feat: rebuild task workspace and guided setup, recovery, and review

docs/FRONTEND_REBUILD_STATUS.md

@@ -0,0 +1,134 @@
+# Frontend rebuild status
+
+Baseline: main `71ede3053b1d71f4c6e4f7053b355ebe301c1a1b`, checked on
+2026-10-10. The roadmap is not complete. Merged PRs #288 and #293 finish
+the first recovery/connection fixes; #286 contributes scoped journey fixes.
+Passing those slices does not establish complete product or platform acceptance.
+
+## Product guidance revision — 2026-10-10
+
+The user's latest direction treats language and information architecture as part
+of the rebuild. Everyday screens should lead with the current state, the action
+available to the user and the next step. Replace defensive explanations and
+chains of negations with concise guidance tied to real controls. Keep decision
+information, such as the effect of deleting a credential or an unresolved write,
+at the relevant action; put implementation details in expandable diagnostics.
+
+Three parallel lanes updated task/recovery flows, connections/extensions and
+review/diagnostic tools. The connection hub now groups task preparation, tools
+and collaboration, and runtime settings around a return-to-task action. MCP,
+Plugin and LSP setup show their actual stage and next step. Empty and error
+states point to available actions; model, Docker and Safe Web refreshes read
+state without repeating a write. Chinese and existing English translations are
+updated together. [Product language guidance](PRODUCT_LANGUAGE.md) records the
+conventions for future work.
+
+Implementation details use expandable records. Commit hook/signing effects,
+shared credential impact, partial file changes and approval scope remain visible
+at the relevant decision. Independent review corrected guidance for staged MCP
+credentials, configured reasoning effort, unavailable browser/UI Evidence
+capabilities and already-applied or denied revert proposals. Existing identity,
+approval, recovery and capability checks remain the behavioral acceptance criteria.
+Model selection waits for refreshes and pending switches to finish, preserving
+the in-flight write state. GitHub credential guidance follows the connection's
+actual credential type, enabled state and system-store availability.
+
+Verification for this revision:
+
+- Full Vitest: **186 files, 1,834 tests passed** on the integrated code. The full
+  run initially found cross-component assertions referring to old text; their
+  selectors were updated while retaining request, binding, recovery and authority
+  assertions. The final full run passed without retries or increased timeouts.
+  Deferred-response regressions cover both refresh/switch orderings; GitHub
+  connection tests cover missing PAT/OAuth credentials and unavailable storage.
+- TypeScript, production build and generated API/transcript-key checks passed.
+  Existing large-chunk build warnings remain.
+- An intermediate CI run interrupted the baseline Windows fixed-command paging
+  test before its command timeout. The same seven native subcases passed locally;
+  the interruption cause remains unconfirmed. Failure-only authority/renewal
+  timing diagnostics now preserve the next failure's evidence, with the existing
+  permissions, time budgets, output and replay assertions unchanged.
+- Production assets served by the real Go API and isolated test database were
+  checked at 1280px and 390px. Verified connection navigation, immutable saved
+  budgets, staged MCP credential guidance and read-only refresh, model status,
+  Copilot return-focus, task draft retention and permission-menu Escape/focus.
+  Connection, extension and model pages had no horizontal overflow at 390px;
+  browser error logs were empty. Original captures are retained under ignored
+  `build/frontend-rebuild/guidance-*` paths.
+- The browser pass used a local fixture with no real model credentials. Credential
+  forms and approval gates were inspected; remote authentication, MCP discovery,
+  permission changes and remote writes were not exercised. Native platform and
+  complete model-to-delivery acceptance remain in the roadmap below.
+
+## Current parallel implementation
+
+All implementation worktrees start from the same baseline. Commits use Qiyuanqiii.
+The integration owner reviews and combines the independently tested changes.
+
+| Lane | Implementation scope | State |
+| --- | --- | --- |
