From fe2b46103fd5d72459081b2040468db0a41120c4 Mon Sep 17 00:00:00 2001 From: huangruiteng <14976749+huangruiteng@users.noreply.github.com> Date: Sun, 20 Sep 2026 02:43:00 +0800 Subject: [PATCH 1/2] docs: propose live research team workspace and agent interaction design Signed-off-by: huangruiteng <14976749+huangruiteng@users.noreply.github.com> --- docs/architecture/rfcs/README.md | 8 + ...lligent-review-presentation-surfaces-v0.md | 12 + ...t-review-presentation-surfaces-v0.zh-CN.md | 9 + .../rfcs/live-team-workspace-v0.md | 374 ++++++++++++++++++ .../rfcs/live-team-workspace-v0.zh-CN.md | 302 ++++++++++++++ .../rfcs/loopx-overall-roadmap-v0.md | 15 +- .../rfcs/loopx-overall-roadmap-v0.zh-CN.md | 11 +- 7 files changed, 729 insertions(+), 2 deletions(-) create mode 100644 docs/architecture/rfcs/live-team-workspace-v0.md create mode 100644 docs/architecture/rfcs/live-team-workspace-v0.zh-CN.md diff --git a/docs/architecture/rfcs/README.md b/docs/architecture/rfcs/README.md index fcf58a98e1..9db0729498 100644 --- a/docs/architecture/rfcs/README.md +++ b/docs/architecture/rfcs/README.md @@ -287,6 +287,14 @@ changes. Lark operation cards. Direct Goal workspace navigation and Overview compose bounded graph/acceptance readback, source navigation and snapshot export. General cross-channel disclosure, living documents and governed amendment/settlement review remain open. +- [Live Team Workspace v0](live-team-workspace-v0.md) + ([中文版](live-team-workspace-v0.zh-CN.md)) + - **RFC status:** Draft presentation slice under intelligent presentation. + - **Delivery on `main`:** Design proposal only; no live team-stream qualification. + - **Current boundary:** A command surface plus spatial research studio, evidence + handoff motion, conclusion revision and replay. L1 binds existing execution + facts to packaged UI; L2 requires original-coordinator two-cycle acceptance; + L3 qualifies semantic zoom and display scale. No new scheduler or authority. - [Human Attention Wishlist v0](human-attention-wishlist-v0.md) ([中文版](human-attention-wishlist-v0.zh-CN.md)) - **RFC status:** Draft, under maintainer review. diff --git a/docs/architecture/rfcs/intelligent-review-presentation-surfaces-v0.md b/docs/architecture/rfcs/intelligent-review-presentation-surfaces-v0.md index 7e3d6b2d16..694236b92d 100644 --- a/docs/architecture/rfcs/intelligent-review-presentation-surfaces-v0.md +++ b/docs/architecture/rfcs/intelligent-review-presentation-surfaces-v0.md @@ -673,6 +673,18 @@ canonical facts, authority, evidence status, or whether an effect committed. Adaptive policies must be inspectable and resettable. Their outputs carry reason codes and preserve a deterministic fallback. +### 8.7 Live team workspace + +[Live Team Workspace v0](live-team-workspace-v0.md) specifies the S5/R2 research +journey: a command surface and spatial studio share typed facts with the list, +semantic zoom and historical replay. Purposeful motion is a product objective; +registration, execution, return, independent acceptance and requester adoption +remain distinct. Preserve source/revision lineage and uncertainty; never infer +activity from decorative motion or confidence from agreement counts. This is a +Draft presentation slice, not a shipped team stream or new scheduling authority. +Its L1–L3 plan and V1–V7 acceptance extend this RFC's Stage 3 presentation journey; +they do not close cross-channel or governed settlement acceptance here. + ## 9. Coverage across the long-horizon lifecycle | Phase | Intelligent surface responsibility | diff --git a/docs/architecture/rfcs/intelligent-review-presentation-surfaces-v0.zh-CN.md b/docs/architecture/rfcs/intelligent-review-presentation-surfaces-v0.zh-CN.md index da11790a72..30d56c5a1e 100644 --- a/docs/architecture/rfcs/intelligent-review-presentation-surfaces-v0.zh-CN.md +++ b/docs/architecture/rfcs/intelligent-review-presentation-surfaces-v0.zh-CN.md @@ -551,6 +551,15 @@ type PresentationArtifactPlan = { Adaptive policy 必须 inspectable、resettable，其输出携带 reason codes，并始终保留 deterministic fallback。 +### 8.7 团队实时工作区 + +[团队实时工作区 v0](live-team-workspace-v0.zh-CN.md) 描述 S5/R2 投研旅程： +指挥台与空间工作室同列表、语义缩放和历史回放共用 typed 事实。有目的的动态 +是产品目标；注册、执行、返回、独立验收与请求方采用保持分离。保留来源 / 版本 +谱系和不确定性，不能由装饰动态推断活动或由共识人数推断置信度。这是 Draft +展示切片，不是已交付的团队流，也不新增调度权限。其 L1–L3 计划与 V1–V7 验收 +展开本 RFC Stage 3 的展示旅程，不关闭本文跨渠道或受治理 settlement 验收。 + ## 9. 覆盖长程工作的完整生命周期 | 阶段 | 智能化界面的职责 | diff --git a/docs/architecture/rfcs/live-team-workspace-v0.md b/docs/architecture/rfcs/live-team-workspace-v0.md new file mode 100644 index 0000000000..e939d5dd84 --- /dev/null +++ b/docs/architecture/rfcs/live-team-workspace-v0.md @@ -0,0 +1,374 @@ +# RFC: Live Team Workspace v0 + +- **RFC status:** Draft; proposed product and presentation decisions. +- **Delivery maturity:** Design only. An isolated synthetic visual exploration does not qualify production execution or streaming. +- **Owners:** existing workspace presentation, collaboration and runtime owners. +- **Created / normative revision:** 2026-09-20. +- **Implementation baseline:** `e7ef75c08`. +- **Language:** [中文版](live-team-workspace-v0.zh-CN.md) is the semantic mirror. Differences are defects. +- **Parent contracts:** [intelligent presentation](intelligent-review-presentation-surfaces-v0.md), [overall roadmap](loopx-overall-roadmap-v0.md) S3/S5/S7 and R2/R3/R7, [project coordination](../../reference/project-coordination.md). + +Sections 1–11 propose the product and acceptance contract; Section 4 records +current implementation facts. Section 12 contains unresolved decisions. Appendix +A separates investigation from implementation. This document owns the live-team +presentation slice, not another scheduler, work graph or capability registry. + +## 1. Decision summary + +Make a working research team **visible, spatial and inspectable** in the existing +Goal workspace. Combine a precise command surface with an expressive research +studio: people can watch collaborators exchange artifacts, challenge findings, +revise conclusions and return to the original coordinator. Dramatic presentation +is an explicit product objective, alongside comprehensibility and correctness. + +Use one semantic projection for the spatial view, accessible list, timeline and +channel summary. Existing typed owners decide execution, acceptance, adoption +and effects. The renderer chooses how to explain those facts. Opening, replaying, +zooming or rearranging the view never starts work or changes authority. + +The first live slice is opt-in through the existing **Team execution** entry. +A later approved Overview composition can bring the scene forward without +removing Tasks/Kanban, Chat or Outputs. No default navigation or runtime changes +are approved by this RFC. A synthetic preview must visibly identify itself. + +## 2. Problem and observable result + +A roster plus status cards cannot answer: **what changed the team's conclusion, +which evidence is it using, who owes the next result, and does the original +research Agent continue?** Streaming more chat creates volume without causality. +A generic workflow editor describes topology but misses the evolution of a +research judgment. An animated office can suggest productive work when the only +fact is that a member was registered. + +The target journey starts in the original research conversation: the coordinator +splits a question, local and cloud members investigate, one challenges an input, +another revises it, an independent check accepts the exact artifact, and the +coordinator adopts it into a new research step. The workspace makes that chain +legible without opening every conversation. It also makes incomplete return, +lost observation and rejected results equally visible. + +Invariants: + +- Registration, grant, admission, execution, return, independent acceptance, + requester adoption and Goal completion are distinct facts, not one progress bar. +- A peer role, visual group, drag gesture or model statement grants no authority. +- An unknown cost, source relationship, dependency or state remains unknown. +- Evidence freshness and input revision travel with the displayed conclusion. +- Reading/replay cannot run validators or providers on every animation tick. +- The original coordinator remains substantive; more members are justified by + independent work or a dependency, not an on-screen Agent count. + +## 3. Scope and visual experience + +### One workspace, three scales + +| Scale | What the user sees | Useful interaction | +| --- | --- | --- | +| Team | Stable member stations around the current question; grouped subteams at larger scale | Focus an active dependency or blocked branch; open a member in place | +| Work | The selected question's dependency paths, owed returns and independent checks | Follow a handoff, inspect a repair, return to the original conversation | +| Evidence | Versioned artifacts and compact conclusion changes | Compare before/after, open an authorized source, inspect acceptance and adoption separately | + +Use semantic zoom: expanding one task replaces unrelated detail with a compact +context, preserving selection and a return path. Camera changes are intentional +and interruptible; a new event cannot move the user's reading position. Mobile +uses a vertical focused chain and accessible list instead of shrinking a graph. + +### Signature interactions + +1. **Evidence travels.