the main research anchors:
+
+| Skill | Use when | Main job | Usually hands off to |
+|---|---|---|---|
+| `scout` | the task frame is still unclear | task framing, baseline discovery, metric and dataset clarification | `baseline` or `idea` |
+| `baseline` | a trustworthy baseline does not yet exist | attach, import, reproduce, repair, and verify the baseline | `idea` |
+| `idea` | the baseline is clear but the next direction is not | generate, compare, and select durable research directions | `experiment` |
+| `experiment` | one selected idea is ready to run | implement and evaluate the main run on one durable line | `analysis-campaign`, `write`, or `decision` |
+| `analysis-campaign` | follow-up experiments are needed | run slices, ablations, robustness checks, or reviewer-facing supplements | `write`, `decision`, or `finalize` |
+| `write` | there is enough evidence to draft | turn accepted evidence into outline, draft, and paper bundle work | `review` or `finalize` |
+| `finalize` | the quest is near closure | consolidate claims, summaries, final state, and closure checks | quest completion approval |
+| `decision` | a durable route choice is needed | make a clear go/stop/branch/reuse decision from evidence | another anchor |
+
+### 6.2 Companion skills
+
+These are auxiliary entry or quality-control skills:
+
+| Skill | Use when | Main job |
+|---|---|---|
+| `figure-polish` | a figure is important beyond debug use | render-inspect-revise a milestone or paper figure |
+| `intake-audit` | the quest already has meaningful prior state | trust-rank old assets and choose the correct next anchor |
+| `review` | a substantial draft already exists | run a skeptical paper-like audit before claiming done |
+| `rebuttal` | reviewer comments or revision requests exist | map reviewer pressure into experiments, text deltas, and response artifacts |
+
+### 6.3 The important design point
+
+The daemon is not supposed to contain a giant hard-coded research scheduler.
+
+Instead:
+
+- the prompt defines the operating contract
+- the skill defines stage-specific discipline
+- the runtime persists state and routes turns
+
+That is the core DeepScientist design choice.
+
+## 7. What each skill usually leaves behind
+
+These are the durable outputs you should expect:
+
+| Skill | Typical durable outputs |
+|---|---|
+| `scout` | updated `brief.md`, updated `plan.md`, literature notes, framing memory |
+| `baseline` | `PLAN.md`, `CHECKLIST.md`, baseline verification notes, confirmed or waived baseline state |
+| `idea` | durable idea draft, selected idea package, rationale for why this route won |
+| `experiment` | implementation changes, run logs, `record_main_experiment(...)`, result evidence |
+| `analysis-campaign` | campaign manifest, slice records, synthesis notes |
+| `write` | selected outline, writing plan, draft, references, claim-evidence map, paper bundle |
+| `finalize` | final summary, closure state, final quest health check |
+| `decision` | durable route decision, next-anchor recommendation |
+| `intake-audit` | trusted-versus-untrusted asset map, next anchor recommendation |
+| `review` | review report, revision log, experiment TODO list |
+| `rebuttal` | review matrix, response letter, text deltas, evidence-update plan |
+| `figure-polish` | final polished figure assets and render-checked outputs |
+
+## 8. Built-in MCP structure
+
+DeepScientist keeps the built-in MCP surface intentionally small.
+
+Only these namespaces are built in:
+
+- `memory`
+- `artifact`
+- `bash_exec`
+
+There is no separate public built-in `git` namespace.
+
+Git-aware behavior is exposed through `artifact`.
+
+### 8.1 `memory`
+
+Purpose:
+
+- reusable knowledge
+- lessons that should survive beyond one turn
+- quest-local or global memory cards
+
+Current built-in tools:
+
+- `memory.write(...)`
+- `memory.read(...)`
+- `memory.search(...)`
+- `memory.list_recent(...)`
+- `memory.promote_to_global(...)`
+
+Use `memory` when the output should be remembered and reused later.
+
+Do not use it for transient progress chatter.
+
+### 8.2 `artifact`
+
+Purpose:
