spec-agent 2.0.5 → 2.0.7

package/.cursor/rules/spec-agent-assistant.mdc

@@ -29,12 +29,23 @@ When user intent is about running `spec-agent`, apply this workflow:
 
 4. Keep responses action-oriented:
    - Provide exact commands.
-   - Explain expected output paths (`output/handoff`, `output/execution`).
+   - Explain expected output paths (`output/handoff`, `output/execution`, `output/prototypes`, `output/prototype_index.json`).
    - Avoid claiming completion without command evidence.
    - Prefer invoking `spec-agent-product-dev-agent` when user asks requirement-to-development end-to-end.
 
 5. Command execution policy (important):
    - Auto-run short safe commands in chat: `spec-agent --version`, `spec-agent status --workspace ./output`, read-only listing/check commands.
-   - For long steps (`pipeline`, `handoff`, `execute`, `round`), do NOT run automatically in chat.
-   - For long steps, output exact command and ask manual run with: "请在终端执行上面命令，执行完成后回复：继续".
+   - Auto-run medium commands in chat when possible: `handoff`, `round`, and feedback updates (`--complete` / `--fail`).
+   - Keep `pipeline` as default manual step for large-doc stability.
+   - For manual steps, output exact command and ask: "请在终端执行上面命令，执行完成后回复：继续".
    - After user replies "继续", read artifacts/logs and provide next single command.
+
+6. Delivery scope policy (current phase):
+   - Frontend-only implementation by default.
+   - Prioritize pages, routes, UI states, forms, interactions, and mock data.
+   - Backend/API/database tasks are marked as "deferred" and not executed in current round unless user explicitly requests backend implementation.
+
+7. Prototype-aware implementation policy:
+   - If `output/prototype_index.json` exists, prioritize prototype-linked tasks.
+   - For task execution, first read `output/handoff/tasks/<TASK_ID>.md` and use "Prototype References" as primary UI evidence.
+   - Treat "Prototype Candidates" as optional references that need human confirmation.

package/.cursor/skills/spec-agent-execution-orchestrator/SKILL.md

@@ -14,6 +14,11 @@ Run the spec-agent execution loop with minimal manual steps:
 4. Apply completion/failure feedback
 5. Continue next round
 
+Current execution scope:
+
+- Frontend implementation tasks only (UI/pages/routes/state/interaction).
+- Backend/API infra tasks are deferred unless explicitly requested by user.
+
 ## When To Use
 
 Use this skill when user asks:
@@ -29,10 +34,9 @@ Use this skill when user asks:
 Execution mode:
 
 - Auto-run short read-only checks in chat, such as `spec-agent status --workspace ./output`.
-- Never auto-run long commands from chat execution tools.
-- Always provide one terminal command at a time.
-- After each command, ask user to reply "继续" before next step.
-- This is mandatory for `handoff`, `execute`, `round`.
+- Auto-run `handoff`, `execute`, `round` when possible.
+- If environment confirmation becomes noisy, switch to one-command manual fallback.
+- For fallback steps, ask user to reply "继续" before next step.
 
 Copy this checklist and keep it updated:
 
@@ -57,7 +61,7 @@ Run:
 spec-agent handoff --workspace ./output --target cursor --include-summaries
 ```
 
-Ask user to run it manually and reply "继续".
+Prefer auto-run; if blocked by environment prompts, ask manual run and reply "继续".
 
 Skip this if `output/handoff/handoff_bundle.json` already exists and user did not request refresh.
 
@@ -69,7 +73,7 @@ Run:
 spec-agent round --workspace ./output --input ./docs --target cursor --max-parallel 4
 ```
 
-Ask user to run it manually and reply "继续".
+Prefer auto-run; if blocked by environment prompts, ask manual run and reply "继续".
 
 If user asks reset:
 
@@ -83,12 +87,16 @@ Inspect:
 
 - `output/execution/inbox/*.md`
 - `output/execution/execution_report.json`
+- `output/prototype_index.json` (if exists)
+- `output/prototypes/*` (if exists)
 
 Return a concise batch summary:
 
 - scheduled task IDs
 - pending/running/succeeded/failed/blocked counts
 - which task to execute first
+- frontend-only kept tasks vs deferred backend tasks
+- high-confidence prototype references per task
 
