zhu1090093659/dsh-web v0.4.17

dsh-web v0.4.17

on GitHub

This release raises the host floor to @deepseek-ai/dsh >= 0.2.2-rc.2: upgrade the host first, then the plugins. Install it with npm i -g @deepseek-ai/dsh@next and confirm with dsh --version, because npm's latest still points at 0.2.2-rc.1 and a host installed from latest does not satisfy this release's floor - its update button refuses the upgrade protectively. Beyond that, this release moves browser-view idle page reclamation, LiangShen's on-demand tool-family paging and the Plugins page batch update one step each, and removes the task board's /goal run option and goal acceptance; for anyone already on 0.2.2-rc.2 this release is also a fix, because 0.4.16's web-ui-agents row fails to import on an rc.2 host. The four satellite repositories align to 0.4.17.

The host floor rises to @deepseek-ai/dsh >= 0.2.2-rc.2; upgrade the host first: dsh.engines.dsh and the @deepseek-ai/dsh peer dependency of every family package move from >=0.2.2-rc.1 to >=0.2.2-rc.2, and the same move carries every workspace @deepseek-ai/* range, the DSH_CLI_VERSION pin in CI and the release mount, the badges of both root READMEs and all 32 package READMEs, and the lockfile. Upgrade the host first: npm i -g @deepseek-ai/dsh@next, confirm with dsh --version that it reads 0.2.2-rc.2, then update the family on the Plugins page. npm's latest still reads 0.2.2-rc.1 (rc.2 ships on the next tag only), so a host installed from latest does not satisfy this floor: if dsh --version still reports 0.2.2-rc.1, the host upgrade has not landed yet and the plugins should wait.
On an rc.1 host the update button shows a protective refusal, which is expected rather than a fault: the update request resolves the target package's declared host floor before touching anything and answers HTTP 412 when the running host does not satisfy it, showing ... requires DSH >=0.2.2-rc.2 (current DSH 0.2.2-rc.1); upgrade DSH before updating (the host-side full form is plugin-manager: <package> requires DSH >=0.2.2-rc.2 (current DSH 0.2.2-rc.1); upgrade DSH before updating). The refusal happens before any write, so nothing is written: the profile, the packages on disk and the task board ledger are untouched, and retrying after the host upgrade is all that is needed.
If 0.4.17 was installed by hand on rc.1, dsh web still starts but the whole dsh-web plugin set does not load: the official boot gate skips bundles that do not satisfy the floor, one bundle at a time, and stderr carries a single line, dsh: skipping profile bundle "@linxin666/dsh-web-all": <reason> (the format is <binName>: skipping profile bundle <JSON package name>: <reason>). Nothing is thrown, no data is deleted, and the remaining bundles boot normally. Raise the host to 0.2.2-rc.2 and restart to recover; the profile needs no edit. Two stopgaps exist: grant the exact-version exemption with dsh plugin --profile <profile> allow-version <package>@<version> --dsh-version 0.2.2-rc.1 --accept-risk, which applies only to that package and that DSH version, or remove the package and roll back to 0.4.16.
For hosts already on 0.2.2-rc.2, this release is a fix: rc.2 removed the @deepseek-ai/dsh-subagent/internal subpath export, and 0.4.16's dsh-agents imported it statically at module top level, so the aggregate's web-ui-agents row threw ERR_PACKAGE_PATH_NOT_EXPORTED: './internal' at import time and was recorded as a degraded row at stage: import - that row's UI and tools were missing while the other rows kept working. This release probes Symbol.for('dsh.subagent.deliverPrompt') first and falls back to subagents.sendMessage(caller, target.id, content, { signal }), so one artifact loads and works on both host generations.
Browser view: pages idle for three days close themselves: TabRecord.lastActiveAt is stamped when a target is adopted and refreshed only by a browser_* call that resolves to that page; BrowserViewService.sweepIdle runs every ten minutes (an unref'd timer, cleared on dispose) and closes pages past the window through the existing closeTab path - the session binding pointing at it is dropped, a panel pinned to it falls back to the remaining tabs, and panel state is broadcast; nothing is scanned and no browser is launched when none is running. The window is the new idleCloseDays setting: default 3 days, minimum 0, 0 disables the capability, read on every sweep, so an edit takes effect on the next sweep without remounting the plugin or reconnecting the browser. Adopted real-browser pages are never candidates: only tabs the plugin created (origin: 'created') qualify, and watching a page in the panel, selecting its chip or projecting its frames does not count as activity. The idle clock lives in memory only, so a host restart restarts every page's window.
LiangShen pages plugin tool families by default: the shipped pagedToolPatterns grows from the MCP bridge alone to ['mcp__*', 'browser_*', 'task_board_*', 'team_*', 'squad*', 'ui_*', 'ssh_*'], while host builtins (shell, the file system, glob/grep, background jobs, skill, ask_user_question, todo_write, web_search, subagent delegation, plan, goal, fact_register, tool_activate and the MCP resource readers) stay resident. A withheld family is announced as one namespace summary in the catalog and needs tool_activate({ namespace }) before its first call: mcp__* takes the first __ segment of the tool name (mcp__codegraph__codegraph_explore becomes codegraph) while the others derive a literal-prefix family (task_board_* to task_board, squad* to squad), and a bare * is refused. At most three activated namespaces are held at once, and activating a fourth returns the least recently used one to its summary; under ptc those summaries are genuinely unreachable - absent from the request's tools:sdk section and not callable inside a run_code program. The settings card exposes the list as the on-demand tool family field, and writing it empty removes the grouping and returns every tool to the resident face (all three rows that page from that one pattern list are rewritten together). Measured on one session, the system prompt dropped from 104,949 to 58,961 characters and the tools:sdk section from 73,675 to 27,666.
The task board loses the /goal run option and goal acceptance: goalRunEnabled and TaskRecord.goalRun, the /goal dispatch, the session/projections goal-phase settlement, the five goalVerification* keys, TaskRecord.skipVerification, the ExecutionRecord.verification contract, the update_goal(action: complete) gate on tools/pre-execute, the judge-model runner and task_board_manage(action: 'reset-verification') are all removed; manual runs, scheduled triggers and rework now run as an ordinary single turn and settle from that turn's turn/end result, with no extra judge-model calls. Existing cards and execution history stay intact: the authoritative ledger $DSH_HOME/task-board/ledger-v2.json keeps schemaVersion: 7, leftover goalRun, skipVerification and verification fields on old rows are dropped at load rather than rejecting the row, and they are never written back; non-strict plugin config loading likewise ignores leftover keys in old profile patches. Passing reset-verification to task_board_manage is now refused as an unknown action.
Plugins page batch update: one request runs the whole batch, and a refused row does not stop the others: updates now go through the official updateBundle(name, version, { requestId }) primitive (the manager verifies the target, removes the installed version, installs with the selection it found and restores on failure), so a batch is this host running that one primitive over its rows in order and the browser sends one request and polls one job. Each row carries its own result (phase, targetVersion, landed, error, restartRequired), the request resolves every id by the single route's own rules before touching anything, and a refusal becomes that row's error instead of failing the batch; a batch is capped at 50 rows, duplicate ids resolve once, and the batch always settles after the last row answers. A package that landed but cannot be loaded in this process is a restart row rather than an error, named on both surfaces - a restart of DSH may still be needed before the new version loads.
A fresh install that reports ERR_PNPM_IGNORED_BUILDS needs the builds approved and the command re-run: on a profile that has never approved build scripts, pnpm ends with [ERR_PNPM_IGNORED_BUILDS] Ignored build scripts: cloudflared@..., cpu-features@..., ssh2@... plus Run "pnpm approve-builds" to pick which dependencies should be allowed to run scripts., dsh plugin add exits 1, and the package is not added to dsh.profile.bundles - installed but not in effect. Run pnpm approve-builds for those three dependencies and re-run the same dsh plugin add.