+
+- quest control plane
+- durable research state
+- Git-aware branch and worktree routing
+- experiment and paper records
+- user-visible interaction continuity
+
+The artifact namespace is large, but it is still one family.
+
+#### A. Generic durable records
+
+- `artifact.record(...)`
+- `artifact.refresh_summary(...)`
+- `artifact.render_git_graph(...)`
+
+#### B. Branch and route control
+
+- `artifact.checkpoint(...)`
+- `artifact.prepare_branch(...)`
+- `artifact.activate_branch(...)`
+- `artifact.submit_idea(...)`
+- `artifact.list_research_branches(...)`
+- `artifact.resolve_runtime_refs(...)`
+
+#### C. Baseline lifecycle
+
+- `artifact.publish_baseline(...)`
+- `artifact.attach_baseline(...)`
+- `artifact.confirm_baseline(...)`
+- `artifact.waive_baseline(...)`
+
+#### D. Experiment and analysis lifecycle
+
+- `artifact.record_main_experiment(...)`
+- `artifact.create_analysis_campaign(...)`
+- `artifact.get_analysis_campaign(...)`
+- `artifact.record_analysis_slice(...)`
+
+#### E. Paper lifecycle
+
+- `artifact.submit_paper_outline(...)`
+- `artifact.list_paper_outlines(...)`
+- `artifact.submit_paper_bundle(...)`
+
+#### F. Reading and interaction continuity
+
+- `artifact.arxiv(...)`
+- `artifact.interact(...)`
+- `artifact.complete_quest(...)`
+
+The most important artifact tool for long-running collaboration is:
+
+- `artifact.interact(...)`
+
+Because it keeps together:
+
+- user-visible updates
+- mailbox polling
+- connector delivery
+- threaded continuity
+- attachment delivery
+
+### 8.3 `bash_exec`
+
+Purpose:
+
+- durable shell execution
+- monitored long runs
+- durable logs
+- stoppable and readable sessions
+
+Current built-in tool:
+
+- `bash_exec.bash_exec(...)`
+
+This one tool supports multiple modes:
+
+- `detach`
+- `await`
+- `read`
+- `kill`
+- `list`
+- `history`
+
+The design rule is simple:
+
+- anything shell-like should go through `bash_exec`
+- do not hide important execution in transient shell snippets
+
+## 9. How the three MCP namespaces divide responsibility
+
+Use this mental model:
+
+- `memory`: remember
+- `artifact`: decide and record
+- `bash_exec`: run and monitor
+
+Examples:
+
+- a reusable lesson from a failed run -> `memory.write(...)`
+- confirming a baseline -> `artifact.confirm_baseline(...)`
+- launching training -> `bash_exec.bash_exec(mode='detach', ...)`
+- notifying the user about the next checkpoint -> `artifact.interact(...)`
+
+If you mix these roles badly, the quest becomes harder to resume and audit.
+
+## 10. How one real turn usually works
+
+A typical turn looks like this:
+
+1. a user or connector message arrives
+2. the daemon restores quest snapshot and history
+3. `PromptBuilder` assembles the current turn prompt
+4. the active skill defines the stage discipline
+5. priority memory is injected
+6. the agent uses `memory`, `artifact`, and `bash_exec`
+7. outputs are persisted into files, artifacts, memory cards, logs, and Git state
+8. `artifact.interact(...)` keeps the user-facing thread continuous
+
+That is why DeepScientist feels more like a persistent workshop than a stateless chat.
+
+## 11. When to change prompt, skill, or MCP code
+
+Use this quick rule:
+
+- change `src/prompts/system.md` when the global operating stance is wrong
+- change `src/prompts/contracts/shared_interaction.md` when continuity behavior is wrong
+- change `src/prompts/connectors/*.md` when one connector behaves wrong
+- change `src/skills/<skill>/SKILL.md` when one stage behaves wrong
+- change `src/deepscientist/prompts/builder.py` when prompt assembly or runtime context selection is wrong
+- change `src/deepscientist/mcp/server.py` when the built-in tool surface itself is wrong
+
+Do not use a giant prompt patch to fix a real MCP contract bug.
+
+Do not use a new MCP tool to fix a stage-discipline problem that belongs in a skill.
+
+## 12. What to read next
+
+- [07 Memory and MCP](./07_MEMORY_AND_MCP.md)
+- [13 Core Architecture Guide](./13_CORE_ARCHITECTURE_GUIDE.md)
+- [06 Runtime and Canvas](./06_RUNTIME_AND_CANVAS.md)
+- [01 Settings Reference](./01_SETTINGS_REFERENCE.md)

