CodexFlow.QueryRuntime.Engine 0.1.0

CodexFlow QueryRuntime

English | 简体中文

A cross-platform, auditable, replayable, sandboxed agent runtime — for .NET.

CodexFlow QueryRuntime is a runtime harness for building AI agents. It is not another full coding-assistant app, and it does not require a web UI, accounts, a database, or a SaaS deployment. It extracts the infrastructure that every serious agent needs — model calls, tool calls, execution policy, trace, replay, sandboxing, and CLI automation — into a runtime you can embed, test, extend, and ship as a single native binary.

The implementation is still early/experimental, but already ships a working slice you can run today. This is a standalone repository split out from the original codexflow repo, focused only on the QueryRuntime suite.

asciinema cast

Why it exists

Most agent projects get stuck in the same place: the demo is easy, but turning it into testable, auditable, reproducible, safely-executable engineering infrastructure is hard. Common pain points:

• LLM providers differ wildly — tool calls, JSON schema, and "thinking" mode behave inconsistently across them.
• Tool execution has no boundary — reading files, writing files, running commands, and network access all get tangled together.
• When an agent run fails, there's no reproducible context — you're left guessing from terminal logs.
• CLI automation and an in-app embedded runtime end up as two separate codebases.
• The line between local execution, a Docker sandbox, and a Kubernetes runner is fuzzy.
• Native AOT publishing, a cross-platform CLI, and plugin loading are in natural tension.

QueryRuntime is the layer in between: more engineered than a 50-line agent demo, lighter than a full SaaS platform. Use it as the foundation for your own agent product, or just use the qre CLI as a tool for CI, codebase analysis, tool-call validation, and replay debugging.

Features

• Unified query/tool loop — collapses scattered round loops, tool execution, and termination logic into one reusable state machine (IQueryRuntimeEngine).
• Read-only tool pack — qre_list_files, qre_read_file, qre_search_files for analysis-style work.
• Verify tool pack — qre_git_status, qre_git_diff, qre_dotnet_build, qre_dotnet_test, executed under a capability policy in a trusted-local context.
• Trace / replay — every run writes a JSONL trace; replay latest defaults to a provider-free / tool-free recorded replay.
• Run artifacts — each run writes events.jsonl, manifest.json, run.json, diff.patch, usage.json, and artifacts/ under .qre/runs/<run-id>/. Large payloads spill to blobs/sha256/..., keeping only digest metadata in the trace.
• Capability policy — the profile (none / readonly / verify) decides which capabilities, commands, network, and mounts are allowed.
• Sandbox runners — LocalProcessSandboxRunner (trusted local dev) and DockerSandboxRunner (container isolation: read-only mount, network deny, non-root, dropped capabilities, seccomp, and more).
• External tool manifests — declare stdio or minimal mcp-stdio tools via .qre/tools/*.json, using an out-of-process, manifest-first design compatible with Native AOT.
• Lazy tool activation — opt into tool_search so the model starts with a small always-on tool surface, searches capabilities, and activates deferred tools for later rounds.
• Python function tools — Python projects can decorate ordinary functions, generate manifests, and register them as QRE tools without taking over the LLM tool-call loop.
• Machine-readable output — --json CLI output, qre trace latest --jsonl, and qre replay latest for scripts, CI, and third-party integration.
• Thinking policy — thinking is disabled by default when tools or JSON output are requested, improving tool-call and schema-output compatibility.
• Native AOT — qre ships as a single native binary (validated on osx-arm64).

Project layout

Runtime projects:

• CodexFlow.QueryRuntime.Engine — the unified query/tool loop execution engine.
• CodexFlow.QueryRuntime.Abstractions — Phase 1 stable contracts (runtime, model, tool registry, trace store, sandbox runner, CLI option DTOs).
• CodexFlow.QueryRuntime.Experimental — a lightweight wrapper and experimental harness.
• CodexFlow.QueryRuntime.Cli — the experimental qre CLI (main entry point).
• CodexFlow.QueryRuntime.Sandbox.LocalProcess — trusted-local ISandboxRunner.
• CodexFlow.QueryRuntime.Sandbox.Docker — Docker container isolation runner.

Test projects:

• CodexFlow.QueryRuntime.UnitTests
• CodexFlow.QueryRuntime.IntegrationTests

This repo intentionally does not include CodexFlow.Core. Core-side bridge coverage belongs in the original CodexFlow repo, with Core consuming QueryRuntime through adapters.

