winagent-sandbox-mcp.git/plans, branch main

release: prepare WinAgent Sandbox MCP v0.1.0

2026-08-14T09:00:29+00:00

Unify the public product name and repository identifier while preserving stable executable, store, AppContainer, namespace, pipe, and mutex identifiers. Make Directory.Build.props VersionPrefix the application-version source of truth, use SDK-generated assembly metadata, and compare helper compatibility by parsed major/minor/build release identity. Rebuild the fixed winagent-sandbox-mcp release staging directory from clean defaults, emit versioned ZIP/checksum assets, verify manifests and fresh extraction, and reject local configuration, user state, and test artifacts. Document release/deployment separation and upgrade responsibilities, add the initial changelog, update operational script paths, and record the v0.1.0 verification results.

test(security): close on-demand elevation acceptance gaps

2026-08-13T14:20:20+00:00

Add an actual-process acceptance harness for the administrative Named Pipe boundary. Exercise production pipe creation and peer PID inspection, remote-form client rejection, Task and protocol identity mismatch, duplicate network application, oversized frames, and an intentional helper-crash recovery point without adding test-only operations to the privileged protocol. Add a reusable PowerShell attack-matrix runner that preserves the helper parent-image check and drives the self-contained published helper through real UAC. Refactor the client pipe setup into internal diagnostic hooks while leaving the production DACL, PIPE_REJECT_REMOTE_CLIENTS, PID comparison, and helper lifetime behavior unchanged. Record the completed UAC rejection, HTTP MCP server-continuity, helper crash, startup Recovery rejection/success, loopback restoration, WFP residue, and regression evidence. Promote the three on-demand elevation traceability rows from PARTIAL to PASS and capture the 80-test Windows Security/Recovery result and final Bundle checksum. Verified: Release build warning/error 0; Core 105 PASS; Planner 10 PASS; administrative boundary 4 PASS; Windows Security/Recovery 80 PASS with skip 0; actual IPC attack matrix PASS; actual Task-time and Recovery-time UAC rejection PASS; Journal/backup/main/helper residue 0; fresh self-contained Bundle show PASS.

feat!: isolate privileged sandbox network operations per task

2026-08-13T13:48:24+00:00

Run agent-mcp with an explicit asInvoker manifest and move WFP and loopback mutations into a fixed requireAdministrator helper under the release admin directory. Add the Agent.Admin client and host projects with task-scoped runas launch, a logon-SID and Administrators-only local named pipe, remote-client rejection, mutual PID validation, strict versioned JSON framing, fixed network-only operations, and bounded messages. Delay ACL and network Task Resource application until the first process_run. Verify executable identity, managed entry points, DNS snapshots, and normal-token WRITE_DAC access before elevation; reuse the helper only within the Task and keep elevation failures terminal without policy replanning. Keep LPAC creation, executable approval, Job confinement, and ACL leases in the normal process. Make startup platform checks read-only, create the fixed Application Profile during apply, preserve dynamic WFP lifetime in the helper, and route persistent loopback recovery through the same elevated boundary. Update recovery journaling, localized elevation and IPC errors, audit events, release publishing, manifests, README guidance, Specifications, Design, implementation records, and requirement traceability for the on-demand elevation model. Add administrative-boundary, file-only no-elevation, elevation-terminal, lazy-resource, crash-recovery, manifest, IPC schema, and operational verification coverage. Verified Release build with zero warnings, Core 105/105, Planner 10/10, administrative boundary 4/4, elevated Windows suite 79/79, and published stdio/HTTP MCP E2E.

feat!: unify sandbox configuration and permission application

2026-08-13T10:15:01+00:00

Replace the split executable approval and sandbox resolution workflows with schema_version 2 and a YAML-driven execution.entries model. Resolve workspace, mount, executable, runtime, network, and resource settings into one config-scoped Applied Configuration snapshot protected by a user-only DACL and written with atomic replacement. Reduce the public CLI to implicit server startup, show, and apply. Remove serve/check, executable and sandbox subcommands, and configuration override flags including --config, --lang, --transport, --path, and --runtime-root. Add migration diagnostics for every removed schema and command surface, and keep server preflight mandatory for both stdio and HTTP. Update effective-policy compilation to validate the applied fingerprint, resolved paths, and executable hashes before startup. Preserve fail-closed behavior while allowing execution entries and runtime roots inside a development workspace to inherit that workspace capability without creating an impossible narrower ACL boundary. Rewrite the README around the consolidated permission matrix and document configuration, application, store location, startup, transport-specific elevation, Python 3.14, PowerShell, and cmd usage. Align specifications, design, implementation plan, traceability, audit report, localization, publish scripts, and operational verification with the breaking schema. Verification: Release build completed with zero warnings and errors; Agent.Core.Tests 104/104 and Agent.Planner.Tests 10/10 passed; administrator Windows Security/Recovery 56/56 and README operational 19/19 passed with no skips; published HTTP MCP completed a cmd/curl/venv-Python task with network deny and zero ACL, WFP, journal, backup, or process residue.

