将 9 个 schema 定义文件从 JSON 转为 YAML 格式，与用户配置格式统一。

改动:
- schemas/v1/*.json → schemas/v1/*.yaml（内容等价，YAML 语法）
- $id 和 $ref 的 URL 保持 .json 后缀作为 URI 标识符（JSON Schema 标准约定）
- loadEmbeddedSchemas: 读 YAML → 转 JSON → 注册到 compiler
- go:embed 改为 *.yaml
- 更新 types.go 注释和 README.md 文件名引用

测试覆盖: 87.0% (全项目测试通过)

Refs #11

新建 internal/runtime 包，实现基于 schema.Workflow 的可配置调度引擎。

核心组件:
- Tool 接口: 统一 agent/command/function 三种工具类型
- AgentTool: 包装 agent.Runner，调用 AI 智能体
- CommandTool: 通过 exec.CommandContext 执行 shell 命令
- ToolRegistry: 支持 name + alias 查找
- ExecutionContext: stage 间数据传递 + loop 状态维护
- Expression: 结构化 ExitCondition 求值（6 operator）+ 字符串表达式
- StageExecutor: 串行执行（when/on_blocked/input_from）
- ParallelExecutor: 并行执行（all/any/first_success 三种 join）
- LoopExecutor: 循环执行（exit_when/judge/max_iterations/max_consecutive_unknown）
- Engine: 顶层入口，消费 *schema.Workflow

Prompt 模板外部化:
- prompts/*.md: 5 个模板文件（coding/review/issue-handling/issue-post/loop-judge）
- PromptBuilder: 从 embed FS 加载 text/template 渲染

测试覆盖: 76.9% (76 个测试用例)
全项目测试通过

Refs #11

Stage 8 of scenario-driven scheduling refactor.

- dispatch.go: 624 -> 152 lines; Service now holds *runtime.Engine and
  *schema.Workflow, Dispatch delegates to engine.Run
- delete prompts.go; 5 prompt templates migrated to
  internal/runtime/prompts/*.md (rendered by prompt_builder.go)
- add runtime_adapter.go: ArtifactResolverAdapter bridges
  pipeline.ArtifactSet <-> runtime.ArtifactResolver, routing LoadResult
  by stage name to typed loaders
- add runtime/default_workflow.{go,yaml}: go:embed built-in default
  workflow; --workflow flag supports external override
- cmd/worker/main.go: serviceFactory signature returns error to surface
  workflow load failures; --workflow flag wired through buildService
- tests: add TestToPipelineResult + 2 end-to-end integration tests
  (BlockedPath, CompletedPath); drop 9 dispatch state-machine tests
  whose behavior is now covered by internal/runtime engine tests
- .gitignore: anchor 'worker' pattern to repo root ('/worker') so
  cmd/worker/ is no longer ignored; add .trae/

Known gap: issue-post-blocked stage removed -- Engine terminates on
Blocked, so the stage could never run. issue-handling BLOCKED now
returns Blocked directly without posting a blocking comment. Future
Engine enhancement (on_blocked: continue) can restore this.

Verification: go build/vet/test all pass; runtime 76.1%, pipeline 70.3%
coverage.

Refs: #11

Complete Issue #11 verification standard: agent role definitions are
now loaded from workflow config, no longer hardcoded in BuiltinAgents().

- delete Config fields: DefaultAgent, ReviewerAgent, IssuePostAgent,
  LoopJudgeAgent, MaxCodingReviewRounds, LoopJudgeStartRound
  (last 3 were dead code; first 3 now derived from workflow)
- delete BuiltinAgents(), AgentByName(), AgentNames(), AgentRole struct
  (dead code: only config_test.go self-tested them)
- delete defaultMaxCodingReviewRounds / defaultLoopJudgeStartRound consts
- add runtime.lookupDisplayNameByPromptTemplate: derives display name
  from workflow.Tools by prompt_template, with fallback chain
  display_name > aliases[0] > name > empty
- modify agent_tool.buildPromptData: derive DefaultAgent/ReviewerAgent/
  IssuePostAgent from EC.Workflow instead of Config
- Config now holds only Repo/IssueNumber/GitHubToken (3 runtime params)
- update tool_test.go: add testWorkflow helper, pass workflow to
  IssueHandling test so prompt derivation works
- add TestLookupDisplayNameByPromptTemplate (5 sub-cases)
- update config_test.go: drop BuiltinAgents tests, simplify defaults test

Backward compatible: default workflow's display_names match the old
hardcoded values (developer/QA lead/PM), so prompts render identically
without --workflow flag.

