verify-deepseekbot
Prove DeepSeekBot behavior in a real isolated DSH Host the way a user does, and keep the evidence. Use before claiming a feature or fix works, to reproduce a reported bug, or to capture PR screenshots; its feature map says how to reach and drive each mapped feature and what observable result proves it.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 852
- Forks
- 63
01 前言 Front Matter
- name
- verify-deepseekbot
- description
- Prove DeepSeekBot behavior in a real isolated DSH Host the way a user does, and keep the evidence. Use before claiming a feature or fix works, to reproduce a reported bug, or to capture PR screenshots; its feature map says how to reach and drive each mapped feature and what observable result proves it.
02 资源文件 78
- 参考文档features/activity-center-idle.md7.1 KB
- 参考文档features/activity-inline-approval.md15.0 KB
- 参考文档features/appearance-recovery.md4.0 KB
- 参考文档features/approval-templates.md2.7 KB
- 参考文档features/assignment-report-continuation.md3.7 KB
- 参考文档features/bot-call.md21.4 KB
- 参考文档features/bot-dm-reply.md16.6 KB
- 参考文档features/bot-group-mention.md2.0 KB
- 参考文档features/bot-mode-dock.md31.4 KB
- 参考文档features/bot-mode.md6.6 KB
- 参考文档features/bot-session-folders.md4.1 KB
- 参考文档features/bot-stop.md2.4 KB
- 参考文档features/browser-lifecycle.md6.7 KB
- 参考文档features/channel-sections.md2.6 KB
- 参考文档features/cloud-host-connections.md3.1 KB
- 参考文档features/composer-footer.md4.8 KB
- 参考文档features/connect-channel.md15.9 KB
- 参考文档features/conversation-timeline.md44.1 KB
- 参考文档features/data-reset.md2.4 KB
- 参考文档features/desktop-window-zoom.md14.5 KB
- 参考文档features/global-materials.md2.3 KB
- 参考文档features/global-search.md2.1 KB
- 参考文档features/group-invitations.md3.1 KB
- 参考文档features/group-send-review.md17.0 KB
- 参考文档features/im-bot-pickers.md2.2 KB
- 参考文档features/im-conversation-mirroring.md3.6 KB
- 参考文档features/im-dm-commands.md3.0 KB
- 参考文档features/image-wallpaper.md4.7 KB
- 参考文档features/local-appearance.md3.7 KB
- 参考文档features/local-computer-video.md24.5 KB
- 参考文档features/marketplace.md5.3 KB
- 参考文档features/memory-diffs.md2.0 KB
- 参考文档features/memory-images.md10.1 KB
- 参考文档features/memory-restore.md5.2 KB
- 参考文档features/native-settings-handoff.md2.1 KB
- 参考文档features/orchestrator-model-switch.md2.6 KB
- 参考文档features/panel-appearance.md4.0 KB
- 参考文档features/personabot-create.md7.7 KB
- 参考文档features/phone-activity.md10.0 KB
- 参考文档features/phone-android-interactions.md40.0 KB
- 参考文档features/phone-bot-create.md17.3 KB
- 参考文档features/phone-browser.md2.0 KB
- 参考文档features/phone-channel-avatars.md3.3 KB
- 参考文档features/phone-channel-files.md6.9 KB
- 参考文档features/phone-chat-backgrounds.md4.4 KB
- 参考文档features/phone-chat-row-actions.md7.6 KB
- 参考文档features/phone-composer.md7.3 KB
- 参考文档features/phone-deletion.md4.7 KB
- 参考文档features/phone-drawer-views.md6.2 KB
- 参考文档features/phone-haptic-feedback.md4.7 KB
- 参考文档features/phone-interface.md4.7 KB
- 参考文档features/phone-multiple-hosts.md5.5 KB
- 参考文档features/phone-notifications.md30.8 KB
- 参考文档features/phone-pairing.md37.4 KB
- 参考文档features/phone-pinned-list.md7.1 KB
- 参考文档features/phone-qol.md8.5 KB
- 参考文档features/phone-routines.md2.4 KB
- 参考文档features/phone-usage.md3.4 KB
- 参考文档features/phone-workspace-navigation.md5.6 KB
- 参考文档features/private-input-origin.md4.5 KB
- 参考文档features/profile-backup.md3.4 KB
- 参考文档features/README.md17.6 KB
- 参考文档features/release-identity.md3.5 KB
- 参考文档features/remote-hosts.md27.0 KB
- 参考文档features/ssh-host.md3.7 KB
- 参考文档features/storage-cleanup.md4.1 KB
- 参考文档features/subagent-presets.md3.3 KB
- 参考文档features/team-mode.md5.8 KB
- 参考文档features/timed-question-recovery.md2.4 KB
- 参考文档features/tob-home.md33.0 KB
- 参考文档features/tool-approval-simple.md6.0 KB
- 参考文档features/video-wallpaper.md2.4 KB
- 参考文档features/vrm-companion.md8.8 KB
- 参考文档features/workspace-grant.md7.2 KB
- 参考文档features/workspace-request-write.md5.1 KB
- 参考文档references/ios-chat-resume.md2.7 KB
- 参考文档references/observation.md3.7 KB
- 参考文档references/performance.md12.5 KB
03 SKILL.md 正文
Verify DeepSeekBot
Unit tests never boot DSH. A claim that a feature works needs a run against a real Host, driven through the user path, with saved evidence. scripts/verify-app.mjs is the lever for that run; node scripts/verify-app.mjs --help lists its flags.