Embedding as a .NET library

Use CodexFlow.QueryRuntime.Abstractions.IQueryRuntimeHostEngine when another .NET application needs QRE to replace its existing in-process runtime. This contract accepts message history, custom AIFunction tools, required-tool steering, provider ChatOptions, trace/workspace paths, and streaming text deltas. The CLI can stay smaller; the host facade is the replacement surface.

Reference the library projects or packages that match the surface you need:

• CodexFlow.QueryRuntime.Abstractions for the stable host-facing contracts.
• CodexFlow.QueryRuntime.Experimental for the current ready-to-use harness, trace writer, built-in tool packs, and IChatClient adapter.
• CodexFlow.QueryRuntime.Models when the host wants QRE to construct a provider-specific IChatClient from endpoint/model settings.

For direct project references during local integration:

<ItemGroup>
  <ProjectReference Include="..\codexflow.queryruntime.engine\CodexFlow.QueryRuntime.Abstractions\CodexFlow.QueryRuntime.Abstractions.csproj" />
  <ProjectReference Include="..\codexflow.queryruntime.engine\CodexFlow.QueryRuntime.Experimental\CodexFlow.QueryRuntime.Experimental.csproj" />
  <ProjectReference Include="..\codexflow.queryruntime.engine\CodexFlow.QueryRuntime.Models\CodexFlow.QueryRuntime.Models.csproj" />
</ItemGroup>

Create the runtime from any Microsoft.Extensions.AI.IChatClient. The host can build the chat client itself, or use QreModelProviderSelector.CreateDefault() from CodexFlow.QueryRuntime.Models to select one of QRE's provider adapters.

using CodexFlow.QueryRuntime.Experimental;
using CodexFlow.QueryRuntime.Models;
using Microsoft.Extensions.AI;
using Qre = CodexFlow.QueryRuntime.Abstractions;

var chatClient = QreModelProviderSelector.CreateDefault().CreateClient(
    apiUrl: configuration["QRE_API_URL"]!,
    apiKey: configuration["QRE_API_KEY"]!,
    model: configuration["QRE_MODEL"]!,
    apiMode: configuration["QRE_API_MODE"]);

Qre.IQueryRuntimeHostEngine runtime =
    new ExperimentalQueryRuntimeHarness(
        new ChatClientExperimentalModelClient(chatClient));

Then call QRE with the host's existing conversation state and tool surface:

using CodexFlow.QueryRuntime.Experimental;
using Microsoft.Extensions.AI;
using Qre = CodexFlow.QueryRuntime.Abstractions;

var customTools = new[]
{
    AIFunctionFactory.Create(
        (string path) => repository.ReadContext(path),
        new AIFunctionFactoryOptions
        {
            Name = "repo_context",
            Description = "Read repository context for a relative path."
        })
};

var result = await runtime.RunAsync(
    new Qre.QueryRuntimeHostRequest
    {
        InitialMessages = history,
        WorkspacePath = workspacePath,
        RunId = runId,
        SessionId = sessionId,
        Tools = customTools,
        RequiredToolName = "repo_context",
        Execution = new Qre.QueryRuntimeExecutionOptions { MaxRounds = 4 },
        Options = chatOptions,
        TextDeltaSink = (delta, ct) => StreamToClientAsync(delta, ct)
    },
    ct);

Important request fields:

• InitialMessages is the host-owned multi-turn context. Use it instead of flattening prior turns into one prompt. When ToolSearch is enabled, QRE prepends its own small discovery system message and leaves the host messages intact.
• Tools exposes custom host tools to the model; set EnableTools = false to temporarily disable tools while keeping the tool list configured.
• ToolProfile can add QRE's built-in workspace tools (readonly, verify, repair) when the host wants them.
• RequiredToolName forces the model to call a specific tool before normal tool mode resumes. With ToolSearch enabled, that required tool remains visible in the first round so provider APIs receive a declared schema.
• Options passes provider options such as model id, temperature, response format, and provider-specific VllmChatOptions. QRE copies per-run options before injecting tools so host-owned options are not mutated.
• Output.RequestJson = true requests provider-level JSON output. Output.Json and Output.Stream are CLI formatting flags; library hosts should inspect QueryRuntimeResult and use TextDeltaSink.
• TextDeltaSink receives streamed assistant text deltas for the host UI or API.
• TimeProvider and QueryIdFactory are available for deterministic tests and replay-oriented host integrations.
• ToolSearch = new QueryRuntimeToolSearchOptions { Enabled = true } enables lazy activation for the host facade. Set TopK, AlwaysOnToolNames, or DeferredToolNames when the host needs tighter control over the initial tool surface.