** A bounded packet moves along an actual handoff edge once + when a delivery is observed. Request, returned artifact and acceptance have + different glyphs. Continuous decorative particles are not a throughput meter. +2. **Disagreement leaves a trace.** A sourced challenge opens a visible return + path; the superseded conclusion remains as a before/after layer. Agreement + count never becomes confidence. Independent origin requires explicit lineage. +3. **The conclusion takes shape.** A central research folio accumulates adopted, + versioned findings. Compare the previous basis with the revised basis; do not + animate a fictional probability or forecast from token output. +4. **Replay the turning point.** A timeline seeks to material events and shows + what was known then, why it changed and what remains unresolved. Replay is + labeled historical, with an explicit return to current observations. +5. **Focus the critical dependency.** Dim unrelated work without hiding failures. + Highlight a real prerequisite path or owed return. If the work owner provides + no dependency relation, show an observed exchange, not an invented causal edge. +6. **Grow from a studio to a portfolio.** At 10/30/100, collapse by explicit + subgoal/team scope, preserve material exceptions and expand one cluster. A + hundred actors are not a hundred simultaneous text panels or subscriptions. + +The preferred art direction is clean and high-tech: a cold dark field, precise +linework, a clear central research focus, stable member zones and event-local +highlights. Avoid overlapping decorative orbits, translucent text panels and +multiple simultaneous focal points. Route edges outside reading regions; keep +only the relevant exchange prominent. Depth must clarify hierarchy. + +The visual vocabulary can be striking: architectural depth, dimensional work +surfaces, purposeful choreography and transitions from member to artifact to +conclusion. Retain LoopX typography, semantic colors and readable DOM text. +Use [Earn The User's Attention](../../development/design.md#earn-the-users-attention) +for the complete viewport. Motion itself may earn attention, but must not +misrepresent activity. No raw reasoning stream, invented inner monologue, +unreviewed public showcase, trade execution or automated investment decision is +part of this slice. + +## 4. Audited current system + +| Current owner / entry | Verified behavior at baseline | Gap for this proposal | +| --- | --- | --- | +| `collaboration_mcp.Delegations`; `control_plane/collaboration/delegation.ts` | Requester-scoped explicit bindings, detached work, original-operation recovery, independent accepted-artifact readback | No complete chronological team event source or adoption-to-continuation proof | +| `delegation_inventory.py`; CLI `delegation operations/inspect` | Bounded hash-address pagination; actual Turn/preflight; unavailable branches retained | Inventory is a live page, not an ordered change stream or fleet snapshot; reads can invoke pinned validation | +| `goal-team-work.tsx` in the existing workspace | On-demand inventory and per-binding prerequisite checks | No spatial scene, replay or continuous team observation | +| `chat_server._turn_events`; frontend `streamChatTurn` | Turn-scoped SSE with event identity, sequence and reconnect cursor | Not a Goal/team aggregate; its cursor cannot order independent Turn journals | +| Peer requests/returns and Goal Chat continuation | Request lineage, receiver decisions, original routes; native pause/recovery | Delivery is not requester adoption; unattended wake and two-cycle G1 remain unqualified | +| Existing task graph and acceptance projections | Typed work dependencies and independent acceptance observations | Research conclusion/source lineage must come from the owning capability, not inferred chat text | + +Provider credential/admission defects remain runtime-owner work. The scene +must expose such failures and cannot fix them by inferring readiness or launching +an alternative provider. Existing integration repairs and this presentation +slice can progress independently until live qualification requires both. + +## 5. Architecture and ownership + +**Placement:** extend the existing presentation contract and workspace. Reuse +collaboration, Turn, Todo/acceptance, manager context and source/evidence owners. +There is no new built-in capability or provider. The built-in local Chat host +provides observation transport; optional model/runtime providers retain their +adapters. Finance-specific metrics stay in the finance/research capability and +arrive as audience-safe artifact/claim projections, never kernel rules. + +```mermaid +flowchart LR + W[Existing work / Turn / peer / acceptance owners] --> P[Typed audience-scoped presentation projection] + E[Capability-owned evidence and revisions] --> P + P --> S[Snapshot plus bounded change transport] + S --> V[Spatial scene / DOM list / timeline] + P --> C[CLI and authorized channel summaries] + V --> A[Existing action preview / execution owner] +``` + +### Observation contract, before a new schema + +Implement one cohesive TS projection when its first live caller is built. +Python may read journals, collect owner facts and serve transport; it must not +invent the semantic lifecycle. First inventory producer fields and receipt +identities. Reuse existing schemas; do not add speculative parallel event names +or a second writable event-sourced task system merely to render an animation. + +The projection needs: scoped Goal and audience; stable actor/work/operation +identity; parent request and explicit dependency references where present; +source owner and revision; event identity and observed time; previous/current +visible fact; freshness/completeness; audience-safe artifact reference and +version; separate validation, delivery and adoption observations. Unknown is +explicit; observation time must not be presented as actual start/finish time. + +A process lock is evidence of an active delegation worker, not proof that its +model is currently generating tokens. A returned model claim is distinct from +host-observed execution and independent validation. UI labels preserve that +provenance. Cost displays retain measured/estimated/unavailable distinctions and +provider scope; motion rate or token count does not imply financial cost. + +### Snapshot and stream + +- A versioned snapshot establishes one audience-scoped display basis. Subscribe + with a bounded cursor; if the basis expired or a gap is detected, resnapshot + explicitly. The existing hash inventory cursor is not a chronological cursor. +- A projection sequence may order display delivery within one stream epoch; + source revisions and request/dependency identities establish causality. Do not + impose wall-clock total order across hosts or reuse Turn cursors globally. +- Deduplicate by stable event identity; replay changes presentation only. Duplicate + delivery, reconnect and seeking cannot repeat side effects or sound/motion. +- Use bounded buffers and coalesce routine activity. Preserve acceptance, + invalidation, failure and decisions; overflow becomes a visible gap plus + resnapshot rather than silent loss or unbounded memory. +- Separate cheap observations from explicit accepted-artifact revalidation. + Show last-verified revision/time and invalidation. Reconnection must not call + every validator repeatedly; a correctness-critical fresh read uses the existing + owner and its admission limits. A historical acceptance badge is never labeled + currently valid when inputs or availability have changed. +- One scoped team connection replaces per-actor browser connections. Reuse the + local