feat(data-agent): add evidence-backed autonomous analysis

.github/workflows/backend-prisma-quality.yml

@@ -5,6 +5,7 @@ on:
     paths:
       - "apps/backend/**"
       - "packages/shared-types/**"
+      - "packages/analysis-task-protocol/**"
       - "pnpm-lock.yaml"
       - "package.json"
       - "scripts/check-backend-capability-boundaries.ts"
@@ -15,6 +16,7 @@ on:
     paths:
       - "apps/backend/**"
       - "packages/shared-types/**"
+      - "packages/analysis-task-protocol/**"
       - "pnpm-lock.yaml"
       - "package.json"
       - "scripts/check-backend-capability-boundaries.ts"
@@ -66,6 +68,15 @@ jobs:
       - name: Prisma generate
         run: pnpm --filter @text2sql/backend run prisma:generate
 
+      - name: Prepare canonical PostgreSQL schema
+        env:
+          PGPASSWORD: postgres
+        run: |
+          if ! psql -h localhost -U postgres -d postgres -tAc "SELECT 1 FROM pg_database WHERE datname='text2sql_ci'" | grep -q 1; then
+            psql -h localhost -U postgres -d postgres -c "CREATE DATABASE text2sql_ci"
+          fi
+          pnpm --filter @text2sql/backend exec prisma migrate deploy
+
       - name: Backend lint
         run: pnpm --filter @text2sql/backend run lint
 
@@ -75,16 +86,39 @@ jobs:
       - name: Backend build
         run: pnpm --filter @text2sql/backend run build
 
-      - name: Backend test
-        run: pnpm --filter @text2sql/backend run test
+      - name: Backend test with coverage
+        run: pnpm --filter @text2sql/backend exec jest --coverage --runInBand
 
       - name: Text2SQL v2 focused coverage report
+        run: pnpm --filter @text2sql/backend run collect:text2sql-v2-focused-coverage-gate:strict
+
+      - name: Text2SQL accuracy contracts and composite gates
+        run: |
+          pnpm run text2sql:no-legacy-compat:check
+          pnpm run governance:terminology:check
+          pnpm --filter @text2sql/backend run collect:text2sql-v2-eval-gate
+          pnpm --filter @text2sql/backend exec jest --runInBand \
+            test/integration/text2sql-accuracy-gate.spec.ts \
+            test/integration/text2sql-accuracy-replay-parity.spec.ts
+          TS_NODE_TRANSPILE_ONLY=1 pnpm --filter @text2sql/backend exec ts-node \
+            --project tsconfig.json \
+            scripts/collect-text2sql-accuracy-gate.ts \
+            > /tmp/text2sql-accuracy-gate.json
+          jq -e '.rollout.releaseDecision == "HOLD"' /tmp/text2sql-accuracy-gate.json
+          jq -e '.summary.realOutcomePairCount == 0' /tmp/text2sql-accuracy-gate.json
+
+      - name: Data Agent composite release manifest
         run: |
-          pnpm --filter @text2sql/backend exec jest \
-            --coverage \
-            --runInBand \
-            text2sql-v2-focused-coverage-gate.spec.ts
-          pnpm --filter @text2sql/backend run collect:text2sql-v2-focused-coverage-gate
+          pnpm --filter @text2sql/backend exec jest --runInBand \
+            test/integration/data-agent-release-gate.spec.ts
+          TS_NODE_TRANSPILE_ONLY=1 pnpm --filter @text2sql/backend exec ts-node \
+            --project tsconfig.json \
+            scripts/collect-data-agent-release-gate.ts \
+            > /tmp/data-agent-release-gate.json
+          jq -e '.version == "data-agent-release-manifest/v1"' /tmp/data-agent-release-gate.json
+          jq -e '.rollout.releaseDecision == "HOLD"' /tmp/data-agent-release-gate.json
+          jq -e '.rollout.gatePass == false' /tmp/data-agent-release-gate.json
+          jq -e '.topology.mode == "single_workflow"' /tmp/data-agent-release-gate.json
 