package/docs/en/99_ACKNOWLEDGEMENTS.md

@@ -13,8 +13,11 @@ We especially thank the following projects for their ideas, research direction,
 - EvoScientist
 - Orchestra-Research
 - Orchestra-Research/AI-Research-SKILLs
+- Overleaf / overleaf/overleaf
+- Monaco Editor / microsoft/monaco-editor
+- Novel / steven-tey/novel
 
-These projects have been important references for thinking about automated research, open-ended exploration, evolutionary search, experiment organization, and long-horizon research agents.
+These projects have been important references for thinking about automated research, open-ended exploration, evolutionary search, experiment organization, long-horizon research agents, collaborative LaTeX editing workflows, browser-based code editing experiences, and rich notebook-style writing interactions.
 
 ## Personal acknowledgements
 

package/docs/en/README.md

@@ -0,0 +1,79 @@
+# DeepScientist Docs
+
+DeepScientist is not just a long-running autonomous scientific discovery system. It is also a persistent research map that lives on your own machine.
+
+2 minutes to install. 2 minutes to bind Weixin. 2 minutes to launch. Extremely fast and easy to use.
+
+It is also a workshop-style collaboration environment: let it keep moving autonomously, or step in anytime to collaborate, edit code, run the terminal yourself, or keep notes and plans in a Notion-style workspace.
+
+Use DeepScientist anywhere: on the server through TUI, in the browser through Web, on the phone through Weixin or QQ, and even on glasses through Rokid Glasses.
+
+## News
+
+- `2026/03/24`: DeepScientist officially releases `v1.5`.
+- `2026/02/01`: the DeepScientist paper is available on [OpenReview](https://openreview.net/forum?id=cZFgsLq8Gs) for `ICLR 2026`.
+
+## More From ResearAI
+
+AI Scientist and AutoFigure projects worth exploring alongside DeepScientist:
+
+| Project | What it does |
+|---|---|
+| [AutoFigure](https://github.com/ResearAI/AutoFigure) | generate paper-ready figures |
+| [AutoFigure-Edit](https://github.com/ResearAI/AutoFigure-Edit) | editable vector paper figures |
+| [DeepReviewer-v2](https://github.com/ResearAI/DeepReviewer-v2) | review papers and drafts |
+| [Awesome-AI-Scientist](https://github.com/ResearAI/Awesome-AI-Scientist) | curated AI scientist landscape |
+
+This page is the shortest path to the right document.
+
+## If you are new
+
+- [00 Quick Start](./00_QUICK_START.md)
+  Start here if you want to install DeepScientist, launch it locally, and create your first project.
+- [12 Guided Workflow Tour](./12_GUIDED_WORKFLOW_TOUR.md)
+  Follow the real product flow from landing page to workspace, step by step.
+- [02 Start Research Guide](./02_START_RESEARCH_GUIDE.md)
+  Read this next if you want to understand each field in the `Start Research` dialog and what it actually submits.
+
+## If you want to launch a project well
+
+- [02 Start Research Guide](./02_START_RESEARCH_GUIDE.md)
+  Explains the current frontend fields, derived contract fields, and practical examples.
+- [01 Settings Reference](./01_SETTINGS_REFERENCE.md)
+  Use this when you need to configure runners, connectors, runtime defaults, or home paths.
+- [11 License And Risk Notice](./11_LICENSE_AND_RISK.md)
+  Read this first if you care about the license boundary, server safety, fabricated outputs, connector leakage, and public exposure risk.
+
+## If you want to collaborate through external surfaces
+
+- [10 Weixin Connector Guide](./10_WEIXIN_CONNECTOR_GUIDE.md)
+  Bind personal WeChat through DeepScientist's built-in QR login and iLink runtime.
+- [03 QQ Connector Guide](./03_QQ_CONNECTOR_GUIDE.md)
+  Use QQ as a practical collaboration surface for progress, commands, and milestone delivery.
+- [04 Lingzhu Connector Guide](./04_LINGZHU_CONNECTOR_GUIDE.md)
+  Bind Lingzhu / Rokid Glasses to DeepScientist.
+
+## If you want to understand how the system works
+
+- [13 Core Architecture Guide](./13_CORE_ARCHITECTURE_GUIDE.md)
+  Read this first if you want a user-facing overview of launcher, daemon, quests, Canvas, memory, and connectors.
+- [14 Prompt, Skills, and MCP Guide](./14_PROMPT_SKILLS_AND_MCP_GUIDE.md)
+  Read this when you want the real turn-time structure: prompt assembly order, stage skills, and built-in MCP tool families.
+- [06 Runtime and Canvas](./06_RUNTIME_AND_CANVAS.md)
+  Explains how the daemon, workspace, canvas, and connector views fit together.
+- [07 Memory and MCP](./07_MEMORY_AND_MCP.md)
+  Explains memory, artifacts, and built-in MCP behavior.
+
+## If something is broken
+
+- [09 Doctor](./09_DOCTOR.md)
+  Start here for diagnostics and common runtime problems.
+- [01 Settings Reference](./01_SETTINGS_REFERENCE.md)
+  Check this if the problem is likely caused by config, credentials, or connector setup.
+
+## If you are developing DeepScientist
+
+- [90 Architecture](./90_ARCHITECTURE.md)
+  High-level system contracts and repository structure.
+- [91 Development](./91_DEVELOPMENT.md)
+  Maintainer-facing workflow and implementation notes.