Upgrade Notes

• The host floor moves from >=0.2.2-rc.1 to >=0.2.2-rc.2: this release raises dsh.engines.dsh and the @deepseek-ai/dsh peer floor, so a host on 0.2.2-rc.1 no longer qualifies. The host goes first, and npm's latest is still 0.2.2-rc.1: install it from the next tag (npm i -g @deepseek-ai/dsh@next) and verify with dsh --version.
• The refusal on an rc.1 host is protective by design: HTTP 412 plus requires DSH >=0.2.2-rc.2 (current DSH 0.2.2-rc.1); upgrade DSH before updating, raised before any write, leaving no half-installed state. Seeing it means the host has not been upgraded yet, and nothing about the plugin needs investigating.
• Recovery after a hand install of 0.4.17 on rc.1: dsh web still starts, stderr carries a single dsh: skipping profile bundle ... line, and the plugin set does not load; raising the host to 0.2.2-rc.2 and restarting recovers it without touching the profile. The stopgaps are dsh plugin --profile <profile> allow-version <package>@<version> --dsh-version 0.2.2-rc.1 --accept-risk for that exact combination, or removing the package and rolling back to 0.4.16.
• The web-ui-agents import failure on rc.2 hosts is fixed here: 0.4.16 lost that row on rc.2 to ERR_PACKAGE_PATH_NOT_EXPORTED once the ./internal subpath disappeared; the dual-cohort probing in this release makes one artifact load on both rc.1 and rc.2.
• Browser view gains idleCloseDays, default 3 days: only pages the plugin created are reclaimed, adopted real-browser pages are untouched, 0 disables the capability, and the idle clock lives in memory so a host restart restarts the window. A page watched in the panel with no tool call still closes when its window expires.
• LiangShen's tool face is grouped by default: the shipped page list is mcp__*, browser_*, task_board_*, team_*, squad*, ui_* and ssh_* while host builtins stay resident; a withheld family is unreachable before tool_activate({ namespace }), at most three stay activated, and writing the on-demand tool family field empty restores the full resident face.
• Retired task board fields and ledger compatibility: this release removes the /goal run option and goal acceptance while the ledger schema stays v7, existing cards and execution history stay intact, retired fields are dropped at load and never written back, and no migration step is needed.
• Batch update and the restart fact: the Plugins page's update all is now one request and one job, a refused row does not stop the others, and a package that landed but cannot load in this process is explicitly marked as needing a DSH restart. The hand-written remove-then-install replacement is gone, and the layer order (the position in dsh.profile.bundles) survives consecutive updates.
• The fresh-install ERR_PNPM_IGNORED_BUILDS trap: dsh plugin add exits 1 without adding the bundle, so run pnpm approve-builds first and retry.
• Platform verification: this release ran the repository gates on Linux (Ubuntu x86_64) against a pinned rc.2 host tree (typecheck, build, test, test:scripts, libs:check, aggregate:check, docs:check, i18n:check, emoji:check among them). Windows and macOS were not verified on physical hardware. The host running at release time was not restarted, so the real GUI observation of the new idleCloseDays settings card and the batch update toolbar lands only after the host is restarted and the page refreshed.