       - name: R6 scenario package tests
         env:
@@ -148,6 +182,8 @@ jobs:
 
           jq -e '.status == "success"' /tmp/health.json
           jq -e '.data.dependencies.postgres.configured == true' /tmp/health.json
+          jq -e '.data.dependencies.text2sqlAccuracy.mode == "shadow"' /tmp/health.json
+          jq -e '.data.dependencies.text2sqlAccuracy.capabilities.parser == "available"' /tmp/health.json
 
           session_json=$(curl -sS -X POST http://127.0.0.1:3100/api/v1/sessions \
             -H "content-type: application/json" \

.gitignore

@@ -2,6 +2,14 @@
 /data/*
 /vibe/*
 /docs/*
+!/docs/text2sql-project-architecture-business-flow-review.md
+!/docs/standards/
+/docs/standards/*
+!/docs/standards/text2sql-accuracy-gate-spec.md
+!/docs/standards/data-agent-autonomous-analysis-spec.md
+!/docs/runbooks/
+/docs/runbooks/*
+!/docs/runbooks/data-agent-task-recovery-and-release.md
 # Node / monorepo artifacts
 node_modules/
 dist/
@@ -20,4 +28,4 @@ __pycache__/
 .env.*.local
 .omx/
 .enter/*
-graphify-out/*
+graphify-out/*

AGENTS.md

@@ -72,13 +72,24 @@ CI 参考：
 - `RAG 配置` 健康检查需同时覆盖 `dry-check`（草稿）与 `persisted-check`（已保存），并校验返回 `checkedAgainst=draft|persisted`、`reasonCode` 可解释，且检测失败不应清空草稿输入。
 - 网关 smoke：`node tests/smoke/nginx-dev-gateway-smoke.mjs` 可区分 frontend/backend/stream 三类上游失败。
 - 健康检查：`GET http://localhost:3002/health` 应可用（后端内部端口检查），且 `dependencies.ragConfig.embedding/rerank` 应可见当前激活 provider+model+configSource 摘要。
+- Text2SQL accuracy 健康摘要还需包含 `dependencies.text2sqlAccuracy.mode/supportedSlices/supportedDialects/guidelineDigest/capabilities`，不得暴露连接、授权明细或真实 Fixture 路径。
+- 自治分析健康摘要还需包含 `dependencies.analysisRuntime.canonicalStore/durableWorkflow/commandOutbox/telemetry`；telemetry 不得含 Prompt、SQL rows、网页正文、token、URL secret、PII 或高基数 metric label。
 - 若本次改动涉及流式/工具调用：需关注 stream 与 tool 相关字段一致性（细节见 LLM 迁移规范）。
 - 若本次改动涉及 Text2SQL v2 read-model/delivery hard-cut：执行
   - `pnpm --filter @text2sql/backend run collect:text2sql-v2-eval-gate`
   - （发布阻断）`pnpm --filter @text2sql/backend run collect:text2sql-v2-eval-gate:strict`
   - `pnpm --filter @text2sql/backend run collect:text2sql-v2-focused-coverage-gate`
   - `pnpm run text2sql:no-legacy-compat:check`
   - 并核对 `rollout.recommendedStage` 与 `rollout.rollbackSuggested`。
+- 若本次改动涉及 Text2SQL 准确率、Receipt、修复或 replay parity：执行
+  - `pnpm --filter @text2sql/backend run collect:text2sql-accuracy-gate`
+  - `pnpm --filter @text2sql/backend run collect:text2sql-v2-eval-gate`
+  - `pnpm --filter @text2sql/backend run collect:text2sql-v2-focused-coverage-gate`
+  - 发布阻断使用 `collect:text2sql-accuracy-gate:strict`；没有受信 real Outcome Receipt 时必须保持 `HOLD`，不得以现有 focused/eval 全过替代。
+- 若本次改动涉及自治 AnalysisTask、Deep Search、Evidence/Claim、Memory/Skill 或 ReleaseManifest：执行
+  - `pnpm --filter @text2sql/backend run collect:data-agent-release-gate`
+  - 核对每个 component 的 exact version/scope/freshness/evidence refs/owner approval，以及 `topology.mode`。
+  - 发布阻断使用 `collect:data-agent-release-gate:strict`；缺真实签名 Text2SQL Outcome 或代表性 analyst Outcome 时必须保持 `HOLD`。
 - 若本次改动涉及 modeling parity 指标：执行 `pnpm --filter @text2sql/backend run collect:modeling-parity-shadow-gate`，确认 `relationshipPlatform/semanticSpine/modelingWorkspace` 三维输出可生成。
 - 若本次改动需要发布门禁（go/no-go）：执行 `pnpm --filter @text2sql/backend run collect:modeling-parity-shadow-gate:strict`，并检查 `rollout.recommendedStage` 与 `rollout.rollbackSuggested`。
 