package/docs/images/weixin/weixin-plugin-entry.svg

@@ -0,0 +1,33 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="1280" height="720" viewBox="0 0 1280 720" fill="none">
+  <rect width="1280" height="720" rx="32" fill="#FAFBFF"/>
+  <rect x="40" y="40" width="1200" height="640" rx="28" fill="#FFFFFF" stroke="#DAE2F3" stroke-width="2"/>
+  <circle cx="108" cy="108" r="34" fill="#2563EB"/>
+  <path d="M92 108C92 99.16 99.16 92 108 92C116.84 92 124 99.16 124 108C124 116.84 116.84 124 108 124C99.16 124 92 116.84 92 108Z" fill="#fff" fill-opacity=".2"/>
+  <path d="M101 116V100H115" stroke="#fff" stroke-width="8" stroke-linecap="round" stroke-linejoin="round"/>
+  <text x="164" y="98" fill="#0F172A" font-family="Arial, sans-serif" font-size="34" font-weight="700">WeChat Version And Plugin Entry Check</text>
+  <text x="164" y="138" fill="#475569" font-family="Arial, sans-serif" font-size="20">Reference precheck from the personal WeChat guide before you bind in DeepScientist</text>
+
+  <rect x="88" y="196" width="492" height="430" rx="28" fill="#F8FAFF" stroke="#D5DDF0" stroke-width="2"/>
+  <text x="120" y="252" fill="#0F172A" font-family="Arial, sans-serif" font-size="28" font-weight="700">Compatibility checklist</text>
+  <rect x="120" y="292" width="428" height="88" rx="20" fill="#EEF2FF"/>
+  <text x="150" y="326" fill="#3730A3" font-family="Arial, sans-serif" font-size="20" font-weight="700">WeChat for iOS</text>
+  <text x="150" y="356" fill="#334155" font-family="Arial, sans-serif" font-size="18">Use 8.0.70 or newer if you need the plugin entry.</text>
+  <rect x="120" y="398" width="428" height="180" rx="20" fill="#fff" stroke="#E1E7F5"/>
+  <text x="150" y="434" fill="#0F172A" font-family="Arial, sans-serif" font-size="20" font-weight="700">Expected path in the reference flow</text>
+  <text x="150" y="474" fill="#334155" font-family="Arial, sans-serif" font-size="18">1. Me</text>
+  <text x="150" y="508" fill="#334155" font-family="Arial, sans-serif" font-size="18">2. Settings</text>
+  <text x="150" y="542" fill="#334155" font-family="Arial, sans-serif" font-size="18">3. Plugins</text>
+  <text x="150" y="576" fill="#475569" font-family="Arial, sans-serif" font-size="16">If this entry is missing, update WeChat first.</text>
+
+  <rect x="640" y="168" width="520" height="486" rx="34" fill="#ECFDF5" stroke="#C3EED3" stroke-width="2"/>
+  <rect x="760" y="204" width="280" height="420" rx="42" fill="#FFFFFF" stroke="#D7EAD9" stroke-width="2"/>
+  <rect x="786" y="246" width="228" height="42" rx="16" fill="#DCFCE7"/>
+  <text x="847" y="273" fill="#166534" font-family="Arial, sans-serif" font-size="20" font-weight="700">WeChat</text>
+  <rect x="786" y="318" width="228" height="54" rx="18" fill="#F8FAFC"/>
+  <text x="818" y="352" fill="#0F172A" font-family="Arial, sans-serif" font-size="18">Me</text>
+  <rect x="786" y="388" width="228" height="54" rx="18" fill="#F8FAFC"/>
+  <text x="818" y="422" fill="#0F172A" font-family="Arial, sans-serif" font-size="18">Settings</text>
+  <rect x="786" y="458" width="228" height="54" rx="18" fill="#DCFCE7" stroke="#22C55E"/>
+  <text x="818" y="492" fill="#166534" font-family="Arial, sans-serif" font-size="18" font-weight="700">Plugins</text>
+  <text x="702" y="616" fill="#475569" font-family="Arial, sans-serif" font-size="17">DeepScientist still binds from its own QR modal.</text>
+</svg>