Chat SSE infrastructure where compatible; provider-specific streaming + can be absent. Without granular events, show last observed state honestly. + +## 6. Research and technical choices + +These primary sources inform the design; they do not qualify LoopX behavior. + +### Agent-interface study + +The visual study prioritizes interfaces **for operating and understanding +agents**, rather than unrelated interfaces built by agents. Creator posts and +sampled demo frames were inspected on X. These are presentation observations +and creator-described capabilities, not independent product or runtime tests. +No popularity, throughput or productivity claim is used as evidence. + +| Agent example / creator source | Observed or described interaction | Transfer to LoopX / limit | +| --- | --- | --- | +| [AgentCraft multiplayer](https://x.com/idosal1/status/2033962188025565595) | RTS map, member roster, selected-agent context; creator describes cross-machine handoff | Stable spatial identity and selection-to-context; adopt the interaction grammar without fantasy-game chrome or implying that proximity grants authority | +| [MagicPath 2.0](https://x.com/skirano/status/2054975534539370708) | Creator presents a shared human/agent canvas and parallel design outputs; demo shows editable product artifacts | Make research artifacts the common workspace: compare contributions beside their sources, rather than filling the scene with conversation windows | +| [Copilot Mission Control](https://x.com/DanWahlin/status/2059401567892377800) | Selected session, tool-oriented scene and activity panel; author describes replay and turns | Couple a legible event stream to the scene, with selection and historical inspection; tools are distinct from team members and successful calls do not certify outcomes | +| [ClawTeam Gource visualization](https://x.com/huang_chao4969/status/2036860913685561421) | Agent/Git activity visualized as an evolving graph alongside execution output | Event-driven spatial changes can make collaboration tangible; file/commit activity alone cannot stand in for peer adoption or research progress | +| [Mission Control](https://x.com/nykdotdev/status/2091755907676082211) | Creator walkthrough presents dispatch, review, runtime and operational views | Preserve explicit review, failure and next-action context under the expressive scene; do not copy a full administration console into the first viewport | +| [Star Office UI](https://x.com/ring_hyacinth/status/2028021181073527273) | Pixel workspace with state-related locations and a compact work/status area | Stable places can convey team presence; pixel-office decoration and synthetic inner monologues are not the chosen research interface | + +**Design synthesis:** a spatial research workbench with a precise command +surface. Keep an enduring place for each member, but put the selected research +question and its artifacts at the visual center. This is a proposed synthesis, +not a claim of unprecedented invention. General WebGL showcases may inform +craft, but do not determine the agent interaction model. + +The next visual iteration must demonstrate three connected moments: + +1. A member returns a versioned evidence object into the shared workbench. + One bounded transfer reveals its sender, receiver and delivery observation; + selection opens the artifact in a stable, readable inspector. +2. An independent member challenges that object. A focused spatial separation + reveals the exact old/new inputs and unresolved discrepancy; unaffected + members stay compact. Spatial depth represents explicit versions or groups, + never hidden reasoning, confidence or fabricated progress. +3. The original researcher adopts the verified revision and changes the + visible conclusion. A time scrubber reconstructs that turning point without + re-running work, then returns to current observations. + +The command surface stays stationary while objects change. Use crisp material +edges, restrained illumination and deliberate transitions for a high-tech feel; +remove redundant orbit lines, repeated status cards and simultaneous animation +loops. Compare the scene with its plain list on comprehension of who owes what, +what changed and what remains unverified. A stronger visual treatment still +requires a new concrete preview; the current exploration is not approved UI. + +### Runtime and rendering references + +| Source | Useful precedent | LoopX choice / tradeoff | +| --- | --- | --- | +| [AutoGen Studio](https://microsoft.github.io/autogen/0.4.7/user-guide/autogenstudio-user-guide/index.html) | Visual composition and interaction with teams | Runtime evidence and evolving conclusions take priority over a workflow editor; do not require users to wire a DAG | +| [LangSmith Studio](https://docs.langchain.com/langsmith/studio) | Graph inspection, chat and time-travel debugging | Adapt replay to owner questions; historical inspection never implicitly reruns work | +| [Anthropic research system](https://www.anthropic.com/engineering/multi-agent-research-system) | Lead/worker research division and parallel investigation | Require original-lead synthesis and useful independent artifacts; no performance uplift transferred from another system | +| [React Flow animated edges](https://reactflow.dev/examples/edges/animating-edges) and [performance](https://reactflow.dev/learn/advanced-use/performance) | Custom nodes/edges, SVG/Web Animations, selective subscriptions | Candidate if pan/zoom/graph selection merits its dependency; start small with DOM/SVG and stable layout, not a continuously moving force graph | +| [PixiJS performance](https://pixijs.com/8.x/guides/concepts/performance-tips) and [accessibility](https://pixijs.com/8.x/guides/components/accessibility) | Batched spatial rendering; explicit accessible DOM overlays | Candidate for measured large-scene bottlenecks; keep text/actions semantic, and compare with SVG on identical workloads before adding it | +| [Rive data binding](https://www.rive.app/blog/data-binding-in-rive-a-shared-language-for-designers-and-developers) | State-driven authored animation | Optional later expressiveness; runtime facts drive animation inputs, never the reverse; assets/runtime add maintenance cost | +| [AG-UI events](https://docs.ag-ui.com/concepts/events) | Streaming lifecycle, state snapshots/deltas and subagent attribution | Optional transport vocabulary mapping after an adapter need; run-finished does not mean LoopX acceptance/adoption, and raw/reasoning payloads are excluded | +| [SSE](https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events) | Event ids and reconnectable server-to-client delivery | Fit for scoped observations; commands retain current HTTP owners; transport reconnection alone cannot prove complete history | + +Prefer the existing React/DOM stack plus SVG/WAAPI for the first live view. +Defer WebGL/WebGPU/3D dependency until an actual depth interaction or measured +workload needs it. Visual ambition is judged by the experience, not shader count. +A renderer change must preserve the same projection and accessibility behavior. + +## 7. Authority, privacy and compatibility + +Selecting a member opens its existing authorized context; a role or ancestor +relationship grants no child access. Resolve audience authorization server-side +before joining or caching facts. Revoke and clear previously visible data when +scope changes; reconnect cannot reuse another audience's cursor. Do not embed +raw prompts, reasoning, credentials, local paths or private provider identifiers +in events, shareable previews or public screenshots. Explicit owner-only source +access remains distinct from public-safe summaries and exports. + +Default-off preserves ordinary chat, bindings, scheduler behavior and budgets. +Existing threads are not replaced to attach this display. No passive subscription +starts, retries, accepts, cancels or wakes work. Actual controls route through +existing typed actions and current admission; viewing a replay cannot activate +historical action buttons. Queue, inbox and steer retain their different delivery +contracts. Keep unknown/unsupported provider facts visible. + +## 8. Migration and rollback + +Ship the first scene behind an explicit view selection using the existing +team-inspection entry. Keep the list available over the same facts. No canonical +record migration is needed; historical journals lacking fields remain incomplete. +Any disposable transport cache must be rebuildable from current authorized +owners, with explicit history gaps if durable events do not exist. + +Disable the scene or stop its subscription to return to list/chat without +interrupting member work. Do not delete journals, rebind sessions, rewrite +historical receipt strength or reset provider state. Default Overview/navigation +changes require the repository's concrete first-screen preview approval before +commit/push; approval of an unrelated copy edit does not approve this scene. + +## 9. Acceptance + +| ID | Required evidence | Passing result / exclusion | +| --- | --- | --- | +| V1 Truth | Synthetic lifecycle, stale artifact, missing adoption and lost connection | All axes remain distinct; no fabricated executing/accepted/complete state | +| V2 Stream | Duplicate, out-of-order source revisions, expired cursor, restart, revocation and overflow | Consistent current readback, visible gaps, no duplicate effects, no audience leakage | +| V3 Experience | Packaged desktop + 390px, keyboard/screen reader, 200% zoom, reduced motion, idle/blocked/unavailable | Usable source/action path and return focus; selection preserved; no idle activity theater | +| V4 Research loop | Original coordinator plus 2–3 real workers; two cycles; one peer dependency, rejection/correction and worker interruption while another progresses | Pinned independent validation, requester adoption and original-lead continuation read back via CLI and packaged UI; no manual result forwarding; R2/G1 rules unchanged | +| V5 Scale | Identical recorded workloads at 10/30/100 actors, high event rate and long replay | Freeze workload/device/targets before measurement; report latency, CPU/GPU, memory, gaps and attention cost; synthetic render load is not live fleet qualification | +| V6 Rollback | Feature off, old thread, pause, renderer error, source loss | Existing work continues under its original owner; list/Chat remain available; no validation storm | +| V7 Comprehension | Observers identify the blocker, changed conclusion, evidence origin and next owner from list versus scene | Predeclare questions and success criteria; record errors and time-to-answer; novelty ratings cannot conceal worse comprehension | + +Candidate V5 targets, to ratify before execution: 60fps target/30fps graceful +floor for foreground motion; p95 interaction below 100ms; p95 projection-receipt +to DOM below 500ms; bounded 10-minute replay and no unbounded heap growth in a +30-minute run. Separate renderer latency from provider/network latency. Record +hardware, viewport, event rate, visible nodes/edges and measurement method. These +are design budgets, not measured performance or release guarantees. + +## 10. Operating behavior + +Quiet, disconnected, stale, recovering and blocked are first-class visible +states. Stop packet animation when disconnected; retain last observations with +their age. Stop animation in hidden tabs; reduced motion removes travel/camera +motion while retaining discrete events. Offer an explicit pause-animation control +separate from pause-coordinator. Do not auto-scroll logs or emit per-token live +region announcements. Unknown cost is not zero. No audio by default. + +Aggregate only over the declared scope and completeness. Show partial inventory +instead of claiming an entire fleet is healthy. Failure stays discoverable while +its cluster is collapsed. Use existing scheduler backpressure; the display has +no wake policy or model refresh loop. Record observation errors without storing +sensitive event bodies in diagnostics. + +## 11. Delivery order and relationship to aggressive R2 progress + +| Slice | Complete useful result | Entry / exit | Owner and rollback | +| --- | --- | --- | --- | +| L0 Design and traceability | Interactive synthetic study, source audit and executable acceptance plan | Design review; V1/V3 concept checks; no live claim | Presentation; discard prototype, retain decisions | +| L1 Live small-team scene | Existing Team execution opens current scoped work, follows returns and exposes exact evidence; list and scene agree | Existing admitted bindings; V1/V2/V3/V6 with real entrypoint | Collaboration projection + Chat/packaged UI; disable scene | +| L2 Continuous research | Original research Agent completes V4 and the user replays correction→adoption→next cycle | Runtime credential/admission repairs and usable L1; qualify each host profile, then V4 | Existing R2/R3 owners; stop new admission and retain results | +| L3 Semantic zoom and scale | Subteams, source/conclusion version comparison and focused replay at measured 10/30/100 display load | Typed lineage producers and L2; V5/V7 and approved first-screen composition | Presentation + evidence owner; reduce detail/revert renderer | + +The next executable implementation is **L1**, a complete observation-to-user +vertical, not an event schema alone. Include source collection, typed projection, +reconnect semantics, packaged UI, independent CLI readback and failure cases in +one delivery plan. Inventory existing receipt identities before choosing a new +transport schema. Keep expensive validation outside routine polling. + +R2 remains the execution priority: resolve runtime admission/authentication in its +current owner; prove two useful cycles; then qualify governed task derivation and +approved member/profile provisioning through existing registration and settings. +Parallel return/join and original-lead continuation are required before raising +concurrency. Persistent static bindings are a bounded starting point, not a +reusable dynamic staffing system. Do not create a second orchestration program, +replace ongoing integration work, weaken gates, or hide execution failures behind +visual completion. Lark consumes the same bounded facts and existing actions; +its complete return journey is separately qualified and is not a 3D UI port. + +## 12. Open decisions + +- **Presentation owner, before L1:** final art direction and motion grammar. + Recommendation: a command surface plus spatial research studio, with semantic + zoom and conclusion revision as the signature. Test populated and failed states. +- **Collaboration/transport owners, before L1:** reuse or extend existing event + receipts for a restartable team stream. Audit durable coverage; choose a scoped + cache only if the current owners cannot supply the needed projection directly. +- **Presentation/performance owners, before L3:** retain SVG or add PixiJS/Rive. + Compare equal scenes, accessibility, bundle size and memory; no dependency is + adopted by this RFC. +- **Research capability owner, before conclusion view goes live:** exact public-safe + claim/source/version/adoption fields and privacy. The kernel must not classify + report independence from prose or infer financial recommendation confidence. + +## Appendix A. Investigation and delivery evidence + +At the named baseline, source inspection confirms the existing inventory, +preflight, team dialog and Turn SSE described in Section 4. Primary-source +research in Section 6 establishes design alternatives, not product acceptance. +A local synthetic interactive study exercises spatial members, one-shot handoff +motion, correction replay, evidence drill-down and degraded-state presentation. +It launches no Agents and reads no private research data. It is not shipped in +the product or used as live runtime evidence. + +**Current checkpoint:** L0 design proposal; L1–L3 unimplemented here. V1/V3 may be +explored with the synthetic study; V2/V4/V5/V6/V7 remain unqualified until their +required production or measured evidence exists. No G1/G3/G4 promotion follows. +The exact delivery PR carries validation and review; roadmap pointers retain +this boundary rather than appending another operational task ledger. diff --git a/docs/architecture/rfcs/live-team-workspace-v0.zh-CN.md b/docs/architecture/rfcs/live-team-workspace-v0.zh-CN.md new file mode 100644 index 0000000000..464c11b23c --- /dev/null +++ b/docs/architecture/rfcs/live-team-workspace-v0.zh-CN.md @@ -0,0 +1,302 @@ +# RFC：团队实时工作区 v0 + +- **RFC 状态：** Draft；产品与展示决策提案。 +- **交付成熟度：** 设计阶段。隔离的合成视觉探索不证明生产执行或流式链路已通过验收。 +- **Owner：** 既有工作区展示、协作与 runtime owner。 +- **创建 / 规范修订：** 2026-09-20。 +- **实现基线：** `e7ef75c08`。 +- **语言：** [English](live-team-workspace-v0.md) 为语义镜像；差异属于缺陷。 +- **上位合同：** [智能展示](intelligent-review-presentation-surfaces-v0.zh-CN.md)、[总路线](loopx-overall-roadmap-v0.zh-CN.md) S3/S5/S7 与 R2/R3/R7、[项目协调](../../reference/project-coordination.md)。 + +第 1–11 节提出产品与验收合同；第 4 节记录当前实现事实。第 12 节保留未决事项， +附录 A 区分调研与实现。