@@ -142,6 +153,8 @@ CI 参考：
 - `pnpm --filter @text2sql/backend run collect:text2sql-v2-focused-coverage-gate` 输出需包含 scoped coverage、关键文件门槛、A-M flow blockers 与 eval fixture 行为测试追溯。
 - strict-completion 语义补齐后，focused coverage 输出还需包含 `strictCompletionRows` 评估结果（metadata grounding / correction grounding / context-pack parity）。
 - runtime-intelligence 生效后，focused/eval 输出还需覆盖 `runtime-plan-consistency`、`artifact-ref-compaction`、`smart-defaults-evidence`、`plain-general-no-sql`、`large-context-compaction`、`validation-diagnostics`、`correction-grounding`、`execution-preview`、`all-stage-stream-lifecycle`；focused coverage 还需包含 `runtimeArtifactProducerRows` 与 `streamLifecycleRows`。
+- accuracy closure 生效后，runtime artifact producer 还必须覆盖 `accuracy_receipts`；sync/stream/run view/save view/replay 共享安全 accuracy summary，correction 固定为 allowlisted AST Patch、最多两次并回到完整 validation。
+- `TEXT2SQL_ACCURACY_MODE=shadow` 只记录候选证据摘要，不得阻断现有 baseline 用户链路或贡献 `GO`；`enforce` 才要求 trusted policy/schema、七层 passed Receipt 与 final ValidationReceipt 后交付。
 - `pnpm run text2sql:no-legacy-compat:check` 必须通过。
 
