haolpku · Oct 21, 2026 · Oct 20, 2026 · Oct 21, 2026
diff --git a/.claude/skills/workspace-ingest/SKILL.md b/.claude/skills/workspace-ingest/SKILL.md
@@ -1,6 +1,6 @@
 ---
 name: workspace-ingest
-version: 1.1.0
+version: 1.2.0
 description: Guide Codex to ingest a workspace into DataMind with source evidence, checked write receipts, and optional atomic publication.
 keywords: [ingest, workspace, 入库, 导入, 文档, 表格]
 ---
@@ -9,7 +9,9 @@ keywords: [ingest, workspace, 入库, 导入, 文档, 表格]
 
 Codex is the main agent: inspect sources, choose surfaces, extract relationships,
 and generate SQL when needed. DataMind executes validated data operations.
-MCP tools have the `datamind_` prefix; Python DataPlane tools omit it.
+MCP tools have the `datamind_` prefix. The operation steps below describe MCP.
+Python DataPlane tools omit the prefix and use a bound RequestContext; use
+Workspace.reader for pinned Python reads.
 
 1. Start `operation_begin(mode="write")`, then fetch this Skill with `skill_get`
    using that operation_id. Reuse the ID for related calls so receipts record the
@@ -28,10 +30,12 @@ MCP tools have the `datamind_` prefix; Python DataPlane tools omit it.
    from raw_file_read. Use `graph_upsert_triples` for explicit standalone facts;
    use `graph_build_lineage` for file dependencies. These writes do not update
    published workspace snapshots.
-5. Every write returns an IngestReceipt: inspect each `results[].status`, error,
+5. Every business data write returns an IngestReceipt: inspect each `results[].status`, error,
    skipped sources and details. A receipt ID alone does not mean success.
    Report the actual surfaces, receipt IDs, workspace revision, truncation and
-   failures. A failed publication keeps the previous revision readable.
+   failures. Build metadata tools (`build_start`, `build_freeze`, `build_export`)
+   return build state rather than ingestion receipts. A failed publication keeps
+   the previous revision readable.
 
 KB still requires a configured embedding provider. Graph/SQL/Skills/Memory do
 not require a DataMind conversation-model key. Do not switch providers silently.

diff --git a/.env.datamind.example b/.env.datamind.example
@@ -1,101 +1,38 @@
-# DataMind — new-style nested env (Phase 1+)
-#
-# Copy to `.env.datamind` (or merge into `.env`) and fill in real values.
-# All new-code ENV vars use the DATAMIND__<section>__<field> prefix with
-# double-underscore as the nested delimiter. The legacy .env.example keys
-# (LLM_API_BASE, etc.) still drive the old main.py / server.py — nothing
-# is broken by this file.
-
-# ---- LLM (wire protocol is explicit) ----
-DATAMIND__LLM__API_BASE=http://35.220.164.252:3888
-DATAMIND__LLM__API_KEY=sk-your-key-here
-DATAMIND__LLM__PROTOCOL=anthropic
-# For an OpenAI-format gateway use:
-# DATAMIND__LLM__PROTOCOL=openai_chat_completions
-DATAMIND__LLM__MODEL=claude-sonnet-4-6
-# Cheaper model used for subagents / background summarisation
-DATAMIND__LLM__FALLBACK_MODEL=claude-haiku-4-5-20251001
-# Defaults to the main protocol. Set only if the same gateway exposes the
-# fallback model through a different wire protocol.
-# DATAMIND__LLM__FALLBACK_PROTOCOL=anthropic
-DATAMIND__LLM__TIMEOUT_S=60
-DATAMIND__LLM__CONNECT_TIMEOUT_S=10
-DATAMIND__LLM__MAX_RETRIES=3
-
-# ---- Embedding ----
-# provider: must match a name registered in capabilities/embedding/providers/
-# Built-ins (to be added in Phase 2): openai, voyage, huggingface, siliconflow
+# DataMind 2: Codex is the main agent; DataMind supplies data tools.
+# Copy to .env.datamind. Keep secrets out of version control.
+DATAMIND__DATA__PROFILE=default
+# DATAMIND__DATA__BASE_DIR=/srv/datamind
+
+# CLI / HTTP / UI Codex host. Omit model/provider to inherit Codex login/config.
+# DATAMIND__CODEX__MODEL=your-supported-codex-model
+# DATAMIND__CODEX__BASE_URL=https://your-gateway.example/v1
+# DATAMIND__CODEX__API_KEY=your-private-key
+# DATAMIND__CODEX__WORKSPACE=/srv/workspace
+# DATAMIND__CODEX__REASONING_EFFORT=low
+DATAMIND__CODEX__APPROVAL_MODE=auto_review
+DATAMIND__CODEX__TIMEOUT_S=600
+DATAMIND__CODEX__MAX_CONCURRENCY=4
+# A trusted unattended host can use deny_all plus an explicitly authorized list.
+# DATAMIND__CODEX__PREAPPROVED_TOOLS=["datamind_operation_begin"]
+
+# KB embeddings are independent of the main Codex model.
 DATAMIND__EMBEDDING__PROVIDER=openai
 DATAMIND__EMBEDDING__API_BASE=https://api.openai.com/v1
-DATAMIND__EMBEDDING__API_KEY=sk-your-key-here
 DATAMIND__EMBEDDING__MODEL=text-embedding-3-small
-# Leave blank to auto-detect dimension at first call
-# DATAMIND__EMBEDDING__DIMENSION=
-DATAMIND__EMBEDDING__BATCH_SIZE=32
-DATAMIND__EMBEDDING__TIMEOUT_S=30
-DATAMIND__EMBEDDING__CONNECT_TIMEOUT_S=10
-DATAMIND__EMBEDDING__MAX_RETRIES=3
-# DATAMIND__EMBEDDING__MAX_BATCH_TOKENS=8000
+# DATAMIND__EMBEDDING__API_KEY=your-embedding-key
+# Or install .[huggingface] and use a local model:
+# DATAMIND__EMBEDDING__PROVIDER=huggingface
+# DATAMIND__EMBEDDING__MODEL=BAAI/bge-small-zh-v1.5
+# DATAMIND__EMBEDDING__DIMENSION=512
 