+| Workspace | Direct task navigation, context/recovery and observation entries, file/terminal/preview switching, existing connection/environment entries | Implemented and locally verified |
+| Task configuration | Shared Go budget/project-config resolution, immutable creation and successor contracts, safe preview and creation UI | Implemented and locally verified |
+| MCP credentials | Go-owned credential presence/write/delete for supported MCP descriptors, first-entry/update/removal UI | Implemented; real OS lifecycle/remote authentication not exercised |
+| Review and Git | Existing PR thread/reviewer operations through exact previews and approvals; registered bisect recipe selection and bounds | Implemented; remote writes tested with fixtures only |
+
+The existing Thread/Run identities, draft and attachment recovery, original
+operation keys, approval bindings, read-only restrictions, process capability
+gates, historical routes and conditional module loading remain acceptance
+requirements. No renderer-owned execution or credential authority is introduced.
+
+## Remaining roadmap
+
+- P1/P2: The slices above have contract/recovery tests and production-browser
+  checks; the full configured-model-to-delivery journey still needs acceptance.
+- P3: Task/project configuration, manual HTTPS MCP bearer credentials and PR
+  discussion/reviewer actions are implemented. Cost caps require a price snapshot
+  and do not represent exact billing. Project exclusion paths, suggested Skills
+  and typed command IDs remain recorded metadata without automatic filtering,
+  installation or execution; see ADR 0170.
+- P4: Plugin version history/rollback/trust revocation, Hooks diagnostics, Docker
+  onboarding, guided UI Evidence, Batch preparation/rework and the scoped macOS
+  runtime assembly remain separate unfinished work.
+- P5: Full project/model-to-delivery, restart/disconnection, long-session,
+  keyboard/responsive and native cross-platform acceptance is still outstanding.
+
+## First implementation evidence
+
+Local verification used Go 1.26.9 and the repository's pinned frontend packages:
+
+- Full Vitest: **184 files, 1,822 tests passed**. The first combined runs exposed
+  missing locale providers in new test fixtures and cold real-module compilation
+  consuming the recovery test's action timeout. Both were corrected without
+  mocking the recovery callback chain or increasing global timeouts; the final
+  full run passed.
+- TypeScript and production build passed. Settings, configuration and diagnostic
+  surfaces remain lazy-loaded. Existing large-chunk build warnings remain.
+- Combined Go checks passed for bounded budgets, initial/restart creation replay,
+  changed-budget conflict, project rejection, successor inheritance, v185-to-v186
+  upgrade, invalid direct SQL inserts, MCP binding and HTTP/OpenAPI contracts.
+  Full projectconfig, protocolregistry, surfacegovernance, releasegate and
+  producte2e packages passed. Focused MCP race and injected-store Desktop checks
+  also passed. This is not a claim of `go test ./...` on every platform.
+- Generated OpenAPI/TypeScript and protocol history checks retain old readers and
+  default fingerprints. Independent review found and repaired legacy zero/large
+  budget reads, explicit-null parsing and known-rejection retention across draft
+  edits/remounts.
+
+A real browser used production assets served by an isolated loopback Go process
+and a separate test database. Verified task/context navigation, return-draft
+retention, file/preview entries, a 390-pixel layout, configuration source display,
+reconnection, immutable saved budgets after the project file changed, and the
+exact-Run GitHub entry. A staged manual HTTPS MCP descriptor displayed actual
+Windows Credential Manager availability and absence of its dedicated test token.
+No secret was entered, OS credential changed, MCP discovery approved or remote
+GitHub write executed.
+
+The isolated service deliberately had no configured real model. Both UI and HTTP
+refused new Thread creation until model setup. A CLI-created fixture supplied the
+saved-budget browser check; HTTP creation success is covered by Go/client tests,
+not asserted as a real-account browser journey. Original browser JPEGs and local
+logs are retained under ignored `build/frontend-rebuild/`; they are not generated
+illustrations or repository fixtures.
+
+These checks do not stand in for native desktop, real remote account, Docker or
+macOS acceptance. The overall rebuild remains incomplete.