 ### D. Governance 术语硬切规范
@@ -182,6 +195,47 @@ CI 参考：
 必跑检查：
 - `pnpm run backend:capability-boundary:check`（落地后）
 
+### F. Text2SQL 准确率门禁规范
+来源：`docs/standards/text2sql-accuracy-gate-spec.md`
+
+适用范围：
+- Text2SQL QueryContract、版本元组、Receipt chain、结果 Oracle、有界 AST 修复与 replay parity。
+- `apps/backend/src/modules/conversation/runtime/evaluation/**`
+- `apps/backend/scripts/collect-text2sql-accuracy-gate.ts`
+
+关键 MUST：
+- 合成评测只能证明 closeout，不能替代真实、受信、签名的 Outcome Receipt。
+- `shadow` 不阻断 baseline；`enforce` 才以七层 Receipt、Execution Permit 与 final ValidationReceipt 为硬条件。
+- 修复最多两次，只允许 AST 等价 allowlist，且每次修复后回到完整 validation。
+- 最终发布决策只允许 `GO/HOLD/NO_GO/ROLLBACK`；无真实证据时必须 `HOLD`。
+
+必跑检查：
+- `pnpm --filter @text2sql/backend run collect:text2sql-accuracy-gate`
+- `pnpm --filter @text2sql/backend run collect:text2sql-accuracy-gate:strict`（发布阻断）
+
+### G. Data Agent 自治分析规范
+来源：`docs/standards/data-agent-autonomous-analysis-spec.md`
+
+适用范围：
+- `apps/backend/src/modules/conversation/analysis/**`
+- `apps/backend/src/modules/platform/{durable,artifacts,observability}/**`
+- `apps/backend/src/modules/knowledge/{research,assets}/**`
+- `apps/frontend/src/{app,components}/analysis/**`
+- `apps/backend/scripts/collect-data-agent-release-gate.ts`
+
+关键 MUST：
+- PostgreSQL Analysis Ledger 是 Task/Revision/Attempt/Event/Artifact/Receipt/Manifest 权威；Temporal、UI、SSE 与 OTel 都是 orchestration/projection。
+- Worker 只能提出 candidate，Commit Guard 才能提交；Correction 追加 Revision 并传播 stale/invalidated，旧 replay 保留。
+- 网页内容始终 untrusted；Report 只能从 Claim projection 生成；生产 Memory/Skill 只能来自治理后的 active KnowledgeAsset。
+- ReleaseManifest 的 required evidence 若 unknown/stale/failed 不得 GO；缺真实签名 Outcome 固定 HOLD；安全失败给出 NO_GO/ROLLBACK。
+
+必跑检查：
+- `pnpm --filter @text2sql/backend run collect:data-agent-release-gate`
+- `pnpm --filter @text2sql/backend run collect:data-agent-release-gate:strict`（发布阻断）
+
+恢复与回滚：
+- `docs/runbooks/data-agent-task-recovery-and-release.md`
+
 说明：
 - 以上仅为执行摘要，细节规则以 standards 原文为准。
 

README.md

@@ -16,6 +16,7 @@ This project is not a prompt-only SQL generator. It is a Text2SQL platform proto
 - Before SQL generation, the agent retrieves schema, glossary terms, historical examples, and semantic assets.
 - Generated SQL must pass read-only, safety, permission, dialect, and execution checks.
 - Each run has a `runId` for tracing, RAG evidence, delivery artifacts, and replay.
+- Long-running autonomous analysis uses a durable `AnalysisTask` ledger, bounded workers, evidence/claim lineage, correction revisions, and a composite release manifest.
 - Frontend and backend stay aligned through shared types and an SSE protocol package for sync responses, streaming responses, and run details.
 
 ## Problems Solved
@@ -35,6 +36,9 @@ This project is not a prompt-only SQL generator. It is a Text2SQL platform proto
 5. **End-to-end demos across multiple datasource types**
    The project supports SQLite, MySQL, PostgreSQL, CSV, and Excel datasources, with frontend workbench pages such as `/data-sources`, `/chat`, `/settings`, and `/modeling` for demos and continued extension.
 
+6. **Durable autonomous analysis without unverifiable autonomy**
+   `/analysis` turns a decision goal into a versioned WorkGraph. PostgreSQL owns task truth, Temporal owns durable orchestration, and deterministic commit guards decide which Evidence, Calculation, Claim, Conflict, and Report artifacts may become visible. Missing signed business or analyst Outcome evidence keeps release status at `HOLD`.
+
 ## Feature Preview
 
 ### Semantic Modeling Workbench
@@ -85,6 +89,7 @@ apps/backend                  NestJS API, Text2SQL runtime, governance, knowledg
 apps/frontend                 Next.js frontend workbench
 packages/shared-types         Shared frontend/backend types
 packages/chat-stream-protocol SSE envelope, parser, terminal guard, and UI projection helpers
+packages/analysis-task-protocol Durable task event, replay, and terminal-state contracts
 infra                         Local PostgreSQL, Redis, and Nginx orchestration
 data                          Local uploads, SQLite files, and runtime data
 docs                          Solutions, standards, troubleshooting, and understanding documents
@@ -118,7 +123,7 @@ flowchart LR
 
 | Domain | Directory | Responsibility |
 | --- | --- | --- |
-| `conversation` | `apps/backend/src/modules/conversation` | Chat entry points, Text2SQL workflow, LangGraph runtime, delivery contract |
+| `conversation` | `apps/backend/src/modules/conversation` | Chat entry points, Text2SQL workflow, autonomous analysis, LangGraph runtime, delivery contract |
 | `governance` | `apps/backend/src/modules/governance` | Workspaces, datasource binding, table permissions, users, and settings governance |
 | `knowledge` | `apps/backend/src/modules/knowledge` | RAG retrieval, semantic assets, glossary, memory, graph, and modeling context |
 | `platform` | `apps/backend/src/modules/platform` | Persistence, query execution, cache, configuration, observability, and read-model guards |
@@ -241,6 +246,7 @@ Default URLs:
 