package/docs/images/weixin/weixin-qr-confirm.svg

@@ -0,0 +1,30 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="1280" height="720" viewBox="0 0 1280 720" fill="none">
+  <rect width="1280" height="720" rx="32" fill="#F7FBF8"/>
+  <rect x="40" y="40" width="1200" height="640" rx="28" fill="#FFFFFF" stroke="#D7E9DC" stroke-width="2"/>
+  <circle cx="108" cy="108" r="34" fill="#16A34A"/>
+  <path d="M97 108L105 116L119 100" stroke="#fff" stroke-width="8" stroke-linecap="round" stroke-linejoin="round"/>
+  <text x="164" y="98" fill="#0F172A" font-family="Arial, sans-serif" font-size="34" font-weight="700">Weixin QR Login Confirmation</text>
+  <text x="164" y="138" fill="#475569" font-family="Arial, sans-serif" font-size="20">One QR image in the modal, phone-side confirmation, automatic persistence</text>
+
+  <rect x="86" y="190" width="484" height="456" rx="28" fill="#FBFCFF" stroke="#DCE4F4" stroke-width="2"/>
+  <text x="118" y="244" fill="#0F172A" font-family="Arial, sans-serif" font-size="28" font-weight="700">DeepScientist modal</text>
+  <rect x="158" y="288" width="340" height="340" rx="26" fill="#fff" stroke="#CBD5E1" stroke-width="2"/>
+  <rect x="198" y="328" width="260" height="260" fill="#111827"/>
+  <rect x="220" y="350" width="60" height="60" fill="#fff"/>
+  <rect x="378" y="350" width="60" height="60" fill="#fff"/>
+  <rect x="220" y="508" width="60" height="60" fill="#fff"/>
+  <rect x="300" y="430" width="40" height="40" fill="#fff"/>
+  <rect x="364" y="494" width="28" height="28" fill="#fff"/>
+  <text x="136" y="660" fill="#475569" font-family="Arial, sans-serif" font-size="17">The popup only needs the QR image and a short hint.</text>
+
+  <rect x="708" y="164" width="472" height="500" rx="36" fill="#ECFDF5" stroke="#C6EDD2" stroke-width="2"/>
+  <rect x="846" y="202" width="196" height="414" rx="34" fill="#FFFFFF" stroke="#D3E7D6" stroke-width="2"/>
+  <text x="889" y="242" fill="#166534" font-family="Arial, sans-serif" font-size="22" font-weight="700">Phone</text>
+  <rect x="874" y="284" width="140" height="82" rx="18" fill="#DCFCE7"/>
+  <text x="908" y="332" fill="#166534" font-family="Arial, sans-serif" font-size="18" font-weight="700">Scan</text>
+  <rect x="874" y="392" width="140" height="82" rx="18" fill="#BBF7D0"/>
+  <text x="896" y="440" fill="#166534" font-family="Arial, sans-serif" font-size="18" font-weight="700">Confirm login</text>
+  <rect x="874" y="500" width="140" height="82" rx="18" fill="#86EFAC"/>
+  <text x="895" y="548" fill="#14532D" font-family="Arial, sans-serif" font-size="18" font-weight="700">Saved automatically</text>
+  <text x="742" y="638" fill="#475569" font-family="Arial, sans-serif" font-size="17">DeepScientist persists the returned token and account ids.</text>
+</svg>