本文只拥有团队实时展示切片，不建立第二套调度器、 +工作图或 capability registry。 + +## 1. 决策摘要 + +让工作的投研团队在既有 Goal 工作区里**看得见、有空间感、可深入核对**。 +将精确的指挥台与富有表现力的研究工作室融合：用户能看见成员交换产物、 +质疑依据、修订结论，并把结果交回原协调员。惊艳的展示本身是明确产品目标， +同时要求可理解、事实准确。 + +空间场景、无障碍列表、时间轴和渠道摘要消费同一语义投影。既有 typed owner +判断执行、验收、采用和副作用；renderer 决定怎样解释事实。打开、回放、缩放、 +移动视图都不能启动工作或改变权限。 + +首个真实切片由既有**团队执行情况**入口显式选择启用。后续经批准的概览编排 +可将现场前置，保留任务 / Kanban、对话和成果。本文不批准默认导航或 runtime +变化。合成预览必须清楚标识演示性质。 + +## 2. 问题与可观察结果 + +成员名单与状态卡片不能回答：**什么改变了团队的结论、正在采用哪份证据、 +谁欠下一份结果、原投研 Agent 是否仍在继续？** 更多流式聊天只是增加信息量， +未必说明因果。通用工作流编辑器描述拓扑，却不表达研究判断的变化；动画办公室 +可能在只有成员注册事实时，制造正在产出价值的错觉。 + +目标旅程从原投研对话开始：主力拆解问题，本地与云端成员研究，一方质疑输入， +另一方修订，独立校验接受准确产物，主力采用它并进入下一步研究。工作区让用户 +不用翻遍全部对话就能理解这条链路，同时看清返回未闭合、观测丢失与结果被拒绝。 + +必须保持的不变量： + +- 注册、授权、准入、执行、结果返回、独立验收、请求方采用、Goal 完成分别记录， + 不能压成一条进度条。 +- peer 角色、空间分组、拖拽或模型声明不授予权限。 +- 未知成本、来源关系、依赖或状态仍显示未知。 +- 证据新鲜度与输入版本随结论展示。 +- 阅读 / 回放不能在每一动画帧运行 validator 或 provider。 +- 原协调员持续承担实质研究；更多成员必须对应独立工作或真实依赖，不能以画面中的数量为目标。 + +## 3. 范围与视觉体验 + +### 一个工作区，三个尺度 + +| 尺度 | 用户看见什么 | 有用的交互 | +| --- | --- | --- | +| 团队 | 当前问题周围的稳定成员工位；规模变大后按明确小队分组 | 聚焦活跃依赖或受阻分支，原地打开成员 | +| 工作 | 所选问题的依赖链、待返回结果与独立校验 | 跟随交接、检查修复、回到原对话 | +| 证据 | 版本化产物与精简的结论变化 | 对比前后，打开获准来源，分别查看验收和采用 | + +用语义缩放展开一个任务，同时把无关细节收为紧凑上下文，保留选择与返回路径。 +镜头变化由明确交互触发且可打断；新事件不能移动用户的阅读位置。手机采用纵向 +聚焦链与无障碍列表，不把大图缩小到无法阅读。 + +### 标志性交互 + +1. **证据流动。** 观察到一次实际交付后，有限的证据载体沿交接边移动一次。 + 请求、返回产物与验收采用不同图形。持续装饰粒子不能充当吞吐率仪表。 +2. **分歧留痕。** 带来源的质疑展开可见返回路径；被替代的结论保留为前后对比层。 + 共识人数不换算成置信度，来源独立性需要显式谱系。 +3. **结论成形。** 中央研究页汇集已采用、带版本的发现，比较旧依据与新依据； + 不依据 token 输出制作虚构概率或预测动画。 +4. **回放转折。** 时间轴定位实质事件，说明当时知道什么、为何改变、还有哪些问题。 + 回放明确标为历史，并提供回到当前观察的入口。 +5. **聚焦关键依赖。** 淡化无关工作但不藏起失败；突出真实前置条件或待返回结果。 + work owner 没有提供依赖关系时，只画观察到的交换，不创造因果边。 +6. **从工作室长成团队群。** 在 10/30/100 规模，按显式子目标 / 团队范围折叠， + 保留重要异常，展开一个群组。100 个成员不等于同时打开 100 份文本面板或订阅。 + +首选美术方向是干净的高科技感：冷黑底、精细线条、清晰的中心研究焦点、稳定 +成员区域和局部事件高亮。避免多层装饰轨道、透明文字卡片与多个焦点同时竞争。 +连线绕开阅读区，只突出相关交换；纵深必须帮助理解层级。 + +视觉可以大胆：建筑式纵深、有体积感的工作面、有目的的编排、从成员到产物再到 +结论的连续转场。沿用 LoopX 字体、语义颜色与可读 DOM 文字。整个视口遵循 +[Earn The User's Attention](../../development/design.md#earn-the-users-attention)。 +动态本身可以值得观看，但不能伪装活动。原始推理流、虚构内心独白、未经审阅的 +公开 showcase、交易执行与自动投资决策不属于此切片。 + +## 4. 当前系统审计 + +| 当前 owner / 入口 | 基线已验证行为 | 本提案尚缺 | +| --- | --- | --- | +| `collaboration_mcp.Delegations`；`control_plane/collaboration/delegation.ts` | requester scoped 显式绑定、独立执行、原 operation 恢复、accepted 产物独立回读 | 完整有序团队事件源；采用到继续研究的证据 | +| `delegation_inventory.py`；CLI `delegation operations/inspect` | 有界 hash 地址分页、真实 Turn/preflight、保留不可用分支 | inventory 是实时分页，不是有序变化流或 fleet 快照；读取可能调用 pinned validator | +| 工作区 `goal-team-work.tsx` | 按需 inventory 与成员启动条件检查 | 空间场景、回放、持续团队观测 | +| `chat_server._turn_events`；前端 `streamChatTurn` | Turn scoped SSE、事件身份 / sequence / 重连 cursor | 不是 Goal/team 聚合流；该 cursor 不能给多个 Turn journal 排序 | +| peer 请求 / 返回与 Goal Chat continuation | 请求谱系、接收方决定、原路返回、native 暂停 / 恢复 | 投递不等于请求方采用；无人值守 wake 与两轮 G1 仍未通过验收 | +| 既有任务图与验收投影 | typed 工作依赖、独立验收观察 | 研究结论与来源谱系应由 capability 提供，不能从聊天文字推断 | + +provider credential / admission 缺陷仍由 runtime owner 处理。场景须揭示这些失败， +不能靠猜 readiness 或启动另一个 provider 消除它们。既有接入修复与展示切片可 +独立推进，到真实资格化时再汇合。 + +## 5. 架构与归属 + +**放置理由：** 扩展既有 presentation contract 和 workspace；复用 collaboration、 +Turn、Todo/acceptance、manager context 与来源 / 证据 owner。不新增 built-in +capability 或 provider。内置本地 Chat host 承载观测传输，可选 model/runtime +provider 保留各自 adapter。金融指标留在 finance/research capability，作为有 +受众边界的 artifact/claim 投影输入，不能变成 kernel 规则。 + +```mermaid +flowchart LR + W[既有 work / Turn / peer / acceptance owner] --> P[Typed 受众范围展示投影] + E[Capability 证据与版本] --> P + P --> S[快照与有界变化传输] + S --> V[空间场景 / DOM 列表 / 时间轴] + P --> C[CLI 与已授权渠道摘要] + V --> A[既有 action preview / 执行 owner] +``` + +### 先明确观测合同，再定 schema + +首个真实调用方落地时，实现一个内聚 TS 投影。Python 可读 journal、收集 owner +事实和承载传输，不能自行创造语义生命周期。先盘点 producer 字段与 receipt +身份，复用既有 schema；不能为了动画添加假想事件大全，或另一套可写的 +事件溯源任务系统。 + +投影需要：Goal 与受众范围；稳定 actor/work/operation 身份；存在时的父请求与 +显式依赖；源 owner 与版本；事件身份和观测时间；前后可见事实；新鲜度与 +完整性；受众安全的 artifact 引用及版本；分离的校验、投递与采用观察。 +未知显式表达，观测时间不能冒充实际开始 / 完成时间。 + +进程锁证明 delegation worker 活跃，不证明模型正在生成 token。模型返回的 +声明与宿主观察到的执行、独立验证分别展示。成本区分实测 / 估计 / 不可用及 +provider 范围；动画速度与 token 数不能隐含货币成本。 + +### 快照与流 + +- 带版本快照确定受众范围内的展示基线，以有界 cursor 订阅。基线过期或发现缺口 + 时显式重读快照；已有 hash inventory cursor 不是时间序 cursor。 +- projection sequence 可以给一个流 epoch 内的投递排序；因果由源版本和 + 请求 / 依赖身份确定。不能按不同 host 的墙钟强排全序，也不能全局复用 Turn cursor。 +- 按稳定事件身份去重；重放只改变展示。重复投递、重连和拖动时间轴不能重做 + 副作用，也不能反复触发同一声效 / 动画。 +- 缓冲有界、合并常规活动，保留验收、失效、失败与决策；溢出显示缺口并重读， + 不能静默丢失或无限增长内存。 +- 将廉价观测与显式 accepted-artifact revalidation 分开，显示上次校验版本 / 时间 + 与失效。重连不能重复调用所有 validator；必须新鲜读取的正确性检查仍走原 owner + 及其准入限制。输入或可用性变化后，历史验收不能标为当前有效。 +- 一个有范围的团队连接替代逐成员浏览器连接。兼容时复用本地 Chat SSE；provider + 可以没有细粒度流，此时如实展示上次观察状态。 + +## 6. 调研与技术选择 + +以下一手资料支撑方案比较，不证明 LoopX 行为通过资格化。 + +### Agent 界面专项调研 + +优先研究**操作和理解 Agent 的界面**，而非 Agent 生成的无关网站。通过 X 阅读 +创作者原帖并抽样查看演示视频画面；以下区分呈现观察与作者描述，不代表独立 +试用或 runtime 验证。热度、吞吐量和效率宣传不作为资格化证据。 + +| Agent 案例 / 创作者来源 | 观察到或作者描述的交互 | LoopX 借鉴与边界 | +| --- | --- | --- | +| [AgentCraft 多人协作](https://x.com/idosal1/status/2033962188025565595) | RTS 地图、成员列表、所选 Agent 上下文；作者描述跨机器交接 | 稳定空间身份与选中后查看上下文；借鉴操作方式，不复制奇幻游戏装饰，也不让空间邻近暗示权限 | +| [MagicPath 2.0](https://x.com/skirano/status/2054975534539370708) | 作者介绍人和 Agent 的共享画布与并行设计产物；视频展示可编辑产品成果 | 让研究产物成为共同工作面，贡献与来源可并排比较，避免把场景塞满聊天窗口 | +| [Copilot Mission Control](https://x.com/DanWahlin/status/2059401567892377800) | 所选 session、工具场景与活动面板；作者描述回放与 turns | 可读事件流与场景联动，支持选择和历史检查；工具与团队成员分开，调用成功不证明结果通过验收 | +| [ClawTeam Gource 可视化](https://x.com/huang_chao4969/status/2036860913685561421) | Agent / Git 活动图随执行演进，旁边保留执行输出 | 用真实事件推动空间变化，体现协作现场感；文件 / commit 活动不能替代 peer 采用或研究进展 | +| [Mission Control](https://x.com/nykdotdev/status/2091755907676082211) | 作者演示分派、review、runtime 与运维视图 | 表现场景下保留明确复核、失败和下一步上下文，不把完整管理后台复制到首屏 | +| [Star Office UI](https://x.com/ring_hyacinth/status/2028021181073527273) | 像素工作区、与状态关联的位置、精简工作 / 状态区域 | 固定位置帮助感知成员存在；像素办公室装饰与虚构内心独白不是本研究界面的方向 | + +**设计综合：空间研究工作台 + 精确指挥界面。** 每个成员有稳定位置，视觉中心 +交给所选研究问题和产物。这是拟议的综合方案，不声称首创。通用 WebGL 作品 +可以辅助打磨表现，但不能决定 Agent 交互模型。 + +下一轮视觉迭代必须连贯演示三个时刻： + +1. 成员把带版本的证据对象交到共享工作台。一次有限交接显示发送者、接收者和 + 投递观察，选中后在稳定、可读的详情区展开产物。 +2. 独立成员质疑该对象。局部空间展开揭示准确的新旧输入与未解决差异；无关 + 成员保持紧凑。纵深表达明确的版本或分组，不能表示隐藏推理、置信度或虚构进度。 +3. 原研究员采用已核验修订，更新可见结论。时间轴还原这个转折，不重新执行 + 工作，并能回到当前观察。 + +指挥界面固定，产物随事实变化。用精确材质边缘、克制照明和有意图的转场呈现 +高科技感；移除冗余轨道、重复状态卡片与同时循环的动画。与普通列表比较用户 +是否能理解谁欠什么、什么变了、什么仍未核验。更强视觉方案仍需新的具体预览， +当前探索不是已批准的产品 UI。 + +### Runtime 与渲染参考 + +| 来源 | 可借鉴之处 | LoopX 选择 / 取舍 | +| --- | --- | --- | +| [AutoGen Studio](https://microsoft.github.io/autogen/0.4.7/user-guide/autogenstudio-user-guide/index.html) | 可视团队组合与交互 | 优先 runtime 证据和演进中的结论，不要求用户连 DAG | +| [LangSmith Studio](https://docs.langchain.com/langsmith/studio) | 图检查、聊天与时间回溯调试 | 回放围绕所有者问题，历史检查不隐式重跑工作 | +| [Anthropic research system](https://www.anthropic.com/engineering/multi-agent-research-system) | lead/worker 研究分工与并行调查 | 要求原主力综合与有用的独立产物，不移植别的系统的效果声明 | +| [React Flow 动态边](https://reactflow.dev/examples/edges/animating-edges)与[性能](https://reactflow.dev/learn/advanced-use/performance) | 自定义节点 / 边、SVG/Web Animations、局部订阅 | 平移 / 缩放 / 图选择值得引入依赖时采用；首版可用 DOM/SVG 稳定布局，不让力导向图持续漂移 | +| [PixiJS 性能](https://pixijs.com/8.x/guides/concepts/performance-tips)与[无障碍](https://pixijs.com/8.x/guides/components/accessibility) | 批量空间绘制、显式 DOM 无障碍覆盖 | 大场景实测瓶颈后的候选；文字 / 动作用语义 DOM，以相同负载和 SVG 比较后再引入 | +| [Rive data binding](https://www.rive.app/blog/data-binding-in-rive-a-shared-language-for-designers-and-developers) | 状态驱动的设计动画 | 后续表现力候选；runtime 事实驱动动画，反向不成立；资产 / runtime 有维护成本 | +| [AG-UI events](https://docs.ag-ui.com/concepts/events) | 流式 lifecycle、状态快照 / delta、subagent 归属 | 有真实 adapter 需求后再映射传输词汇；run-finished 不等于 LoopX 验收 / 采用，排除 raw/reasoning payload | +| [SSE](https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events) | 事件 id 与可重连的服务端推送 | 适合范围内观测；命令沿用 HTTP owner；传输恢复不能证明历史完整 | + +首个真实视图优先既有 React/DOM + SVG/WAAPI。等真实纵深交互或实测负载需要时 +再引入 WebGL/WebGPU/3D。视觉野心按体验评判，不按 shader 数量。更换 renderer +必须保持同一投影与无障碍语义。 + +## 7. 权限、隐私与兼容 + +选择成员只打开既有获准上下文；角色或父子关系不自动授予子范围访问。服务端先 +解析受众权限再 join / cache 事实；scope 变化后撤回并清除已展示数据，重连不能 +复用其他受众 cursor。事件、分享预览与公开截图不能带 raw prompt、推理、凭证、 +本地路径或私有 provider 身份。明确的 owner-only 来源访问与公开安全摘要 / 导出 +保持区分。 + +默认关闭时保持普通 Chat、绑定、scheduler 与预算行为。不能为展示替换旧线程。 +被动订阅不启动、重试、验收、取消或唤醒工作。真实控制走既有 typed action 与 +当前准入；历史回放中的 action 按钮不能执行。queue、inbox、steer 保留不同投递 +合同，provider 未知 / 不支持字段保持可见。 + +## 8. 迁移与回滚 + +首版在既有团队检查入口显式选择，列表使用同一组事实。无需迁移 canonical +记录，缺少字段的旧 journal 保持不完整。可丢弃 transport cache 必须能从当前 +获准 owner 重建；没有持久事件时明确显示历史缺口。 + +关闭场景或订阅可回到列表 / Chat，不能中断成员工作。不得删除 journal、重绑 +session、提高历史 receipt 的语义强度或重置 provider。默认概览 / 导航变化 +必须在 commit/push 前完成仓库要求的具体首屏预览审批；对另一处文案精简的 +批准不等于批准本场景。 + +## 9. 验收 + +| ID | 所需证据 | 通过标准 / 排除项 | +| --- | --- | --- | +| V1 事实 | 合成 lifecycle、过期 artifact、缺失采用、连接丢失 | 各轴分离，不虚构 executing/accepted/complete | +| V2 流 | 重复、源版本乱序、cursor 过期、重启、撤权、溢出 | 当前回读一致，缺口可见，无重复副作用或跨受众泄露 | +| V3 体验 | packaged 桌面 + 390px、键盘 / 屏幕阅读器、200% 缩放、减弱动态、安静 / 受阻 / 不可用 | 来源 / action 路径与返回焦点可用，选择保留，不用闲置动画伪装工作 | +| V4 研究闭环 | 原 coordinator + 2–3 真实 worker，两轮，一次 peer 依赖、拒绝 / 纠偏和一个 worker 中断而另一个继续 | pinned 独立验证、请求方采用、原主力继续均由 CLI 与 packaged UI 回读；无人肉转发；R2/G1 标准不变 | +| V5 规模 | 相同记录负载的 10/30/100 actor、高事件率、长回放 | 测前冻结负载 / 设备 / 目标；报告延迟、CPU/GPU、内存、缺口、注意力成本；合成渲染压力不等于真实 fleet 资格 | +| V6 回滚 | 功能关闭、旧线程、暂停、renderer 错误、源丢失 | 原 owner 继续拥有工作；列表 / Chat 可用，不触发校验风暴 | +| V7 理解 | 对比列表与场景，观察者识别阻塞、结论变化、证据来源和下一 owner | 预先声明问题 / 成功标准，记录错误与答题时间；新奇评分不能掩盖理解退步 | + +V5 候选目标在实验前定稿：前台动态目标 60fps / 优雅降级底线 30fps；p95 交互 +低于 100ms；p95 投影接收到 DOM 低于 500ms；有界 10 分钟回放，30 分钟运行 +无无界堆增长。renderer 延迟与 provider / 网络分开统计，记录硬件、视口、 +事件率、可见节点 / 边及测量方法。这些是设计预算，不是已测性能或发布保证。 + +## 10. 运行合同 + +安静、失联、陈旧、恢复中与受阻都是明确状态。失联停止证据移动，保留上次观察 +和时间；隐藏 tab 停止动画，reduced motion 去掉移动 / 镜头，仍保留离散事件。 +“暂停动画”与“暂停协调员”分别提供。日志不自动滚动，live region 不逐 token +播报，未知成本不显示 0，声音默认关闭。 + +聚合受声明 scope 与完整性约束；分页不完整时不能宣称整个 fleet 健康。群组折叠 +仍能发现失败。复用 scheduler backpressure，展示没有 wake policy 或模型刷新循环。 +记录观测错误时不把敏感事件正文写入诊断。 + +## 11. 交付顺序与激进推进 R2 的关系 + +| 切片 | 完整有用结果 | 进入 / 退出 | Owner 与回滚 | +| --- | --- | --- | --- | +| L0 设计与可追溯性 | 可交互合成研究、源审计、可执行验收计划 | 设计评审；V1/V3 概念检查；不声明 live | Presentation；丢弃原型，保留决策 | +| L1 真实小团队现场 | 既有团队执行入口展示有范围的工作、返回与准确证据；列表 / 场景一致 | 已准入绑定；真实入口的 V1/V2/V3/V6 | Collaboration 投影 + Chat/packaged UI；关闭场景 | +| L2 持续研究 | 原投研 Agent 完成 V4，用户回放纠偏→采用→下一轮 | runtime 凭证 / 准入修复及可用 L1；逐 profile 资格化后 V4 | 既有 R2/R3 owner；停新准入、保留结果 | +| L3 语义缩放与规模 | 子团队、来源 / 结论版本对比、实测 10/30/100 展示负载下的聚焦回放 | typed 谱系 producer 与 L2；V5/V7、批准首屏编排 | Presentation + evidence owner；降细节 / 回退 renderer | + +下一个可执行实现是 **L1**，完整的观测到用户 vertical，而不是单独一个 event +schema。同一交付计划包含源收集、typed 投影、重连语义、打包前端、独立 CLI +回读和失败用例。选择新 transport schema 前先盘点既有 receipt 身份，昂贵校验 +离开常规轮询。 + +执行优先级仍是 R2：当前 runtime owner 解决准入 / 认证，证明两轮有用闭环， +随后沿既有注册与 settings 验证受治理任务派生及经批准的成员 / profile 创建。 +提高并发前必须处理并行返回 / join 和原主力继续。静态绑定是有界起点，不是可 +重复动态配员系统。不能另建编排项目、替换正在进行的接入修复、降低 gate 或用 +视觉完成掩盖执行失败。Lark 消费同样的有界事实与既有 action，完整返回旅程 +单独验收，不要求把 3D UI 搬到 Lark。 + +## 12. 未决事项 + +- **Presentation owner，L1 前：** 最终美术方向与 motion grammar。建议融合指挥台 + 与空间研究工作室，以语义缩放、结论修订为特色；验证有内容与失败状态。 +- **Collaboration / transport owner，L1 前：** 复用或扩展哪些 receipt 支持可重启 + 团队流。先审计持久覆盖；现有 owner 无法直接提供需要的投影时才选 scoped cache。 +- **Presentation / performance owner，L3 前：** 保持 SVG 还是引入 PixiJS/Rive。 + 对比相同场景、无障碍、bundle 与内存；本文不引入依赖。 +- **Research capability owner，结论视图上线前：** 公开安全 claim/source/version/ + adoption 字段及隐私。kernel 不从文字分类报告独立性，也不推断金融建议置信度。 + +## 附录 A：调研与交付证据 + +在所列基线上，源码审计确认第 4 节的 inventory、preflight、team dialog 和 Turn +SSE。第 6 节一手资料支撑设计选项，不构成产品验收。一个本地合成交互研究展示 +空间成员、单次交接动态、纠偏回放、证据深入与降级状态；不启动 Agent，不读取 +私人投研数据，未作为产品发布，也不充当真实 runtime 证据。 + +**当前检查点：** L0 设计提案；本文未实现 L1–L3。合成研究可探索 V1/V3； +V2/V4/V5/V6/V7 在获得其要求的生产或测量证据前都未资格化。不晋升 G1/G3/G4。 +准确交付 PR 承载验证与 review；总路线只投影此边界，不再追加运行任务账本。 diff --git a/docs/architecture/rfcs/loopx-overall-roadmap-v0.md b/docs/architecture/rfcs/loopx-overall-roadmap-v0.md index 21219ad3a4..5f013ec9d0 100644 --- a/docs/architecture/rfcs/loopx-overall-roadmap-v0.md +++ b/docs/architecture/rfcs/loopx-overall-roadmap-v0.md @@ -39,7 +39,7 @@ P0 blocks correctness or continuity in the current user journey. P1 enables repe | **S2 Typed kernel and durable authority · P0/P1** | Effect/Todo/quota/recovery owners, TS migration and store candidates exist; writer cutover/provider promotion remain incomplete | Migrate one real transaction/recovery lifecycle at a time, with semantic counterexamples before cutover/deletion. R1 correctness precedes migration volume. Require real backend, concurrency/fence, ambiguous commit, retention/export recovery, bridge costs and D1–D3 evidence | | **S3 Goal planning and multi-Agent collaboration · P0/P1** | Vision/replan, peer frontiers, claim/lease, directory, manager_context and explicit continuation exist; general handoff/shared amendment remain incomplete | R2 proves peer dependency; R3 closes parallel joins, pipelines, help/review, continuation and automatic return; R4 delivers one intent-preserving amendment class. Cover cycles, invalidated inputs, rejection/deferral, lease transfer, competing bases and aggregate acceptance | | **S4 Runtime/host/daemon · P0/P1** | Attached/managed, Turn, broker, runtime connectors and Desktop repairs exist; registration does not establish executable capacity | Qualify multi-Turn supervision for one real supported combination; restart/cancel/drain/stop retain work and fence old executors. Then expand host parity, unique service-profile ownership, clean installation and upgrades; show unsupported adapter capabilities | -| **S5 Frontend, Lark and human interaction · P0/P1** | Local chat, settings, proposals and partial Goal Channel verticals exist; shared audience/session/work readback needs qualification | One journey spans settings, work graph, handoff, blockers, cost, corrections, artifacts and return. Shared typed projections; reconnect/repeated-click/stale/original-route cases. Then intelligent review, keyboard accessibility, bilingual terminology, actionable errors and offline degradation; interrupt only for actual decisions | +| **S5 Frontend, Lark and human interaction · P0/P1** | Local chat, settings, proposals and partial Goal Channel verticals exist; shared audience/session/work readback needs qualification | One journey spans settings, work graph, handoff, blockers, cost, corrections, artifacts and return. Shared typed projections; reconnect/repeated-click/stale/original-route cases. [Live team workspace](live-team-workspace-v0.md) makes exchange, revision and original-coordinator continuation visible. Then intelligent review, keyboard accessibility, bilingual terminology, actionable errors and offline degradation; interrupt only for actual decisions | | **S6 Materials, evidence, memory and learning · P1** | Authority registry, material lifecycle/frontier, decision context, reward memory and turn recall exist; direction baseline and parts of attribution remain proposed | Connect material revision→same-Agent read→decision reference→artifact/outcome. Expose expiry/revocation/source loss and forgetting policy. Handoff preserves decision-relevant summaries and authorized artifacts; qualify OpenViking/Obelisk as optional providers. Prove causal utility with controls, not relevance alone | | **S7 Budget, scheduling and fleet scale · P0 observation/P1–P2 expansion** | Quota/scheduler and partial usage aggregates exist; full provider cost, distributed reservations and hundred-Agent concurrency need evidence | Separate configured budget, admission, consumption and estimates; unknown is not zero and replay cannot double-charge. R7 pagination/bounded summaries, provider/host limits, fairness, backpressure, event wake and isolation; report registration/activity/throughput and cost per accepted outcome separately | | **S8 Capabilities, extensions and domain integration · P1/P2** | Capability catalog, extension lifecycle, hooks, engineering/research/content/office capabilities and computer-use contracts exist | First exercise the shared control plane with existing issue-fix/PR-review and material/research callers. Every provider has readiness/version/permissions/default-off/uninstall/rollback/isolation and real-entry evidence. New domain effects start with one simulated operation, not a marketplace or workflow DSL | @@ -295,6 +295,19 @@ This qualifies a local execution-facts readback, not the full R2 ladder: assignm receipt integration, provisioning, remote probes, two-cycle continuation and Lark qualification remain open under their existing owners. +**Live team experience is a core S5 outcome.