DC Refactoring Phase 1 Implementation Record

文档属性：历史实施记录。后续阶段和当前状态以当前架构规范与 v2.1 规范为准。

Date: 2026-10-02

Completed

Added immutable Hub PolicySnapshot protected by RWMutex.
Added monotonic security policy version persisted in SQLite.
Hub startup now rebuilds the snapshot from SQLite; snapshot load failure aborts startup rather than defaulting to High.
Security setting writes use SQLite transactions and only publish the new snapshot after the durable write succeeds.
Added a compile-time Tool Registry with fixed RequiredRisk for Low/Medium/High capabilities.
Existing API security checks now resolve required risk through the registry instead of duplicating risk numbers in handlers.
Security-setting audit records include old/new effective mode and snapshot version.

Existing API routes and Agent WebSocket protocol were not changed in this phase.
No production Hub/Agent restart was performed.
When Server.Policy is initialized, security checks use the in-memory snapshot and do not query SQLite on the request hot path.
Existing tests that construct Server without a snapshot retain the legacy storage fallback for test/backward compatibility; production startup always initializes the snapshot.

go test ./... passed.
go test -race ./... passed.
Hub and Agent clean builds passed.
No production service restart was performed; live rollout remains a separate controlled step to preserve the no-interruption requirement.

Added edit_block to the Agent execution engine.
old_text must match exactly once; zero/multiple matches are rejected without mutation.
Replacement uses a same-directory temporary file, preserves file permission bits, fsyncs the file, renames atomically, and syncs the directory.
Added Agent RPC action edit_block, Hub /api/v1/devices/file/edit_block, MCP edit_block, and stdio proxy support.
edit_block is registered as Medium risk and therefore follows the existing Hub policy snapshot gate.
Added optional sub_action protocol field without changing existing message semantics.
Added tests for unique replacement, duplicate/missing rejection, permissions, Low-mode denial, and MCP registration.

Phase 2 rollout remains deliberately separate from production process replacement. Existing production Agent connections have not been restarted.

Added a per-Agent SessionManager with a bounded session count (default 16).
Added fixed-size 2 MiB stdout/stderr RingBuffers; old output is discarded when capacity is exceeded.
Added long-running process actions: start_process, read_process_output, interact_with_process, force_terminate, list_sessions.
Processes run in their own process group so termination/force termination covers the command tree.
Added periodic completed-session cleanup.
Added Hub RPC and HTTP /api/v1/process/session dispatch with action-based lifecycle operations.
Added session tests for ring retention, output capture, termination, duplicate IDs, and session limits.
Existing WebSocket request/response protocol remains compatible; new process actions are additive.

Production Hub/Agents were not restarted during Phase 3 implementation.

Added per-node operation arbitration in cluster.Node.OpMu.
Low-risk read actions use RLock and may run concurrently.
Medium/High mutations use exclusive Lock, preventing overlapping writes/control operations against the same node.
Unknown actions fail closed as High risk for the concurrency gate.
Session IDs are generated with 128-bit crypto/rand entropy when omitted by the API caller; PID is never used as identity.
Session ownership is bound to a non-reversible SHA-256 fingerprint of the authenticated API credential and checked by Agent for read/input/terminate operations.
Session TTL is enforced; stdin is capped at 64 KiB per call.
Agent restart still leaves all sessions in process memory only, so old session state is not restored.

All validation passed with go test ./..., go test -race ./..., Hub/Agent builds, and git diff --check. No production process was restarted.

Added optional protocol/agent capability metadata: ProtocolVersion, AgentVersion, Capabilities.
Hub records the capability set per connected Agent; when a modern Agent explicitly advertises capabilities, unsupported actions fail with capability missing rather than silently degrading.
Old Agents remain compatible because capability metadata is optional and capability enforcement is only applied when a capability set is advertised.
Added structured read-only filesystem operations: list_directory, get_file_info, start_search.
Directory listing supports bounded recursive traversal and a maximum entry count.
File metadata exposes type, size, mode and modification time.
Search is context-cancellable and bounded by a result cap; symlinks are skipped to avoid traversal surprises.
Added Hub endpoints /api/v1/file/list, /api/v1/file/info, /api/v1/file/search.
Added executor tests for directory listing, file metadata and bounded search.

Phase 4 validation: full tests, race tests, Hub/Agent builds and git diff --check passed. Production Hub/Agents were not restarted.