 - Gateway: `http://localhost:3000`
 - Datasource entry: `http://localhost:3000/data-sources`
+- Autonomous analysis: `http://localhost:3000/analysis`
 - Frontend direct debugging: `http://localhost:3001`
 - Backend health check: `http://localhost:3002/health`
 
@@ -267,6 +273,14 @@ node tests/smoke/nginx-dev-gateway-smoke.mjs
 - `/modeling`: datasource modeling, relationships, and semantic views
 - `/prompts`: prompt template management
 
+### Autonomous Analysis
+
+1. Open `/analysis` in a selected workspace.
+2. Create a Goal Contract with objective, decision use, datasource scope, deliverables, risk, and budgets.
+3. Start the task and inspect the versioned WorkGraph, mandatory obligations, live event cursor, budget, Evidence, Conflict, and Report projections.
+4. Pause, resume, cancel, revise, or reconnect without treating a closed browser tab as task cancellation.
+5. Use `taskId` replay and ReleaseManifest for audit; UI progress and telemetry are projections, not Outcome truth.
+
 ## Quality Gates
 
 Repository-wide:
@@ -303,8 +317,12 @@ pnpm run text2sql:no-legacy-compat:check
 pnpm --filter @text2sql/backend run collect:text2sql-v2-eval-gate
 pnpm --filter @text2sql/backend run collect:text2sql-v2-focused-coverage-gate
 pnpm --filter @text2sql/backend run collect:modeling-parity-shadow-gate
+pnpm --filter @text2sql/backend run collect:text2sql-accuracy-gate
+pnpm --filter @text2sql/backend run collect:data-agent-release-gate
 ```
 
+`collect:data-agent-release-gate:strict` is the final release blocker. Without fresh, signed, owner-approved Text2SQL and representative analyst Outcome evidence, the expected decision is `HOLD`, even when synthetic, focused, and eval gates pass.
+
 ## Important Rules
 
 - Prisma schema changes must start in `apps/backend/prisma/schema.prisma`, then migrations must be generated by the Prisma CLI.
@@ -325,3 +343,5 @@ pnpm --filter @text2sql/backend run collect:modeling-parity-shadow-gate
 - `docs/standards/llm-stream-tool-migration-spec.md`: LLM stream and tool calling migration standard
 - `docs/standards/governance-terminology-spec.md`: governance terminology hard-cut standard
 - `docs/standards/backend-business-capability-topology-spec.md`: backend capability topology standard
+- `docs/standards/data-agent-autonomous-analysis-spec.md`: autonomous AnalysisTask, Evidence/Claim, KnowledgeAsset, telemetry, and ReleaseManifest standard
+- `docs/runbooks/data-agent-task-recovery-and-release.md`: task recovery, replay, correction, release, and rollback runbook

apps/backend/.env.example

@@ -3,6 +3,37 @@ NODE_ENV=development
 SQLITE_PATH=../../data/sqlite/text2sql.db
 SQLITE_ALLOWED_DIRS=../../data/sqlite,../../data/uploads/datasources