fix(sandbox): correct cmd descendant handling and protected ACL policy

2026-08-13T09:20:33+00:00

Replace the earlier cmd-wide descendant incompatibility diagnosis with the verified distinction between drive-qualified paths and PATH/workspace-relative invocation. Document the cmd YAML configuration, executable resolution, --no-runtime approval flow, and safe descendant invocation examples in the user README. Keep the System32 protected DACL intact for both the OS Runtime Root and approved entry files nested below it. Skip redundant file-level ACL rewrites for approved System32 executables while retaining LPAC confinement, resolved-path checks, SHA-256 approval, WFP policy, and Job Object enforcement. Expand Windows coverage with capability positive/negative matrices, protected-DACL regression tests, cmd/PowerShell/venv Python script execution, nine Entry-to-Descendant combinations, network none/local probes, and randomized ACL round-trip testing. Add read-only CLI/README verification and published HTTP MCP agent_run verification scripts. Update Design, Audit Report, and Requirement Traceability with the corrected root cause, current evidence, withdrawn diagnostic conclusions, and the remaining partial item for per-combination descendant self-recorded LPAC SID/Job evidence. Verified: Release build 0 warnings/errors; Core 103 PASS; Planner 10 PASS; Windows 75 PASS with skip 0; README operational suite 19 PASS; published HTTP MCP agent_run completed 3/3 with cmd, curl, C:\tmp\.venv Python, network-none enforcement, ACL restoration, and CMD_HTTP_MCP_OK.

feat: revise sandbox usability around semantic policy and LPAC process trees

2026-08-11T20:44:54+00:00

Rewrite the specification and design vocabulary around User Sandbox Configuration, Environment Resolution, Executable Approval, Effective Sandbox Policy, and Windows Enforcement. Move the Internal Tool catalog to the design without treating its count as a requirement, and update README, implementation planning, traceability, audit history, prompts, and the Python skill. Introduce schema_version 1 profiles, workspaces, mounts, executable aliases, network modes, and resource presets. Add local override merging, resolution and approval stores with atomic user-only persistence, fail-closed fingerprints, localized diagnostics, and executable/sandbox management CLI commands. Change Windows execution to LPAC entry approval with inherited descendant confinement. Separate read and execute ACLs, preserve protected OS DACLs, apply Job process/memory/timeout limits to the tree, enforce WFP network modes, prevent credential environment inheritance, disable out-of-tree MSBuild reuse, and harden transactional ACL/recovery cleanup. Isolate planning prompts from execution instructions, reject split coding plans that cannot satisfy task-scoped evidence, preserve configured thinking, and prevent optional task_update failures from corrupting completion checks. Add and update configuration, resolution, approval, LPAC, descendant, readonly/development, Registry/COM, WFP, recovery, .NET/Python/PowerShell process-tree, and Operational E2E coverage. Publish reproducible bundles while preserving deployment-local configuration, prompts, skills, logs, state, approvals, and resolutions.

fix: implement native Ollama thinking semantics

2026-08-11T15:21:23+00:00

Model Ollama chat thinking separately from final content, preserve it only across assistant tool-call continuation, and keep private reasoning out of user results, session context, and logs. Accept native boolean and low/medium/high/max think values, normalize YAML booleans, and serialize explicit overrides using the native top-level chat field. Stop forcing think=false for structured planning, skill selection, and context summaries; corrected retries now retain the configured thinking mode. Update the specification, design, audit report, and requirement traceability for OLL-14. Add request/response, configuration, observability, retry, and tool-continuation coverage. Verified with 92 passing Agent.Core tests and a deepseek-v4-flash:cloud think:true MCP operational run.

Record final audit and conformance limits

2026-08-11T13:50:23+00:00

Audit and align v1 runtime design

2026-08-11T12:44:53+00:00

Implement P2/P3: Plan-Run flow separation, PlannerService, and Windows test fixes