Each run uses one named instance (--name <n>, default default). Its state and evidence live in .humanlayer/verify/<n>/ (ignored by Git). Give each task its own name so two agents never drive the same Host.
Launch
pnpm build
node scripts/verify-app.mjs launch --name <n>
launch wraps scripts/dev-instance.mjs with a fresh DSH_HOME under the OS temp dir and a free port, then records the launcher JSON in .humanlayer/verify/<n>/instance.json. It refuses to start a second Host for a running name. Ready means health.ok is true in its output.
The model field says which model credential the Host inherited. Without one (missing (model calls will fail)), every UI and API path works, but no Bot replies. Features that need a reply say so in their file. Credential setup is in docs/agents/ax-model.md. In a cloud container whose proxy injects the opencode.ai credential, OPENCODE_GO_API_KEY=proxy-injected node scripts/dev-model.mjs setup --from-env once is enough.
Doctor
node scripts/verify-app.mjs doctor --name <n>
It checks that the recorded PID is alive, the Host runs from this worktree with the pinned DSH version, and an authenticated /api call succeeds. It exits non-zero when any check fails. Run it before the first drive and again after any drive that fails or surprises you. On failure, collect the observation below before stopping this task’s Host; preserve the failure evidence before relaunching.
Observe
node scripts/verify-app.mjs observe --name <n>
Run after a failed or surprising drive, with the page still open. It saves Host health and browser-reported Client diagnostics as timestamped JSON. Exit 0 means the shell was observed; 1 means failure; 2 means incomplete observation. Feature verification still needs the user action, visible result and canonical read. Follow the observation guide to inspect Bot activity, tool traces and runtime logs without asking the Human to copy them.
Drive
Drive through the Web Client with shot. Each call opens a fresh headless browser, logs in with the launcher URL, dismisses the preview notice, the API key prompt and the tutorial (--keep-overlays keeps them), runs the actions in order, then captures:
node scripts/verify-app.mjs shot <label> --name <n> --bot-mode \
--fill "Message DeepSeek Bot=What is 6 times 7? Answer with digits only." \
--press Enter --wait-text 42
- Address elements by accessible name (
--click,--fill,--wait). Read names from the last<label>.aria.txt. When a name carries dynamic text, such as a Channel row with its last message, use--click-textwith a stable part. --right-click <name>opens an element's context menu, such as a Channel row's menu in the Messages sidebar; follow it with--clickon a menu row.--wait-textmatches rendered text only, not the composer value. Pick a reply that the prompt does not contain.- Use
rpc <namespace/method> [json]to read canonical state (Channel timelines, Bot lists) after a UI action, or to seed preconditions through the same public API the Client calls. It is evidence of state, never a substitute for the UI step it follows. - For scripted journeys, use the feature file’s launch contract. Some accept
BH_E2E_INSTANCE=.humanlayer/verify/<n>/instance.json, the model suite accepts--name <n>, and others own their Host; check before running so the journey targets this task’s instance.
Measure performance
For a change that touches the timeline, the composer or message rendering, or a report that
Bot mode is slow, hot or laggy, measure before and after. Native iOS follows
iOS performance verification; the Web path uses
scripts/perf-idle-timeline.mjs.
It seeds a long Group without a model, mounts the whole history, and reports idle main-thread
time, typing latency per key and the hottest functions. Follow
the performance guide for the A/B protocol and how to read it.
Add --budget scripts/perf-budgets/timeline.json so the run fails on a regression; the rules
for when to measure and what evidence a PR needs are in
performance observation and verification.
Native mobile
The Web Client drives above do not cover the Expo app in apps/mobile. Before starting a
Simulator, Metro, a native build or a QA Host, follow shared device QA.
Submit the feature's steps to the queue; only its designated runner drives the shared devices.
The Mac runner uses scripts/e2e-mobile-sim.mjs and the Maestro flows in apps/mobile/.maestro/
(phone-notifications.md). Read every captured screenshot.