New Features

• [plugin-manager] Batch updates go through the official updateBundle as one ordered run in this host: one request, one job, one result per row, a refused row leaves the rest running, and a restart-required row is recorded as a landed fact
• [remote-bridge] Add session.attachment.path, serving the bytes of one image by the absolute path that session's own transcript named, with a whitelist built only from that session's offload placeholders and the image references its durable tool results published; a path is served only when it matches byte for byte, realpath() resolves it to itself, a regular file is there and its first bytes match one of four image signatures (the extension is never consulted)

Bug Fixes

• [agents,liangshen] Bridge the runtime contracts of a DSH 0.2.2-rc.2 host: dsh-agents drops its static @deepseek-ai/dsh-subagent/internal import and probes both delivery paths, passes both top-level fields and the nested request to startActivation, accepts drainChildren, drainDescendants and started.dispose(), and steps aside while the official agent-team row is present; dsh-liangshen's dispatch guard falls back to subagents.rootOf(caller) when agents.roots() is absent
• [sdk] Carry @deepseek-ai/dsh-error@0.2.2-rc.2 as a root host face, pin zod to 4.6.6 in dsh-agents and dsh-remote-web-ui, and have dsh-agents supply the Team identity prompt-context position 125 that rc.2 dropped from the central order table
• [agents] Name the refused field when a member or squad draft is not usable: gateMemberRequestBody and gateTeamRequestBody return one reason instead of the fixed "not a usable definition", so a client that renames or adds a draft field can tell which one was not accepted
• [remote-bridge] Read the JSON image reference of a durable tool result too ({ path, image: { ... } }, the persisted shape of a run_code program's own return value), gated on an image object whose mediaType starts with image/ so a file a tool merely named never enters the whitelist
• [sdk] The ERR_PACKAGE_PATH_NOT_EXPORTED: './internal' import failure of the aggregate's web-ui-agents row on a 0.2.2-rc.2 host, removed by the two dual-cohort changes above

Behavior Changes

• [browser-view] Pages the plugin created close themselves once idle past the window: the window is the new idleCloseDays setting (default 3 days, minimum 0, 0 disables), read per sweep; adopted real-browser pages are never candidates and watching a page in the panel is not activity; the clock lives in memory, so a host restart restarts it
• [liangshen] On-demand tool families ship enabled: pagedToolPatterns defaults to ['mcp__*', 'browser_*', 'task_board_*', 'team_*', 'squad*', 'ui_*', 'ssh_*'] while host builtins stay resident, a withheld family is absent from the wire and the generated SDK and cannot be called inside a run_code program before tool_activate({ namespace }), at most three stay activated, and writing the list empty in the settings removes the grouping
• [task-board] The /goal run option and goal acceptance are removed: execution is an ordinary single turn and settlement reads turn/end, with no judge-model calls and no reset-verification unstick action; the ledger schema stays v7, existing cards and execution history stay intact, and retired fields are dropped at load and never written back

Other Changes

• [market] Pin the 2026-11-08, 11-09, 11-10 and 11-11 satellite registration rounds and this round's ascending-id index rebuild into the submodule gitlinks and regenerate market/dist, leaving the store catalog at 66 skins, 12 pets, 124 plugins and 33 presets
• [task-board] Rename the detail run fold key from detail.settings.advanced to detail.settings.run and give it the create dialog's title (Run Mode / 运行方式 / Режим запуска), keeping the fold and its single session-reuse switch
• [team-hub] Rewrite seven places in PROTOCOL.md so settlement follows the host's own execution verdict rather than automated acceptance, with field names, the settle endpoint signature, the in_review/failed states and every numeric constant unchanged
• [docs] Drive browser and GUI verification with the session's browser_* tools, keeping Playwright as the fallback and as the committed mount gate pnpm test:mount, and record the owner of that default in a decision note
• [lib] Rebuild the committed lib/ artifacts of dsh-market and dsh-web-all with the changes in this range
• [satellites] Advance the four submodule gitlinks (dsh-skins, dsh-pet, dsh-community-plugins, dsh-presets) to their 0.4.17 release commits
• [remote-bridge] Name the attachment gate cases by the role the gate requires (a test rename only, no behavior change)
• [notes] Record the 0.2.2-rc.2 cohort and its native-first overlap inventory, among other decision notes