-# ---- Retrieval ----
-# strategy: simple | multi_query | hybrid  (registered in capabilities/kb/providers/)
 DATAMIND__RETRIEVAL__STRATEGY=simple
 DATAMIND__RETRIEVAL__TOP_K=5
 DATAMIND__RETRIEVAL__CHUNK_SIZE=512
 DATAMIND__RETRIEVAL__CHUNK_OVERLAP=64
-DATAMIND__RETRIEVAL__RERANK=false
-# DATAMIND__RETRIEVAL__RERANK_MODEL=
-
-# ---- Graph ----
-# backend: networkx (built-in) | neo4j (future)
 DATAMIND__GRAPH__BACKEND=networkx
-# DATAMIND__GRAPH__DSN=bolt://user:pw@localhost:7687
-DATAMIND__GRAPH__EMBED_ENTITIES=false
-
-# ---- Database ----
-# dialect: sqlite | mysql | postgres  (registered in capabilities/db/providers/)
 DATAMIND__DB__DIALECT=sqlite
-# SQLite example (blank DSN = auto under storage/<profile>/demo.db)
-# DATAMIND__DB__DSN=
-# MySQL example:
-# DATAMIND__DB__DIALECT=mysql
-# DATAMIND__DB__DSN=mysql+pymysql://user:password@host:3306/dbname
-# Postgres example:
-# DATAMIND__DB__DIALECT=postgres
-# DATAMIND__DB__DSN=postgresql+psycopg://user:password@host:5432/dbname
 DATAMIND__DB__READ_ONLY=true
 DATAMIND__DB__ROW_LIMIT=1000
-DATAMIND__DB__QUERY_TIMEOUT_S=10.0
-
-# ---- Memory ----
-# backend: sqlite (built-in) | redis (future) | postgres (future)
 DATAMIND__MEMORY__BACKEND=sqlite
-# DATAMIND__MEMORY__DSN=
-DATAMIND__MEMORY__SHORT_TERM_TURNS=20
-DATAMIND__MEMORY__LONG_TERM_ENABLED=true
-
-# ---- Data profile ----
-# Switches data/profiles/<profile>/ and storage/<profile>/ in lockstep.
-DATAMIND__DATA__PROFILE=default
-
-# ---- Agent loop backend ----
-# "native": built-in tool-use loop on the anthropic SDK — talks directly
-#           to whatever DATAMIND__LLM__API_BASE points at. Default. No
-#           extra process required.
-# "sdk":    uses claude-agent-sdk (which spawns the `claude` CLI) and
-#           routes through a local claude-code-router (CCR) instance.
-#           Best when your upstream gateway only speaks OpenAI format —
-#           CCR translates Anthropic↔OpenAI. Unlocks Hooks / Subagents /
-#           Compaction / Plan mode.
-# Both backends use the SAME 23 DataMind tools and emit the SAME SSE
-# events, so server / CLI / frontend don't care which is active.
-DATAMIND__AGENT__BACKEND=native
-# Only relevant when BACKEND=sdk:
-# DATAMIND__AGENT__CCR_BASE_URL=http://127.0.0.1:13456
-# DATAMIND__AGENT__CCR_API_KEY=dummy
-DATAMIND__AGENT__MAX_TURNS=12
-DATAMIND__AGENT__MAX_TOOL_CALLS=24
-DATAMIND__AGENT__MAX_INPUT_TOKENS=120000
-DATAMIND__AGENT__MAX_TOOL_RESULT_CHARS=12000
-DATAMIND__AGENT__MAX_TOOL_RESULT_ROWS=100
-DATAMIND__AGENT__WALL_CLOCK_TIMEOUT_S=300
+DATAMIND__HOOKS__ENABLED=true
+# DATAMIND__HOOKS__PATH_ALLOWLIST_EXTRA=["/srv/authorized-documents"]
diff --git a/.github/workflows/python-ci.yml b/.github/workflows/python-ci.yml
@@ -132,6 +132,10 @@ jobs:
           python scripts/ci_required_tests.py -q
           datamind/tests/test_direct_data_plane.py
           datamind/tests/test_codex_host.py
+          datamind/tests/test_unified_host.py
+          datamind/tests/test_cli_codex.py
+          datamind/tests/test_http_api.py
+          datamind/tests/test_experiment_faults.py
           plugins/datamind-context/tests/test_datamind_mcp.py
 
   gate:

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,15 @@
 # Changelog
 
+## 2.0.0 — Unified Codex host
+
+- CLI, HTTP and browser UI now use one official Codex SDK host for reads and writes.
+- Removed StoreAgent/RetrieveAgent classes, assembly, role prompts and Claude SDK/CCR adapter.
+- Data services assemble independently of all agent loops and conversation credentials.
+- Read requests bind profile/session, access mode and one published revision below the prompt boundary.
+- Request-scoped traces preserve actual MCP evidence and receipts under Codex code mode; cancellation closes active SDK turns.
+- Migrated examples/benchmarks, retained checkpoint and fault tests, and updated gateway/config migration guidance.
+
+
 ## 1.2.0 — Direct Codex data layer
 
 - Local Codex plugin and optional official `openai-codex` SDK host share direct