Update sm-flow workflow docs

2026-07-06 09:15:51 +08:00
parent 6ccfd33ec5
commit 3e5c6a159c
20 changed files with 1645 additions and 61 deletions
@@ -0,0 +1,29 @@
+# Acceptance: diagnosis-playbook-skills
+
+## Implemented
+
+- Added six classpath diagnosis playbook skills under `src/main/resources/skills/`.
+- Added classpath skill catalog loading and full skill reading.
+- Added `read_skill` as a Spring AI tool.
+- Wired skill catalog and tool into Chat and AIOps Planner/Executor paths.
+- Kept Chat Verifier isolated from skills.
+- Added focused tests for skill loading, unknown skill handling, and method tool injection.
+
+## Verification
+
+Static/unit verification:
+
+```powershell
+mvn -q "-Dtest=SkillCatalogServiceTest,ChatServiceSequentialAgentTest,AiOpsServiceTest,DiagnosisTraceEvaluatorTest" test
+```
+
+Result: passed.
+
+## Not Verified
+
+- Live LLM behavior with actual `read_skill` tool calls was not run.
+- Java2AI `SkillsAgentHook` integration was not attempted because dependency classes were not locally verified.
+
+## Archive Status
+
+OpenSpec change not archived yet. User should confirm whether to archive `diagnosis-playbook-skills`.
@@ -0,0 +1,32 @@
+# Brief: diagnosis-playbook-skills
+
+## Background
+
+The MVP Agent has trace persistence, evidence tools, verifier gates, and fixed diagnosis eval cases, but scenario-specific diagnosis procedures were still embedded in broad prompts and knowledge-base documents.
+
+## Goal
+
+Introduce versionable diagnosis playbook skills that agents can discover from a compact catalog and read on demand through a `read_skill` tool.
+
+## Scope
+
+- Six initial diagnosis playbooks: payment timeout, MySQL connection pool, Redis timeout, slow response, JVM memory risk, and AIOps alert.
+- Classpath skill catalog loader.
+- `read_skill` Spring AI tool.
+- Prompt catalog injection for Chat and AIOps Planner/Executor paths.
+- Focused tests.
+
+## Non-goals
+
+- No external API or database schema changes.
+- No replacement of `lookup_knowledge`.
+- No Verifier skill loading.
+- No direct dependency on Java2AI `SkillsAgentHook` until local package names are verified.
+
+## Scale
+
+standard
+
+## OpenSpec
+
+`openspec/changes/diagnosis-playbook-skills/`
@@ -0,0 +1,23 @@
+# Decisions: diagnosis-playbook-skills
+
+## Key Decisions
+
+- Use "diagnosis playbook skills" as the canonical term: skill is the runtime loading unit, playbook is the diagnosis workflow content.
+- Keep factual knowledge in `knowledge_base/`; skills contain workflow, evidence order, stop conditions, and report rules.
+- Implement a project-local progressive disclosure mechanism first because local Maven cache did not confirm Java2AI skill hook package names.
+- Keep Verifier unchanged so it only validates existing tool evidence.
+- Do not persist `read_skill` as evidence in `tool_invocation`; diagnostic facts must still come from evidence tools.
+
+## Interface Impact
+
+L2 internal interface:
+
+- New `SkillCatalogService`.
+- New `ReadSkillTool`.
+- Chat/AIOps internal method tools include `read_skill`.
+- No HTTP, DTO, database, or external response contract changes.
+
+## Verification
+
+- `mvn -q "-Dtest=SkillCatalogServiceTest,ChatServiceSequentialAgentTest,AiOpsServiceTest,DiagnosisTraceEvaluatorTest" test`
+- Result: passed.