 ### Step 4: Collect outcomes
 
@@ -108,7 +116,7 @@ Fail tasks:
 spec-agent execute --workspace ./output --fail T003 --error "compile failed"
 ```
 
-Both feedback commands are also manual terminal steps.
+Feedback commands can be auto-run; manual fallback only when needed.
 
 ### Step 6: Continue next batch
 
@@ -140,7 +148,7 @@ Use this output template:
 - If inbox is empty but pending remains, check blocked dependencies and report them.
 - Prefer small rounds (2-6 tasks) to reduce rollback cost.
 - Use `orchestrator_context.json` as decision reference when available.
-- Do not chain multiple long commands in one auto-executed step.
+- Avoid chaining multiple heavy commands in one step.
 
 ## Additional Reference
 

package/.cursor/skills/spec-agent-onboarding-agent/SKILL.md

@@ -13,12 +13,17 @@ Turn spec-agent into a callable assistant workflow with minimal user friction:
 3. Guide first successful pipeline run
 4. Optionally bootstrap handoff + execute loop
 
+Current rollout scope:
+
+- Frontend-only requirement delivery for fast customer demos.
+- Backend implementation is deferred by default.
+
 Execution mode:
 
 - Auto-run short checks in chat (`spec-agent --version`, `spec-agent status --workspace ./output`).
-- Do not auto-run long commands in chat.
-- Print command -> ask user terminal execution -> wait for "继续".
-- Apply to `pipeline`, `handoff`, `execute`, `round`.
+- Auto-run `handoff`, `round`, and feedback updates when possible.
+- Keep `pipeline` as manual-first for large documents.
+- For manual fallback, print command -> ask user terminal execution -> wait for "继续".
 
 ## Step 1: Installation Check
 
@@ -48,7 +53,13 @@ spec-agent status --workspace ./output
 spec-agent round --workspace ./output --input ./docs --target cursor --max-parallel 4
 ```
 
-Ask user to run each command manually in terminal, and continue only after user replies "继续".
+`pipeline` manual-first; `round` auto-run when possible.
+
+If requirement docs include prototype images, verify these artifacts after `pipeline`:
+
+- `output/prototypes/*`
+- `output/prototype_index.json`
+- `output/prototype_index.md`
 
 ## Step 3: Multi-Agent Execution Bootstrap (Optional)
 
@@ -84,4 +95,5 @@ spec-agent execute --workspace ./output --fail T003 --error "compile failed"
 - If user is on private registry, provide one-time `--registry` command variant.
 - Do not skip verification commands.
 - Use `spec-agent round` as the primary one-round action.
-- For long steps, always use manual terminal execution handoff.
+- `handoff` / `round` can be auto-run; `pipeline` remains manual-first.
+- When `prototype_index.json` exists, remind user to use task-level prototype references in `output/handoff/tasks/*.md`.

package/.cursor/skills/spec-agent-product-dev-agent/SKILL.md

@@ -13,6 +13,12 @@ Turn requirement input into executable development workflow:
 3. Generate handoff and execution context
 4. Execute one round and guide task implementation
 
+Current delivery scope: frontend-only MVP.
+
+- Implement pages/routes/components/interactions first.
+- Use mocked data contracts for API-dependent screens.
+- Backend/API/database work is deferred unless user explicitly asks for backend.
+
 ## Trigger Scenarios
 
 Use this skill when user says:
@@ -26,10 +32,11 @@ Use this skill when user says:
 Execution mode for this skill:
 
 - Auto-run short checks in chat: `spec-agent --version`, `spec-agent status --workspace ./output`.
-- For long-running commands, do not auto-execute in chat.
-- Always print one long command at a time and ask user to run it in terminal.
+- Auto-run medium commands in chat: `spec-agent handoff ...`, `spec-agent round ...`, and feedback `round --complete/--fail`.
+- Keep `spec-agent pipeline ...` as default manual terminal step.
+- For manual steps, print one command at a time and ask user to run it.
 - Wait for user confirmation ("继续") before giving next command.
-- Long-running commands include: `pipeline`, `handoff`, `execute`, `round`.
+- Manual-by-default command: `pipeline`.
 
 ### Step 1: Ensure installation
 
@@ -69,7 +76,7 @@ spec-agent handoff --workspace ./output --target cursor --include-summaries
 spec-agent round --workspace ./output --input ./docs --target cursor --max-parallel 4
 ```
 
-Run as two separate manual steps and wait for "继续" after each command.
+Prefer auto-run for both commands; if environment blocks execution, fall back to manual run and wait for "继续".
 
 ### Step 5: Development guidance
 
@@ -77,8 +84,21 @@ Read:
 
 - `output/execution/inbox/*.md`
 - `output/orchestrator_context.json`
+- `output/prototype_index.json` (if exists)
+- `output/prototypes/*` (if exists)
+
+Then guide user to execute current inbox tasks in coding sessions, with frontend-first filtering.
+
+Prototype-first task guidance:
+
+- In each `output/handoff/tasks/<TASK_ID>.md`, use `Prototype References` as the main UI evidence.
+- `Prototype Candidates (medium/low)` should be treated as optional, require human confirmation.
+- If a task has no high-confidence prototype, ask for manual prototype confirmation before heavy UI coding.
+
+Frontend-first task filter:
 
-Then guide user to execute current inbox tasks in coding sessions.
+- Keep: 页面、路由、组件、表单、校验、状态管理、交互细节、UI 可用性。
+- Defer: 后端接口实现、数据库结构、服务部署、鉴权网关等服务端任务。
 
 ### Step 6: Feedback loop
 
@@ -106,4 +126,4 @@ spec-agent round --workspace ./output --input ./docs --complete T001,T002 --fail
 - Do not skip pipeline if dispatch artifacts are missing.
 - Do not manually edit `run_state.json`; only update via CLI.
 - Always use `orchestrator_context.json` as current truth.
-- Do not auto-run long commands from chat execution tools.
+- Do not auto-run `pipeline` when document is large or execution confirmation is frequent.