For ASP.NET Core dependency injection, register the chat client and host engine behind the abstraction:

using CodexFlow.QueryRuntime.Experimental;
using CodexFlow.QueryRuntime.Models;
using Microsoft.Extensions.AI;
using Qre = CodexFlow.QueryRuntime.Abstractions;

builder.Services.AddSingleton<IChatClient>(sp =>
{
    var configuration = sp.GetRequiredService<IConfiguration>();
    return QreModelProviderSelector.CreateDefault().CreateClient(
        configuration["QRE_API_URL"]!,
        configuration["QRE_API_KEY"]!,
        configuration["QRE_MODEL"]!,
        configuration["QRE_API_MODE"]);
});

builder.Services.AddScoped<Qre.IQueryRuntimeHostEngine>(sp =>
    new ExperimentalQueryRuntimeHarness(
        new ChatClientExperimentalModelClient(sp.GetRequiredService<IChatClient>())));

Use the lower-level CodexFlow.QueryRuntime.Engine.IQueryRuntimeEngine only when the host needs to own event sinks and trace file creation directly. Use the qre CLI when you want a subprocess/native-binary workflow instead of an in-process library replacement.

See docs/IQueryRuntimeEngine.md for the full engine and facade contract. See docs/toolsearch.md for the lazy tool activation design.

Install

Option 1: Download a prebuilt binary (recommended)

Grab the single-file qre binary for your platform from GitHub Releases — no .NET SDK required:

# macOS (arm64) example
curl -L -o qre.tar.gz \
  https://github.com/iwaitu/codexflow.queryruntime.engine/releases/latest/download/qre-osx-arm64.tar.gz
tar -xzf qre.tar.gz
chmod +x qre
./qre --version

Supported targets: osx-arm64, osx-x64, linux-x64, linux-arm64, win-x64.

Option 2: Build from source

dotnet build CodexFlow.QueryRuntime.slnx
dotnet run --project CodexFlow.QueryRuntime.Cli -- --version

Local Native AOT publish:

dotnet publish CodexFlow.QueryRuntime.Cli -c Release -r osx-arm64 \
  -p:PublishAot=true -p:SelfContained=true
export PATH="$PWD/CodexFlow.QueryRuntime.Cli/bin/Release/net10.0/osx-arm64/publish:$PATH"
qre --version

Option 3: Build the NuGet library package

QRE's NuGet output is the core Engine library. The qre CLI is distributed through the native release artifacts; it is not packed as a .NET tool package by default.

scripts/qre-pack-nuget.sh Release

The package is written to artifacts/nuget:

• CodexFlow.QueryRuntime.Engine

Override the package version with QRE_PACKAGE_VERSION:

QRE_PACKAGE_VERSION=0.1.1 scripts/qre-pack-nuget.sh Release

Quickstart

1. Offline smoke (no LLM key needed)

Verify the CLI / trace / JSON output works end to end:

qre run --workspace . --response "offline smoke" --json "analyze this repo"

Emits a single qre.run.completed JSON object:

{"type":"qre.run.completed","finalText":"offline smoke","runId":"20260602145703992","termination":"NoToolCalls","profile":"none","tools":[],"traceFilePath":"./.qre/runs/20260602145703992/events.jsonl","totalRounds":1,"totalToolCalls":0,"totalDurationMs":52}

2. Read-only codebase analysis

qre run --workspace . --profile readonly --max-rounds 3 \
  "Find the most important runtime entry points and explain them."

For a smaller initial tool schema, enable lazy activation:

qre run --workspace . --profile readonly --tool-search --tool-search-top-k 3 \
  "Find the files that define the runtime engine."

With --tool-search, the first model round sees only tool_search. The search tool returns scored matches with risk, matched fields, required/optional args, and activation status; activated tools are injected on later rounds.

3. Inspect the trace and replay it

qre trace latest --workspace . --jsonl
qre replay latest --workspace . --json

replay latest defaults to a recorded replay: it reads the recorded model responses and tool results from the trace — it does not call the provider or re-execute the original tools.