The Windows runner uses Android device QA and
scripts/android-qa.mjs for its queued emulator/USB-phone lane, XML snapshots and scrcpy video.
Web-export checks qualify their named web paths; native-only behavior remains untested until
its device result is recorded.
Evidence
shot writes <label>.png, <label>.aria.txt and <label>.console.txt to .humanlayer/verify/<n>/evidence/, and <label>.failure.png plus <label>.failure.aria.txt when a step fails. Its JSON output lists every step it performed.
A proof has three parts: the user action, the visible result, and the canonical state behind it (a timeline entry, a persisted Bot, a Git Memory HEAD). A successful API call alone does not prove rendering, and a screenshot alone does not prove persistence. Read the console file: a page error during the journey is a finding even when the screenshot looks right. PR screenshots follow docs/agents/pr-ui-visual-evidence.md: publish reviewed derivatives to Cloudflare R2 using pnpm evidence:publish, never commit evidence media to Git or an evidence branch, and remove disposable local copies only after successful publication and verified public readback, and only when no active QA/release hold remains. Merging a PR does not replace either prerequisite; disposable originals additionally wait until merge. Keep URLs, hashes and provenance in the small task index.
Cleanup
Cleanup is required after successful and failed verification, before task completion or handoff. Follow verification resource cleanup for ownership checks, interrupted runs, retained environments and shared E2E/QA protection. Use the command below only for this task's verified instance with a disposable Profile after confirming that no child or other process uses that Profile. The helper waits only for the recorded Host before removing the directory. If other users exist or ownership is uncertain, use separate shutdown and deletion steps from the runbook. If Profile state must survive, stop the verified Host separately and preserve its closed Profile.
node scripts/verify-app.mjs stop --name <n>
It signals the recorded PID, waits for exit, deletes that instance's matching temporary
DSH_HOME and leaves the evidence directory in place. Recheck that the process and its owned
children have exited, its port is free and the evidence files still exist after stop.
Feature map
features/README.md indexes the mapped features. Before you verify a change, read the file for each feature the change touches and drive every entry point it lists. When the app no longer matches a file, fix the file in the same change, or report the product regression. A journey that cannot pass keeps its failure and records the reason in Gotchas.
When changing a user entry point, add or update its feature file and index row in the same PR.
Name every changed sub-feature and its visible/canonical proof. Run
node scripts/check-feature-map.mjs (also in lint): it checks index coverage, required sections,
local links and script paths, and inventories unmapped journeys. It cannot prove that selectors,
product semantics or runtime behavior are current. Re-run the touched journey and record its
commit/environment/result in the PR; keep unrun branches explicitly untested.
04 同仓库更多 Skills
slide-authoring
Technical reference for writing or editing open-slide pages — file contract, 1920×1080 canvas, type scale, layout, palette/visual direction, assets, speaker notes, stepped reveals, page transitions, and morph transitions. Consult this whenever you are about to write or modify any file under `slides/<id>/`, including from inside the `create-slide` or `apply-comments` workflows, or for any ad-hoc slide edit. Triggers on phrases like "edit slide", "tweak this page", "fix the layout", "change the palette", "reveal one by one", "add a transition", "morph transition", "write the speaker notes", "generate the script / talk track", "investigate the slide framework", "how do slides work here".
current-slide
Resolve which slide, page, and (optionally) selected element the user is currently viewing in the open-slide dev server. Consult this whenever the user references "this page", "this slide", "this element", "the slide I'm on", "the current page", or any deictic reference to slide content without naming it. Re-read `node_modules/.open-slide/current.json` at the start of every such turn — the user navigates between turns, so a value you read earlier in the conversation is almost certainly stale.
create-theme
Use this skill when the user wants to create, draft, author, or extract a slide theme in this open-slide repo. Triggers on phrases like "create a theme", "make a theme called X", "extract a theme from <slide>", "build a theme from these images". Produces two paired files under `themes/` — `<id>.md` (palette, typography, layout, fixed Title/Footer components, motion) and `<id>.demo.tsx` (a runnable demo slide that the dev-UI Themes panel previews). Do NOT use for editing real slides — only for authoring the theme bundle.
create-slide
Use this skill when the user wants to create, draft, author, or generate new slides / a presentation in this open-slide repo. Triggers on phrases like "make slides about X", "make a deck about X", "create a presentation", "draft slides for", "new slide", "a deck with speaker notes / a script", or when the user asks to add content under `slides/`. Do NOT use for editing the framework itself — only for authoring content inside `slides/<id>/`.
apply-comments
Apply pending @slide-comment markers written by the open-slide inspector tool. Use when the user asks to "apply comments", "process slide comments", "apply the inspector comments", or references markers left inside `slides/<id>/index.tsx`.
writing-shape
Writing, exploit: shape raw material into an article, paragraph by paragraph.