docs/PRODUCT_LANGUAGE.md

@@ -0,0 +1,45 @@
+# 产品语言与操作引导
+
+Universal Code 面向在项目中完成任务的人。页面说明帮助用户理解当前状态、
+选择操作并继续工作。文案与入口、表单、结果反馈一起设计和验收。
+
+## 页面结构
+
+1. 用用户要完成的事情命名入口，例如“连接模型”“查看改动”“恢复任务”。
+2. 页面开头用一句话说明当前可以做什么，把主要操作放在说明旁边。
+3. 空状态给出第一步，失败状态给出原因和页面中实际可用的下一步。
+4. 多阶段功能显示当前阶段和后续操作。例如 MCP 接入依次完成登记、
+   凭据配置、发现、审查，然后回到任务使用。
+5. 将 ID、指纹、协议版本和收据细节放在可展开的“详细记录”或诊断区域。
+   与选择直接相关的项目、模型、费用和操作范围保持可见。
+
+## 表达方式
+
+优先采用“当前状态 → 下一步”的短句，按钮名称与引导保持一致。
+用肯定句描述实际作用与生效范围，把有关实施方法的说明留在工程文档中。
+
+| 场景 | 用户看到的指引 |
+| --- | --- |
+| 任务创建结果待确认 | 原请求已保存。选择“核对结果”，继续处理这次请求。 |
+| 模型列表读取失败 | 模型列表读取失败，请选择“重新读取模型”。 |
+| Plugin 导入完成 | 安装包已导入。查看包内容，选择要启用的能力。 |
+| MCP 凭据保存成功 | 本机凭据已保存。继续发现服务器，检查连接结果。 |
+| 查询输出只有一部分 | 当前显示已保留的输出片段。展开命令记录查看来源。 |
+| 查看预算历史 | 本次执行使用创建时保存的预算。后续项目修改用于新任务。 |
+
+文案跟随实际接口和状态。例如保存本机凭据与验证远端连接分别显示各自结果；
+写入结果待确认时，下一步应读取原操作结果。重新提交、继续执行和删除等
+操作使用各自已有的确认与恢复流程。
+
+保留用户作决定需要的信息：删除的具体对象、共享凭据影响的连接、授权范围、
+费用来源和生效时间。使用具体对象和直接语句，让用户在操作处理解影响。
+
+## 一致性与验收
+
+- 常用页面使用“任务”“执行”“项目”；内部标识在详细记录中展示。
+- 已有中英文内容同步维护，相同操作沿用相同按钮和结果名称。
+- 每条指引对应当前状态下可访问的入口或实际可执行的步骤。
+- 原有查询、权限、审批、绑定和恢复行为继续通过相关交互测试。
+- 在真实浏览器中检查主流程、空状态、失败状态、键盘操作及窄屏布局。
+
+这一约定随前端重建维护。产品体验以实际操作和用户反馈为准。

docs/PROJECT_MEMORY.md

@@ -1,5 +1,27 @@
 # Prayu Project Memory
 
+> 2026-10-10 product guidance revision: the frontend rebuild includes user-facing
+> language and information hierarchy. The connection hub groups task preparation,
+> tools/collaboration and runtime settings; MCP/Plugin/LSP show actual setup stages.
+> Task, approval, recovery and delivery copy explains the current state and next
+> available action. Technical records expand on demand, while decision effects
+> stay visible. See [product language](PRODUCT_LANGUAGE.md) and the
+> [integrated verification](FRONTEND_REBUILD_STATUS.md): 186 frontend files / 1,834
+> tests, production build and an isolated real-browser pass. Remaining P4/P5 work
+> continues under the same rebuild roadmap.
+
+> 2026-10-10 frontend rebuild checkpoint: V2 exposes task navigation, exact-Run
+> recovery/observation, connection management and PR discussion actions. New task
+> drafts carry bounded budgets; Go resolves narrowing-only project configuration
+> and pins the result through schema v186, creation replay and successor Runs.
+> Supported manual HTTPS MCP registrations can manage bearer credentials through
+> the shared Go API and OS store. Approval, scope and runtime capability gates
+> remain independent. See [implementation status and evidence boundaries](FRONTEND_REBUILD_STATUS.md)
+> and [task configuration ADR](adr/0170-task-budget-and-project-configuration.md).
+> This checkpoint does not complete P4 environment/extension workflows or native
+> cross-platform acceptance. Project exclusion paths and suggested skills/actions
+> remain recorded metadata, without automatic filtering, installation or execution.
+
 > 2026-10-09 recovery and connection workflow checkpoint (#292): Continuity
 > Fork/Resume and checkpoint Fork open the returned Run in V2 without inheriting
 > a source Thread binding. Late completion refreshes record and Thread lists