Calling a real LLM provider

Risk: with a real provider, your prompt, the model context, and the file contents read by tools are sent to the endpoint you configure. Do not run real LLM analysis against sensitive private repos before evaluating the provider/proxy data policy.

Via environment variables

export QRE_API_URL="https://your-provider.example/v1"
export QRE_API_KEY="sk-..."
export QRE_MODEL="your-model"
export QRE_API_MODE="chat-completions"   # or responses / anthropic-messages

qre run --workspace . --profile readonly \
  "Summarize the repository architecture and list the top 3 risks."

Via command-line flags (equivalent)

qre run --workspace . \
  --api-url "https://your-provider.example/v1" \
  --api-key "$QRE_API_KEY" \
  --model "your-model" \
  --api-mode "chat-completions" \
  --profile readonly \
  "Summarize the repository architecture."

OpenAI-compatible endpoint (DashScope example)

export QRE_API_URL="https://dashscope.aliyuncs.com/compatible-mode/v1"
export QRE_API_KEY="sk-..."
export QRE_MODEL="deepseek-v4-pro"
export QRE_API_MODE="chat-completions"

qre run --workspace . --profile none --thinking off --json \
  "Output exactly this text and nothing else: OPENAI_COMPAT_OK"

Anthropic Messages-compatible endpoint

export QRE_API_URL="https://your-anthropic-compatible.example"
export QRE_API_KEY="sk-..."
export QRE_MODEL="your-claude-style-model"
export QRE_API_MODE="anthropic-messages"

qre run --workspace . --profile readonly --thinking off \
  "Explain the module boundaries of this repository."

--api-mode selects the provider factory's call style. Common values:

• chat-completions
• responses
• anthropic-messages

The CLI's real-provider path builds clients via QreVllmChatClientFactory, which recognizes model families (Qwen, OpenAI GPT, Gemini, Claude, Kimi, MiniMax, GLM, DeepSeek, …) from the model name. Unknown models currently fall back to a default client rather than a strict provider-neutral adapter.

Calling qre from a .NET app

See examples/RepoDoctor for a full, cross-platform example: it invokes qre as a local agent-runtime CLI, registers a custom .NET stdio tool, invokes that tool through qre tool invoke, streams a real-provider model answer to the host app console, and then follows up with a recorded replay. If QRE_API_URL / QRE_API_KEY / QRE_MODEL are not set, it reads the VllmAgent provider section from the sibling CodexFlow appsettings file.

cd examples/RepoDoctor
dotnet run -- /path/to/repo

Use dotnet run -- --offline /path/to/repo to validate the subprocess, streaming, trace, and replay path without an LLM key. Offline mode uses a static model response, so live provider mode is the path that exercises the required custom tool invocation.

Python live-provider tool-call streaming

examples/PythonToolDoctor shows the same binary-first integration style from Python. It reads provider settings from the sibling CodexFlow appsettings.json by default, streams a real LLM response, forces one qre_list_files tool call with --required-tool, and verifies the tool call from the recorded trace.

python examples/PythonToolDoctor/doctor.py /path/to/repo

Node.js live-provider tool demo

examples/NodeFunctionTools shows how Node.js functions become QRE stdio tools. The recorded demo generates Node manifests, registers node_count_files, invokes it through QRE, then streams a real LLM provider answer using that tool result.

python scripts/generate-node-tool-demo.py

Registering New Tools

QRE registers new out-of-process tools through workspace-local manifests under .qre/tools/*.json. See examples/ExternalTools for a minimal stdio tool, manifest, discovery command, and --required-tool smoke. For Python projects, examples/PythonFunctionTools shows how @qre_tool functions generate manifests and become QRE tools. For Node.js projects, examples/NodeFunctionTools shows the same flow using native ESM and explicit JSON schemas.

qre tool register --workspace . --manifest examples/ExternalTools/echo_tool.manifest.json
qre tool list --workspace . --profile readonly --external --json

python examples/PythonFunctionTools/repo_tools.py --manifest-dir .qre/generated-tools
qre tool register --workspace . --manifest .qre/generated-tools/py_count_files.json

node examples/NodeFunctionTools/repo_tools.mjs --manifest-dir .qre/generated-tools
qre tool register --workspace . --manifest .qre/generated-tools/node_count_files.json

Verify tools and capability policy

The verify profile adds controlled local command execution on top of the read-only tools:

qre run --workspace . --profile verify --max-rounds 4 \
  "Run the focused QueryRuntime tests and summarize failures."