Verification: go build/vet/test pass; runtime 76.5%, config 100% coverage.

Refs: #11

Complete Issue #11 documentation standard: explain how to customize
scheduling strategy via workflow config.

- update 'why code-bee' bullets: replace '极简内核' with '场景驱动调度'
- update tech stack table: scheduler now describes workflow-driven orchestration
- add architecture overview with ASCII pipeline diagram
- update usage example: add --workflow flag, fix output message
- update project structure: add internal/runtime/ and internal/schema/,
  refresh package descriptions
- add '自定义 Workflow' section: config structure, 3编排原语, 3 tool types,
  role name derivation, links to schema/default-workflow/prompts
- update roadmap: mark scenario-driven scheduling, composable pipeline,
  tool aliases as completed; add on_blocked recovery item

Refs: #11

- PromptBuilder 加载外部目录 .md 模板，overlay 同名内置模板
- Engine 通过 Functional Options 接收 WithPromptsDir
- Service 透传 EngineOption，CLI 新增 --prompts-dir flag
- 新增 overlay 单元测试 + 自定义 workflow 端到端集成测试

Closes #11 (prompts externalization gap)

Two-stage flag parsing had a bug: parseWorkflowFlag and parsePromptsDirFlag
each registered only a single flag in their own FlagSet. Go's flag package
stops parsing at the first unregistered flag, so when --repo (registered
only in runCLI's FlagSet) appeared before --workflow/--prompts-dir, the
probe functions returned empty strings and the flags were silently ignored
(falling back to builtin workflow / builtin prompts).

Discovered via smoke test: --prompts-dir /nonexistent/prompts did not
error out, and the binary proceeded to dispatch with builtin prompts.

Fix: merge the two probe functions into parseExtraFlags, which registers
all known flags (--repo/--issue/--version as placeholders) so parsing
proceeds past them. Add table-driven regression test covering 6 arg
orderings including the previously broken 'flags after --repo' case.

The same bug affected --workflow since its introduction; this commit
fixes both.

…tion level

Replace the single ASCII tree diagram with three Mermaid diagrams that
show how the architecture is actually implemented:

1. Layered component architecture (flowchart TB) — CLI/pipeline/runtime/
   schema four layers with all key structs (Engine, ToolRegistry,
   StageExecutor, AgentTool, PromptBuilder, ExecutionContext,
   ArtifactResolverAdapter, etc.) and interface-implementation edges
   (ArtifactResolver <- Adapter, Runner <- agent.Runner).

2. Dispatch call chain (sequenceDiagram) — full call timeline from
   runCLI -> Service.Dispatch -> Engine.Run -> StageExecutor ->
   AgentTool -> runner.Run, including result-file reset/read/write
   timing and early-termination semantics.

3. Config loading & data flow (flowchart LR) — workflow dual-source
   (go:embed builtin + --workflow override) and prompt overlay
   (--prompts-dir covers same-name builtin), plus result-file flow
   between stages via ExecutionContext.

Add an Engine termination semantics table documenting how
Blocked/ManualRequired/Completed trigger early return, with a note on
the known issue-post-blocked gap.

- internal/runtime/expression_test.go: goimports formatting
- internal/runtime/loop_executor_test.go: goimports formatting
- internal/schema/loader.go: goimports formatting
- internal/runtime/tool_test.go: SA4006 assign of tool is never used

Found by CI golangci-lint.

…orts

- expression.go: extract evalRegexCondition helper (18→reduced)
- stage_executor.go: extract handleOnBlocked helper (19→reduced)
- agent_tool.go: extract fillCoding/Review/IssuePost/LoopJudgePromptData helpers (30→reduced)
- parallel_executor.go: extract aggregateStatus helper, remove unused base param (28→reduced)
- loop_executor.go: extract executeBody & handleJudge helpers (31→reduced)
- validator.go: check AddResource error, remove unused workflowSchemaPath const
- main.go: add nolint:errcheck for fs.Parse in probe FlagSet
- loop_judge_test.go: add nolint:unused for readLoopHistoryForTest

All found by CI golangci-lint (errcheck, gocyclo, unused, goimports).

Extract postRound helper for per-round maintenance (ConsecutiveUnknown,
judge, LastFeedback). Drops Execute from 18 to <=15 gocyclo threshold.

Found by CI golangci-lint gocyclo check.