docs/adr/0170-task-budget-and-project-configuration.md

@@ -0,0 +1,100 @@
+# ADR 0170: Product task budgets and immutable project configuration
+
+Date: 2026-10-09
+
+## Status
+
+Accepted for schema v186, shared Go product creation, read-only previews and the
+standalone task configuration presentation. Application routing is integrated at
+the existing Thread creation seam; this ADR grants no execution capability.
+
+## Contract
+
+HTTP and in-process Desktop `run_creation.v1` / `thread_creation.v1` requests may
+include `budget` with optional `max_turns`, `max_tool_calls`, `max_tokens`,
+`max_cost_usd` and `timeout_seconds`. Go canonicalizes omitted fields to the
+existing defaults. Turns remain 1..10,000; tools remain 1..1,000,000; tokens are
+0..1,000,000,000; cost is 0..100,000 USD; timeout is 0..604,800 seconds. Zero
+disables an optional token/cost/timeout dimension. Positive cost must be at least
+one micro-USD and is rounded once to micro-USD, matching the monetary ledger.
+These are execution limits, never exact billing estimates or permission grants.
+An active operator price snapshot remains necessary before a cost-capped model
+call; task creation and preview make no model or account calls.
+
+The normalized operator ceiling is pinned as `RunConfig.requested_budget`;
+`Run.Budget` contains the effective narrowed limits. The canonical creation
+fingerprint binds that operator input, independently of the live repository
+file. Default-budget requests retain historical fingerprints. Initial-state
+replay checks the durable operation before loading project files and returns the
+original snapshot across restart. Changed explicit budgets under the same key
+conflict. The existing refusal to replay a creation after its Run has advanced
+remains; original-request observation provides that later recovery path.
+
+Web/Desktop creation and Standard Code's initial Run use the same Go loader and
+resolver as CLI: `.prayu/config.yaml` is bounded, strictly decoded, inert input;
+profiles and read-only selection restrict admission and project budgets must be
+strictly below the operator ceiling. Any rejected field blocks the complete
+creation. No HTTP/Desktop ignore switch is added. The CLI's explicit
+`--ignore-project-config` contract remains. The selected profile is checked
+against the project profile set, which may contain several supported profiles.
+This adds no global or user configuration file or renderer-selected host path.
+
+Thread successors and explicit continuity branches inherit the stored effective
+budget, normalized operator ceiling and project snapshot. They never reload
+changed project configuration. Their existing authority reset is unchanged.
+
+Schema v186 replaces only the controlled-creation trigger's fixed budget/root
+limits. Its Mission/Session/mode/root/event/network and all-denied execution
+checks are preserved. The new branch verifies bounded requested limits and
+exact requested/effective/project relationships. New product snapshots reject
+budget or project replacement in SQLite. Historical migrations remain frozen;
+v185 upgrade preserves old Runs and creation operations without fabricated
+operator input. Clean-install schema artifacts are generated from the full plan.
+
+## Read surfaces and provenance
+
+`POST /api/v1/task-configuration/preview` uses the **read bearer**, accepts only
+workspace identity, profile and budget, and has no idempotency/mutation effect.
+`GET /api/v1/runs/{run_id}/task-configuration` reads the immutable saved snapshot.
+Both return `task_configuration.v1`, `capability_grant=false`, normalized
+requested/effective budgets, closed field sources, a safe project summary and
+fingerprints. A rejected preview has field-specific reasons and no partial
+project/effective fingerprint. Decoder details and repository bytes are never
+returned; exclusion paths and Skill suggestions are represented as counts.
+Malformed JSON, unknown/duplicate fields and unsupported bounds fail closed.
+
+Sources describe normalized effective values: product defaults, operator
+overrides or project restrictions. Historical Runs without a retained operator
+ceiling use `snapshot` provenance rather than inventing their input history.
+Their read projection preserves legal historical CLI ceilings, including zero
+or omitted tool limits and values above current product input bounds. The
+renderer validates these saved numbers independently of creation inputs and
+requires requested/effective snapshot values to match; explicit JSON null is
+never a zero/default value. Preview responses cannot claim snapshot provenance.
+The configuration fingerprint includes the safe projection and full project
+snapshot digest; later file edits do not alter it.
+
+The standalone React `TaskConfiguration` component edits draft request fields
+and shows the Go preview, or reads an existing Run snapshot without edits. Reads
+are aborted on task changes and late results cannot replace the current view.
+Connection identity also gates the first render so a replacement API client
+cannot briefly expose the previous connection's saved configuration.
+Draft inputs and validity belong to the workspace's parent draft state so closing
+the panel cannot silently revert an invalid attempted budget to defaults.
+Creation rechecks the file in Go; a preview is never admission authority.
+
+## Explicit limits and follow-up gap
+
+Budget and profile/read-only admission constraints actually take effect.
+`exclude_paths` currently has **no runtime access-filter consumer**; it is pinned
+metadata, not an enforced file-access restriction. Skill suggestions and typed
+test/format action IDs are also recorded metadata and never automatically
+installed, enabled or executed. The UI states these limits separately. Adding
+an exclusion consumer across every relevant read/write adapter requires a later
+scoped contract and regression matrix; this slice does not claim that work.
+
+This changes neither runtime permission selection nor leases, approvals,
+execution backends, secrets, command launch, model eligibility or account access.
+Tests separate actual Go creation/migration/replay/succession checks from React
+presentation fixtures. Native Desktop delivery and live model billing remain
+outside the evidence supplied by this slice.