Query the policy decision without executing the tool:

qre policy check --workspace . --profile verify \
  --tool qre_dotnet_test --json \
  -- dotnet test CodexFlow.QueryRuntime.slnx --no-restore

Run a policy-gated trusted-local command:

qre sandbox exec --workspace . --profile verify --json -- git status --short

The verify profile is still trusted local execution, not OS-level isolation. LocalProcessSandboxRunner does not actually block a child process from reaching the network or writing files — that requires a Docker/Kubernetes/VM runner to become a trusted execution boundary.

Test

dotnet test CodexFlow.QueryRuntime.UnitTests/CodexFlow.QueryRuntime.UnitTests.csproj

Docker sandbox integration tests are gated (require a local Docker daemon):

RUN_QUERY_RUNTIME_DOCKER_TESTS=true dotnet test \
  CodexFlow.QueryRuntime.IntegrationTests/CodexFlow.QueryRuntime.IntegrationTests.csproj \
  --filter "FullyQualifiedName~DockerSandboxRunnerIntegrationTests"

Gated real-provider integration tests:

RUN_QUERY_RUNTIME_REAL_INTEGRATION_TESTS=true dotnet test \
  CodexFlow.QueryRuntime.IntegrationTests/CodexFlow.QueryRuntime.IntegrationTests.csproj \
  --filter "FullyQualifiedName~ExperimentalHarnessRealLlmPhaseTests"

Use cases

• Local codebase analysis — let a model read repo structure, search files, and summarize architecture risks or migration suggestions.
• Agent tool-call validation — record model requests, responses, tool requests, and tool results as a single JSONL trace for regression testing.
• Read-only review in CI — --json output is script-consumable, ideal for offline smokes and read-only review.
• Teaching, evaluation, and replay — replay a provider-free / tool-free decision trajectory.
• A cross-platform agent product foundation — embed it in desktop apps, IDE plugins, CLIs, web backends, or enterprise platforms.

Security note

.qre/runs/<run-id>/events.jsonl can contain prompts, model responses, tool arguments, tool results, and the contents of files that were read (potentially private code or secret-shaped strings). The repo .gitignore already ignores .qre/. Add the same rule when dogfooding QueryRuntime in other repos, and redact before uploading .qre/ as a CI artifact.

Current limitations

This is a runtime harness still being refined — not yet a mature, secure execution platform:

• The CLI provider path still relies on QreVllmChatClientFactory's model-family heuristic routing, not a fully provider-neutral adapter.
• Replay supports recorded replay, but is not yet benchmark-grade deterministic replay (deterministic IDs, clock injection, and cross-version trace migration are still hardening items).
• Run-scoped diff.patch is scoped to paths edited by repair tools, but if one of those same files already had uncommitted changes before the run, the patch represents HEAD to the final file state, including that pre-existing same-file delta.
• The sandbox has a Docker runner, but Kubernetes / remote runners and a broader platform matrix are not done.
• Native AOT has a blocking linux-x64 CI lane and release packaging workflow, but signing and protected-branch required-check configuration are repository settings that must be verified before the first pre-release.
• usage.json is an estimate (ceil(chars / 4.0)), not a billing source of truth.
• mcp-stdio currently supports only a one-shot tools/call, without a full initialize lifecycle.

Documentation

• docs/queryruntime-technical-guide.md — technical guide (positioning, architecture, usage, roadmap).
• docs/IQueryRuntimeEngine.md — unified execution engine design.
• docs/toolsearch.md — lazy tool activation design.
• docs/queryruntime-harness-open-source-strategy.md — open-source harness strategy.
• docs/queryruntime-pre-release-work-plan.md — pre-release work plan (中文).
• docs/archive/queryruntime-next-development-plan.completed-2026-06-04.md — archived completed development plan.
• docs/queryruntime-tool-partition-matrix.md — tool partition matrix.
• docs/tool-capabilities.md, docs/threat-model.md.

License

MIT

In one sentence

CodexFlow QueryRuntime aims to be a cross-platform, auditable, replayable, sandboxable, embeddable agent runtime harness — making it easier to build your own coding agent, CI agent, IDE agent, or internal enterprise agent platform, instead of forcing you to adopt a full SaaS app.