** The [Live Team Workspace +RFC](live-team-workspace-v0.md) joins a precise command surface with a spatial +research studio: visible artifact exchange, sourced disagreement, conclusion +revision, replay and semantic zoom. It is a presentation slice of existing +owners, not a second orchestrator. The next complete frontend slice connects +requester-scoped execution/return facts to one recoverable projection and the +packaged workspace, with independent CLI readback and honest degraded states. +Then qualify the original research coordinator through the same two real cycles +required above. Existing runtime admission/provider repairs retain their owner; +a synthetic scene is not readiness evidence. Expand display load 10→30→100 only +under frozen rendering/attention budgets, separately from live R7 concurrency. +Current status is design proposal; no G1 or default-screen promotion. + ### R3: Semantic Requests and Automatic Return - **Owner:** manager RFC M2/M3; migrate existing `manager_context` request/tracking/return into one typed collaboration transaction, incorporating the #4094 adapter. diff --git a/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md b/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md index 810c686049..b8bdd681dc 100644 --- a/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md +++ b/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md @@ -39,7 +39,7 @@ LoopX 的目标是让人用本地前端或 Lark 提出、修订和验收复杂 | **S2 typed 内核与 durable authority · P0/P1** | Effect/Todo/quota/recovery owner、TS 事务迁移与 store 候选已存在；writer 和 provider 晋升仍未全闭合 | 每次迁移一个真实事务/恢复生命周期，先语义反例再切换/删除旧 owner；R1 正确性先于迁移数量。真实 backend、并发/fence、ambiguous commit、保留/导出恢复、bridge 成本及 D1–D3 资格 | | **S3 目标规划与 multi-Agent 协作 · P0/P1** | Vision/replan、peer frontier、claim/lease、directory、manager_context 和显式接续有基础；通用 handoff/共享修订未闭环 | R2 必须证明 peer 依赖；R3 完成并行汇合、流水线、求助/复核、接续、自动回报；R4 做一个保持 intent 的 amendment class。检查依赖环、输入失效、拒绝/延期、lease 转移、同基线竞争及 aggregate acceptance | | **S4 runtime/host/daemon · P0/P1** | attached/managed、Turn、broker、runtime connector 和 Desktop 修复存在；“registered”不等于可执行 | 选择一个真实合格组合完成多 Turn supervision；restart/cancel/drain/stop 不丢工作且旧 executor 被 fence。之后扩 host parity、service-profile 唯一 owner、干净安装与版本升级；按 adapter 能力显示不支持项 | -| **S5 前端、Lark 与人机交互 · P0/P1** | 本地对话、settings、proposal 和部分 Goal Channel vertical 已有；统一受众/会话/工作回读仍需资格 | 用一个团队旅程贯穿设置、工作图、handoff、阻塞、成本、修订、产物和回报；共享 typed projection，验证重连/重复点击/stale/原路反馈。再做 intelligent review、无障碍键盘流程、中英术语、错误可恢复和离线降级；只在真实决策处打断人 | +| **S5 前端、Lark 与人机交互 · P0/P1** | 本地对话、settings、proposal 和部分 Goal Channel vertical 已有；统一受众/会话/工作回读仍需资格 | 用一个团队旅程贯穿设置、工作图、handoff、阻塞、成本、修订、产物和回报；共享 typed projection，验证重连/重复点击/stale/原路反馈。再做 intelligent review、无障碍键盘流程、中英术语、错误可恢复和离线降级；只在真实决策处打断人；[团队实时工作区](live-team-workspace-v0.zh-CN.md)让交换、修订与原协调员继续推进可见 | | **S6 材料、证据、记忆与学习 · P1** | authority registry、material lifecycle/frontier、decision context、reward memory、turn recall 已有；方向基线和部分归因仍是提案 | 先打通“材料 revision→同 Agent 阅读→决策引用→产物/结果”；失效、撤销、来源消失与遗忘策略可回读。handoff 保存影响决策的摘要与授权 artifact；OpenViking/Obelisk 按可选 provider 资格化。utility 的因果收益另以对照证明，不把相关性当提升 | | **S7 预算、调度与 fleet 规模 · P0 观测/P1–P2 扩展** | quota/scheduler 与部分 usage aggregate 存在；全 provider 成本、分布式资源预留及百 Agent 并发尚需证据 | 先区分配置预算、准入、消耗与估算；未知成本不记零、重复事件不双记。R7 分页/有界摘要、provider/host 限流、公平性、背压、事件唤醒与失败隔离；分别报告注册数/活跃数/吞吐量和每个验收成果成本 | | **S8 能力、扩展与领域集成 · P1/P2** | 已有 capability catalog、extension 生命周期、hook、工程/研究/content/office 能力及 computer-use 合同 | 优先用现有 issue-fix/PR-review 和材料/研究 caller 检验共享控制面；每个 provider 带 readiness、版本、权限、默认关闭、卸载/回滚、失败隔离与真实入口证据。新 domain effect 从模拟单操作闭环开始，不先建市场或通用工作流 DSL | @@ -272,6 +272,15 @@ executor/profile 检查启动条件。任务准入、当前 pinned 验收绑定 检查，不新增状态账本，也不启动执行。这只验收本地执行事实的读回；计划分配回执 整合、通用创建、远端探针、两轮持续协作及 Lark 等价仍由原 owner 继续推进。 +**团队现场是 S5 核心产品目标。** [团队实时工作区 RFC](live-team-workspace-v0.zh-CN.md) +融合精确指挥台与空间研究工作室：展示产物交换、有来源的分歧、结论修订、回放与 +语义缩放。它复用既有 owner，不建立第二套编排。下一完整前端切片将 requester +范围的执行 / 返回事实接入可恢复投影与打包工作区，提供独立 CLI 回读和真实降级 +状态；随后按上述标准验收原投研 coordinator 的两轮真实协作。既有 runtime +准入 / provider 修复保留原 owner；合成场景不能充当 readiness 证据。展示负载 +10→30→100 在测前冻结渲染 / 注意力预算，和 R7 实际并发分开资格化。当前为设计 +提案，不晋升 G1 或默认首屏。 + ### R3：语义请求与自动回报 - **Owner：** 管家 RFC M2/M3；从已有 `manager_context` request/tracking/return 迁移到单一 typed collaboration 事务，纳入 #4094 adapter。 From c6d78c464b7eb9c91dd262ca1853c02110777c46 Mon Sep 17 00:00:00 2001 From: huangruiteng <14976749+huangruiteng@users.noreply.github.com> Date: Sun, 20 Sep 2026 03:15:38 +0800 Subject: [PATCH 2/2] docs: require one real correction journey before expanding team visuals Signed-off-by: huangruiteng <14976749+huangruiteng@users.noreply.github.com> --- docs/architecture/rfcs/README.md | 5 +- .../rfcs/live-team-workspace-v0.md | 50 ++++++++++++++++--- .../rfcs/live-team-workspace-v0.zh-CN.md | 40 ++++++++++++--- .../rfcs/loopx-overall-roadmap-v0.md | 5 +- .../rfcs/loopx-overall-roadmap-v0.zh-CN.md | 6 +-- 5 files changed, 86 insertions(+), 20 deletions(-) diff --git a/docs/architecture/rfcs/README.md b/docs/architecture/rfcs/README.md index 9db0729498..87a5e8d993 100644 --- a/docs/architecture/rfcs/README.md +++ b/docs/architecture/rfcs/README.md @@ -292,8 +292,9 @@ changes. - **RFC status:** Draft presentation slice under intelligent presentation. - **Delivery on `main`:** Design proposal only; no live team-stream qualification. - **Current boundary:** A command surface plus spatial research studio, evidence - handoff motion, conclusion revision and replay. L1 binds existing execution - facts to packaged UI; L2 requires original-coordinator two-cycle acceptance; + handoff motion, conclusion revision and replay. L1 requires one real + objection→revision→acceptance→adoption journey with evidence and intervention/stop + feedback in packaged UI; L2 requires original-coordinator two-cycle acceptance; L3 qualifies semantic zoom and display scale. No new scheduler or authority. - [Human Attention Wishlist v0](human-attention-wishlist-v0.md) ([中文版](human-attention-wishlist-v0.zh-CN.md)) diff --git a/docs/architecture/rfcs/live-team-workspace-v0.md b/docs/architecture/rfcs/live-team-workspace-v0.md index e939d5dd84..da8276e9a9 100644 --- a/docs/architecture/rfcs/live-team-workspace-v0.md +++ b/docs/architecture/rfcs/live-team-workspace-v0.md @@ -285,6 +285,41 @@ commit/push; approval of an unrelated copy edit does not approve this scene. ## 9. Acceptance +### First live outcome: one understandable correction + +The immediate product target is one **real objection → revision → independent +acceptance → requester adoption**, visible through the packaged workspace with +usable evidence and intervention/stop controls. This is L1's exit condition; +it is not postponed to a later visual polish or continuity phase. A synthetic +script, current-status roster, or stream connection alone cannot satisfy it. + +| Moment / user question | Required production evidence and interaction | +| --- | --- | +| Who objected, and to what? | An attributed request or review outcome bound to the challenged claim/artifact version and its evidence; do not classify free text into an authoritative objection state | +| What changed? | Explicit lineage between the original and revised artifacts, with inspectable inputs and a readable difference; differing hashes alone do not establish a correction | +| What passed acceptance? | Independent acceptance for that exact revision, with verifier/source identity and freshness; superseded inputs invalidate the current badge rather than preserving a reassuring historical label | +| Who adopted it? | Requester-owned evidence referencing the accepted revision and the resulting synthesis or next work; delivery, receiver task adoption, requester read/consumption and agreement counts are insufficient | +| Can I intervene or stop? | An existing authorized intervention route with visible queued/delivered/applied distinctions, and a stop/pause route naming its actual target and returned effect; pending or failed control stays visible | + +Exercise this same real episode through the frontend and independent CLI +readback. Show unexecuted work, rejected acceptance, missing adoption, stale +artifacts and lost observation without completing the chain visually. A user +must be able to locate the relevant evidence and tell what the control changed. +Pausing the coordinator does not stop dispatched workers: name those workers and +the remaining execution scope. If worker cancellation is unsupported, state it +explicitly; a whole-team stop claim requires worker-owner termination readback. + +The current producer gap is substantive: `consume_return` records consumption, +not version-bound requester adoption; research-specific adoption checks in an +example are not a generic production projection. Extend the existing owning +contract and its real caller where necessary, rather than inventing frontend +completion from prose. Do not claim L1 complete until this gap is closed for the +selected episode. Motion is retained only when it clarifies these transitions; +remove effects that obscure absent execution, absent acceptance or source loss. +Broader semantic zoom, extra actors and renderer experiments follow this exit. + +### Qualification matrix + | ID | Required evidence | Passing result / exclusion | | --- | --- | --- | | V1 Truth | Synthetic