 CORS_ALLOWED_ORIGINS=http://localhost:3000
+# Local/test compatibility only. Production defaults to oidc_bearer and refuses
+# to start with dev_headers.
+AUTH_MODE=dev_headers
+AUTH_POLICY_VERSION=trusted-principal-v1
+# Required when AUTH_MODE=oidc_bearer. OIDC subject must equal PlatformUser.id.
+AUTH_OIDC_ISSUER=
+AUTH_OIDC_AUDIENCE=
+AUTH_OIDC_JWKS_URL=
+AUTH_OIDC_ALLOWED_ALGORITHMS=RS256
+AUTH_OIDC_CLOCK_TOLERANCE_SECONDS=5
+ANALYSIS_ARTIFACT_MAX_BYTES=262144
+ANALYSIS_TASK_ARTIFACT_MAX_BYTES=8388608
+ANALYSIS_DURABLE_PROVIDER=temporal
+TEMPORAL_ADDRESS=127.0.0.1:7233
+TEMPORAL_NAMESPACE=default
+TEMPORAL_ANALYSIS_TASK_QUEUE=text2sql-analysis-v1
+TEMPORAL_CONNECTION_TIMEOUT_MS=5000
+ANALYSIS_EVENT_POLL_INTERVAL_MS=500
+ANALYSIS_MULTI_WORKER_MODE=off
+# Bounded Deep Search is fail-closed until both a versioned workspace policy
+# and a provider key are configured. Provider-generated answers stay disabled.
+ANALYSIS_RESEARCH_ENABLED=false
+ANALYSIS_RESEARCH_PROVIDER=tavily
+ANALYSIS_RESEARCH_ALLOWED_DOMAINS=
+ANALYSIS_RESEARCH_SEARCH_TIMEOUT_MS=10000
+ANALYSIS_RESEARCH_EXTRACT_TIMEOUT_MS=20000
+ANALYSIS_RESEARCH_MAX_CONTENT_BYTES=262144
+ANALYSIS_RESEARCH_RETENTION_DAYS=30
+KNOWLEDGE_ASSET_LEGACY_FIXTURE_MODE=false
+TAVILY_API_KEY=
+TAVILY_API_BASE_URL=
 DATASOURCE_UPLOAD_DIR=../../data/uploads/datasources
 DATASOURCE_UPLOAD_MAX_BYTES=10485760
 DATASOURCE_CONNECT_TIMEOUT_MS=5000
@@ -67,6 +98,20 @@ R1_GATE_MIN_SUCCESS_RATE=0.95
 R1_GATE_MAX_REJECTION_RATE=0.2
 R1_GATE_MAX_HARD_FAILURE_RATE=0.05
 
+# Text2SQL accuracy release evidence. Real fixture data and signing private keys
+# stay outside the repository; this service only reads a controlled root and
+# verifies public keys supplied as a JSON object keyed by key id.
+TEXT2SQL_ACCURACY_FIXTURE_ROOT=../../data/text2sql-accuracy
+TEXT2SQL_ACCURACY_TRUSTED_PUBLIC_KEYS_JSON={}
+TEXT2SQL_ACCURACY_EVIDENCE_MAX_AGE_MS=3600000
+TEXT2SQL_ACCURACY_MODE=shadow
+TEXT2SQL_ACCURACY_SUPPORTED_SLICES=sanitized-sqlite-reference
+TEXT2SQL_ACCURACY_GUIDELINE_DIGEST=data-agent-system-design-2026-07-17
+TEXT2SQL_EXECUTION_TIMEOUT_MS=10000
+TEXT2SQL_EXECUTION_MAX_ROWS=200
+TEXT2SQL_EXECUTION_MAX_BYTES=2097152
+TEXT2SQL_EXECUTION_MAX_AST_NODES=20000
+
 LANGSMITH_TRACING=false
 LANGSMITH_API_KEY=
 LANGSMITH_ENDPOINT=https://api.smith.langchain.com

apps/backend/jest.config.ts

@@ -12,6 +12,8 @@ const config: Config = {
   coverageDirectory: "coverage",
   testEnvironment: "node",
   moduleNameMapper: {
+    "^@text2sql/analysis-task-protocol$":
+      "<rootDir>/../../packages/analysis-task-protocol/src",
     "^@text2sql/chat-stream-protocol$":
       "<rootDir>/../../packages/chat-stream-protocol/src",
     "^@text2sql/chat-stream-protocol/(.*)$":