2026-08-10T20:17:32+00:00

## P3: Session/Planning Foundation ### PlanAsync / RunAsync flow separation (Design §8.4) - AgentRuntime.ExecuteAsync now follows a two-phase lifecycle: 1. Planning: generate Initial TaskGraph via Ollama structured output 2. Running: execute ready Tasks (dependencies satisfied) in critical-path then creation order - Sandbox Task Resource (BeginTask/EndTask) is applied at RunAsync time, not PlanAsync time (Design constraint) - QueuedTaskWorker transitions to Planning; ExecuteAsync transitions to Running internally ### PlannerService (new: src/Agent.Core/Runtime/PlannerService.cs) - PlanAsync: uses Ollama ChatStructuredAsync with JSON Schema to generate a TaskGraph with per-Task requires_external_effect - ReplanAsync: rebuilds TaskGraph after policy deny, preserving completed tasks (Design §8.4) - Fallback: single-task plan with conservative estimates when structured output fails or returns empty - Pure analysis tasks get RequiresExternalEffect=false with at least 1 Task (Design constraint) ### PlanResult / PlanSchema (new: src/Agent.Core/Runtime/PlanResult.cs) - JSON model for LLM structured output during Planning - JSON Schema with tasks array: description, dependencies (by index), budget_class, PERT estimates, requires_external_effect ### requires_external_effect integration (P2 + P3) - Removed keyword heuristic RequiresExternalEffect() from AgentRuntime - Removed tool-name-counting RequiredToolCalls() from AgentRuntime - ToolUseGuard now receives PlannerTask.RequiresExternalEffect (set by Planner during Planning, not by keyword matching) - requires_external_effect is NOT always true; it is a per-Task attribute determined by the Planner (Design constraint) ### RunAsync / ready Task execution - SelectNextReadyTask: dependencies satisfied, not completed/failed, critical path first then creation order - Each PlannerTask executed via ExecutePlannerTaskAsync (LLM chat-with-tools iteration loop) - TaskGraph.AddUsage records per-Task token usage for CPM tracking ### Replan implementation - Policy deny (InvalidOperationException with "denied by sandbox policy") triggers ReplanAsync, up to max 3 replan attempts - Failed task is marked Failed in TaskGraph before replan - Registry transitions back to Running for replan iteration ## P3-A: AgentSession + TaskGraph ownership ### AgentSession expansion (src/Agent.Core/Sessions/AgentSession.cs) - Added Summary, ExecutionResults, TaskGraphs (ConcurrentDictionary) - TaskGraph is per-MCP-Task (1 agent_run = 1 TaskGraph), NOT shared per-Session (Design constraint) - Added CompactConversation, CompactExecutionResults, SetSummary methods for context summary compaction ### ExecutionResult (new: src/Agent.Core/Sessions/ExecutionResult.cs) - Records tool execution results in session for context summary - Fields: TaskId, ToolName, Success, Output ### SessionStore (src/Agent.Core/Sessions/SessionStore.cs) - Added Sessions property for TaskGraphToolHandler to resolve graphs by TaskId across sessions ### sessionId threading - IInternalToolBackend.BeginTaskAsync signature changed to include sessionId parameter - InternalToolDispatcher, WindowsToolBackend, TaskGraphToolHandler updated to thread sessionId through - task_update schema includes requires_external_effect property - McpHost and Program.cs pass SessionStore to TaskRegistry ## P3: Context Summary (Design §9.3) - Context utilization = latest prompt_eval_count / num_ctx (NOT cumulative InputTokens) (Design constraint) - TryContextSummary: compacts old Conversation/ExecutionResults when utilization exceeds 80% threshold - AgentRuntimeSettings: added NumCtx, ContextSummaryThreshold, ConversationKeepLastN, ExecutionResultsKeepLastN - OllamaResult: added PromptEvalCount field - OllamaClient: passes PromptEvalCount from OllamaChatResponse ## P3: ExecutionResults Session recording - ExecutePlannerTaskAsync records each tool execution result via session.AppendExecutionResult for context summary and audit ## Windows test fixes ### SandboxPolicyException catch (src/Agent.Windows/Tools/WindowsToolBackend.cs) - ExecuteAsync now catches SandboxPolicyException and returns ToolExecutionResult(false, ..., RetryAllowed: false) instead of propagating the exception - Fixes Dir_mkdir_outside_approved_root_is_denied and File_edit_rename_rejects_outside_approved_root tests ### LoopbackExemption cleanup (tests/Agent.Windows.Tests/SecurityPolicyTests.cs) - Recovery_after_crash: added pre-clean and try/finally to ensure LoopbackExemption is always removed, even on assertion failure - Independent_process_crash: added pre-clean to remove residual loopback exemption from prior test failures - Fixes 7 Independent_process_crash scenarios that were failing due to LoopbackExemption residue from Recovery_after_crash ## Test updates (tests/Agent.Core.Tests/PromptAndRuntimeTests.cs) - FakeOllamaClient implementations updated with ChatStructuredAsync and ChatWithToolsAsync overrides to support Plan/Run flow - RecordingOllamaClient: returns single-task plan JSON for ChatStructuredAsync, tracks all call types - ToolCallingOllamaClient: returns plan with requires_external_effect=true for file edit tasks - Test expectations updated for Plan+Run call counts and token accounting (Plan tokens not added to registry usage) ## Documentation updates - Design.md §7.7: requires_external_effect is set by Planner during Planning, not keyword heuristic - Design.md §8.4: Plan/Run phase separation, replan after policy deny - Design.md §9.3: context utilization formula clarified - ImplementationPlan.md §10.3: Plan/Run flow, replan, Context Summary ## Test results (all passing) - Agent.Core.Tests: 61/61 - Agent.Planner.Tests: 10/10 - Agent.Windows.Tests (elevated): 41/41 - Build: 0 errors, 0 warnings