lifecycle, stale artifact, missing adoption and lost connection | All axes remain distinct; no fabricated executing/accepted/complete state | @@ -322,15 +357,16 @@ sensitive event bodies in diagnostics. | Slice | Complete useful result | Entry / exit | Owner and rollback | | --- | --- | --- | --- | | L0 Design and traceability | Interactive synthetic study, source audit and executable acceptance plan | Design review; V1/V3 concept checks; no live claim | Presentation; discard prototype, retain decisions | -| L1 Live small-team scene | Existing Team execution opens current scoped work, follows returns and exposes exact evidence; list and scene agree | Existing admitted bindings; V1/V2/V3/V6 with real entrypoint | Collaboration projection + Chat/packaged UI; disable scene | +| L1 One real correction | Objection→revision→independent acceptance→requester adoption is understandable, with evidence, intervention and accurately scoped stop controls | Section 9 first live outcome; V1/V2/V3/V6 and episode-level V7 through real frontend + CLI | Collaboration/evidence/control owners + Chat/packaged UI; disable scene without implying work stopped | | L2 Continuous research | Original research Agent completes V4 and the user replays correction→adoption→next cycle | Runtime credential/admission repairs and usable L1; qualify each host profile, then V4 | Existing R2/R3 owners; stop new admission and retain results | | L3 Semantic zoom and scale | Subteams, source/conclusion version comparison and focused replay at measured 10/30/100 display load | Typed lineage producers and L2; V5/V7 and approved first-screen composition | Presentation + evidence owner; reduce detail/revert renderer | The next executable implementation is **L1**, a complete observation-to-user -vertical, not an event schema alone. Include source collection, typed projection, -reconnect semantics, packaged UI, independent CLI readback and failure cases in -one delivery plan. Inventory existing receipt identities before choosing a new -transport schema. Keep expensive validation outside routine polling. +vertical with the complete Section 9 correction episode. Include missing +version/adoption producers, source collection, typed projection, reconnect +semantics, packaged UI, evidence navigation, intervention/stop feedback, +independent CLI readback and failure cases in one delivery plan. Inventory +existing receipt identities before choosing a new transport schema. Keep expensive validation outside routine polling. R2 remains the execution priority: resolve runtime admission/authentication in its current owner; prove two useful cycles; then qualify governed task derivation and @@ -345,8 +381,8 @@ its complete return journey is separately qualified and is not a 3D UI port. ## 12. Open decisions - **Presentation owner, before L1:** final art direction and motion grammar. - Recommendation: a command surface plus spatial research studio, with semantic - zoom and conclusion revision as the signature. Test populated and failed states. + Recommendation: a command surface plus spatial research studio, judged first + on one real correction. Defer expanded semantic zoom; test populated and failed states. - **Collaboration/transport owners, before L1:** reuse or extend existing event receipts for a restartable team stream. Audit durable coverage; choose a scoped cache only if the current owners cannot supply the needed projection directly. diff --git a/docs/architecture/rfcs/live-team-workspace-v0.zh-CN.md b/docs/architecture/rfcs/live-team-workspace-v0.zh-CN.md index 464c11b23c..6ff5ca50e9 100644 --- a/docs/architecture/rfcs/live-team-workspace-v0.zh-CN.md +++ b/docs/architecture/rfcs/live-team-workspace-v0.zh-CN.md @@ -232,6 +232,34 @@ session、提高历史 receipt 的语义强度或重置 provider。默认概览 ## 9. 验收 +### 首个真实结果：看懂一次纠偏 + +眼前的产品目标是：让用户在打包前端里看懂一次**真实的“提出异议 → 修订 → +独立验收 → 请求方采用”**，能够找到证据、介入或停止。这是 L1 的退出条件， +不能推迟到后续视觉打磨或持续运行阶段。合成脚本、当前状态名单或连通事件流 +都不能单独满足它。 + +| 时刻 / 用户问题 | 必须具备的生产证据与交互 | +| --- | --- | +| 谁对什么提出异议？ | 有归属的请求或复核结果，绑定受质疑的结论 / 产物版本及证据；不能从自由文本推断权威异议状态 | +| 到底改了什么？ | 原产物与修订产物之间明确的谱系，可检查输入并读懂差异；哈希不同本身不能证明发生了纠偏 | +| 什么通过了验收？ | 针对准确修订版本的独立验收，带 verifier / 来源身份与新鲜度；输入被替代后不能继续显示令人安心的历史通过标签 | +| 谁采用了它？ | 请求方拥有的采用证据，引用通过验收的版本及形成的综合结论或下一步工作；投递、接收方采用任务、请求方已读 / 消费、赞成人数都不够 | +| 能介入或停止吗？ | 既有获准干预入口，区分排队 / 投递 / 应用；停止或暂停入口说明实际目标和返回效果，待处理 / 失败控制保持可见 | + +前端和独立 CLI 回读同一次真实过程。尚未执行、验收失败、缺少采用、产物过期 +和观测丢失都不能在画面里补成闭环。用户必须能找到相关证据，并知道控制改变了 +什么。暂停协调员不会停止已派发成员：须显示这些成员及仍在执行的范围。成员 +不支持取消时明确说明；宣称整队停止需要成员执行 owner 的终止回读。 + +当前生产者缺口有实质意义：`consume_return` 只记录消费，不是绑定版本的请求方 +采用；示例中的投研专用采用检查，也不是通用生产投影。必要时扩展既有 owning +contract 及其真实调用方，不能在前端从文字制造完成状态。所选过程补齐这些 +证据前，不宣称 L1 完成。动态只有让上述变化更清楚才保留；掩盖未执行、未验收 +或来源失联的效果应移除。更广的语义缩放、成员扩张和 renderer 探索在此后推进。 + +### 资格化矩阵 + | ID | 所需证据 | 通过标准 / 排除项 | | --- | --- | --- | | V1 事实 | 合成 lifecycle、过期 artifact、缺失采用、连接丢失 | 各轴分离，不虚构 executing/accepted/complete | @@ -263,14 +291,14 @@ V5 候选目标在实验前定稿：前台动态目标 60fps / 优雅降级底 | 切片 | 完整有用结果 | 进入 / 退出 | Owner 与回滚 | | --- | --- | --- | --- | | L0 设计与可追溯性 | 可交互合成研究、源审计、可执行验收计划 | 设计评审；V1/V3 概念检查；不声明 live | Presentation；丢弃原型，保留决策 | -| L1 真实小团队现场 | 既有团队执行入口展示有范围的工作、返回与准确证据；列表 / 场景一致 | 已准入绑定；真实入口的 V1/V2/V3/V6 | Collaboration 投影 + Chat/packaged UI；关闭场景 | +| L1 一次真实纠偏 | 看懂异议→修订→独立验收→请求方采用，可查证据、干预并使用范围准确的停止控制 | 第 9 节首个真实结果；真实前端 + CLI 的 V1/V2/V3/V6 和本过程 V7 | Collaboration / evidence / control owner + Chat/packaged UI；关闭场景不暗示工作停止 | | L2 持续研究 | 原投研 Agent 完成 V4，用户回放纠偏→采用→下一轮 | runtime 凭证 / 准入修复及可用 L1；逐 profile 资格化后 V4 | 既有 R2/R3 owner；停新准入、保留结果 | | L3 语义缩放与规模 | 子团队、来源 / 结论版本对比、实测 10/30/100 展示负载下的聚焦回放 | typed 谱系 producer 与 L2；V5/V7、批准首屏编排 | Presentation + evidence owner；降细节 / 回退 renderer | -下一个可执行实现是 **L1**，完整的观测到用户 vertical，而不是单独一个 event -schema。同一交付计划包含源收集、typed 投影、重连语义、打包前端、独立 CLI -回读和失败用例。选择新 transport schema 前先盘点既有 receipt 身份，昂贵校验 -离开常规轮询。 +下一个可执行实现是 **L1**，覆盖第 9 节完整纠偏过程的观测到用户 vertical。 +同一交付计划包含缺失的版本 / 采用 producer、源收集、typed 投影、重连语义、 +打包前端、证据导航、干预 / 停止反馈、独立 CLI 回读和失败用例。选择新 transport +schema 前先盘点既有 receipt 身份，昂贵校验离开常规轮询。 执行优先级仍是 R2：当前 runtime owner 解决准入 / 认证，证明两轮有用闭环， 随后沿既有注册与 settings 验证受治理任务派生及经批准的成员 / profile 创建。 @@ -282,7 +310,7 @@ schema。同一交付计划包含源收集、typed 投影、重连语义、打 ## 12. 未决事项 - **Presentation owner，L1 前：** 最终美术方向与 motion grammar。建议融合指挥台 - 与空间研究工作室，以语义缩放、结论修订为特色；验证有内容与失败状态。 + 与空间研究工作室，首先以一次真实纠偏评判；扩大语义缩放后置，验证有内容与失败状态。 - **Collaboration / transport owner，L1 前：** 复用或扩展哪些 receipt 支持可重启 团队流。先审计持久覆盖；现有 owner 无法直接提供需要的投影时才选 scoped cache。 - **Presentation / performance owner，L3 前：** 保持 SVG 还是引入 PixiJS/Rive。 diff --git a/docs/architecture/rfcs/loopx-overall-roadmap-v0.md b/docs/architecture/rfcs/loopx-overall-roadmap-v0.md index 5f013ec9d0..7c119a1fd4 100644 --- a/docs/architecture/rfcs/loopx-overall-roadmap-v0.md +++ b/docs/architecture/rfcs/loopx-overall-roadmap-v0.md @@ -300,8 +300,9 @@ RFC](live-team-workspace-v0.md) joins a precise command surface with a spatial research studio: visible artifact exchange, sourced disagreement, conclusion revision, replay and semantic zoom. It is a presentation slice of existing owners, not a second orchestrator. The next complete frontend slice connects -requester-scoped execution/return facts to one recoverable projection and the -packaged workspace, with independent CLI readback and honest degraded states. +one real objection→revision→independent acceptance→requester adoption episode +to the packaged workspace, with inspectable evidence, intervention/stop feedback, +independent CLI readback and honest degraded states. Then qualify the original research coordinator through the same two real cycles required above. Existing runtime admission/provider repairs retain their owner; a synthetic scene is not readiness evidence. Expand display load 10→30→100 only diff --git a/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md b/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md index b8bdd681dc..77a8cae363 100644 --- a/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md +++ b/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md @@ -274,9 +274,9 @@ executor/profile 检查启动条件。任务准入、当前 pinned 验收绑定 **团队现场是 S5 核心产品目标。** [团队实时工作区 RFC](live-team-workspace-v0.zh-CN.md) 融合精确指挥台与空间研究工作室：展示产物交换、有来源的分歧、结论修订、回放与 -语义缩放。它复用既有 owner，不建立第二套编排。下一完整前端切片将 requester -范围的执行 / 返回事实接入可恢复投影与打包工作区，提供独立 CLI 回读和真实降级 -状态；随后按上述标准验收原投研 coordinator 的两轮真实协作。既有 runtime +语义缩放。它复用既有 owner，不建立第二套编排。下一完整前端切片在打包工作区 +呈现一次真实异议→修订→独立验收→请求方采用，提供可检查证据、干预 / 停止 +反馈、独立 CLI 回读与真实降级状态；随后按上述标准验收原投研 coordinator 的两轮真实协作。既有 runtime 准入 / provider 修复保留原 owner；合成场景不能充当 readiness 证据。展示负载 10→30→100 在测前冻结渲染 / 注意力预算，和 R7 实际并发分开资格化。当前为设计 提案，不晋升 G1 或默认首屏。