vv-agent 0.1.57

Vector Vein inspired agent framework with cycle runtime, tools and memory management

vv-agent

中文文档

A lightweight agent framework extracted from VectorVein's production runtime. Cycle-based execution with pluggable LLM backends, tool dispatch, memory compression, and distributed scheduling.

Architecture

AgentRuntime
├── CycleRunner          # single LLM turn: context -> completion -> tool calls
├── ToolCallRunner       # tool dispatch, directive convergence (finish/wait_user/continue)
├── RuntimeHookManager   # before/after hooks for LLM, tool calls, memory compaction
├── MemoryManager        # automatic history compression when context exceeds threshold
└── ExecutionBackend     # cycle loop scheduling
    ├── InlineBackend    # synchronous (default)
    ├── ThreadBackend    # thread pool with futures
    └── CeleryBackend    # distributed, per-cycle Celery task dispatch

Core types live in vv_agent.types: AgentTask, AgentResult, Message, CycleRecord, ToolCall.

Task completion is tool-driven: the agent calls task_finish or ask_user to signal terminal states. No implicit "last message = answer" heuristics.

Setup

cp local_settings.example.py local_settings.py
# Fill in your API keys and endpoints in local_settings.py

uv sync --dev
uv run pytest

Quick Start

CLI

uv run vv-agent --prompt "Summarize this framework" --backend moonshot --model kimi-k2.5

# With per-cycle logging
uv run vv-agent --prompt "Summarize this framework" --backend moonshot --model kimi-k2.5 --verbose

CLI flags: --settings-file, --backend, --model, --verbose.

Programmatic

from vv_agent.config import build_openai_llm_from_local_settings
from vv_agent.runtime import AgentRuntime
from vv_agent.tools import build_default_registry
from vv_agent.types import AgentTask

llm, resolved = build_openai_llm_from_local_settings("local_settings.py", backend="moonshot", model="kimi-k2.5")
runtime = AgentRuntime(llm_client=llm, tool_registry=build_default_registry())

result = runtime.run(AgentTask(
    task_id="demo",
    model=resolved.model_id,
    system_prompt="You are a helpful assistant.",
    user_prompt="What is 1+1?",
))
print(result.status, result.final_answer)

SDK

from vv_agent.sdk import AgentSDKClient, AgentSDKOptions

client = AgentSDKClient(options=AgentSDKOptions(
    settings_file="local_settings.py",
    default_backend="moonshot",
    default_model="kimi-k2.5",
))
result = client.run("Explain Python's GIL in one sentence.")
print(result.final_answer)

SDK Workspace Override (Session/Task)

AgentSDKOptions.workspace is the SDK default workspace. You can override it per one-shot run, or bind a fixed workspace to a session.

Priority for workspace resolution is:

1. Explicit workspace passed to run(...) / query(...) / create_session(...)
2. AgentSDKOptions.workspace

from vv_agent.sdk import AgentSDKClient, AgentSDKOptions

client = AgentSDKClient(options=AgentSDKOptions(
    settings_file="local_settings.py",
    default_backend="moonshot",
    default_model="kimi-k2.5",
    workspace="./workspace/default",
))

# One-shot override: this run uses ./workspace/task-a
run = client.run(prompt="Create notes.md", workspace="./workspace/task-a")

# Session override: all turns in this session stay in ./workspace/session-b
session = client.create_session(workspace="./workspace/session-b")
session.prompt("Create todo.md")
session.follow_up("Append one more todo item")
session.continue_run()

Notes:

• AgentSession.workspace is fixed at session creation time.
• prompt()/continue_run()/follow_up() all execute in that same session workspace.
• session.cancel() requests cancellation for the currently running prompt in that session.
• Top-level SDK helpers vv_agent.sdk.run(...) and vv_agent.sdk.query(...) also accept workspace=....

Shell Runtime Configuration (Windows)

bash runtime defaults are a startup/session configuration, not tool-call arguments.

• Global defaults: AgentSDKOptions.bash_shell, AgentSDKOptions.windows_shell_priority, AgentSDKOptions.bash_env
• Per-agent override: AgentDefinition.bash_shell, AgentDefinition.windows_shell_priority, AgentDefinition.bash_env
• Recommended Windows priority: ["git-bash", "powershell", "cmd"]
• On Windows, bash-tool child processes default PYTHONUTF8=1 and PYTHONIOENCODING=utf-8 unless already overridden via the parent environment or bash_env.
• run(...) and create_session(...) both inherit startup shell defaults.
• The bash tool schema description includes a runtime shell hint (resolved shell kind + invocation prefix), so the model sees which shell command style is expected before calling the tool.
• The runtime shell hint is frozen per task/session-run to keep tool schemas stable across cycles and preserve LLM prompt cache efficiency.
• SDK/CLI-generated tasks now also attach structured system_prompt_sections metadata to the system message, so Anthropic prompt-cache breakpoints can keep the stable prompt prefix hot while treating current time and session-memory blocks as volatile.

from vv_agent.sdk import AgentDefinition, AgentSDKClient, AgentSDKOptions

client = AgentSDKClient(
    options=AgentSDKOptions(
        settings_file="local_settings.py",
        default_backend="moonshot",
        windows_shell_priority=["git-bash", "powershell", "cmd"],
        bash_env={"PIP_INDEX_URL": "https://pypi.tuna.tsinghua.edu.cn/simple"},
    ),
    agents={
        "desktop": AgentDefinition(
            description="Desktop helper",
            model="kimi-k2.5",
            # Optional hard override for this agent only:
            bash_shell=None,
            bash_env={"HTTP_PROXY": "http://127.0.0.1:7890"},
        )
    },
)

Execution Backends

The cycle loop is delegated to a pluggable ExecutionBackend.

Backend	Use case
InlineBackend	Default. Synchronous, single-process.
ThreadBackend	Thread pool. Non-blocking submit() returns a Future.
CeleryBackend	Distributed. Each cycle dispatched as an independent Celery task.

CeleryBackend

Two modes:

• Inline fallback (no RuntimeRecipe): cycles run in-process, same as InlineBackend.
• Distributed (with RuntimeRecipe): each cycle is a Celery task. Workers rebuild the AgentRuntime from the recipe and load state from a shared StateStore (SQLite or Redis).

from vv_agent.runtime.backends.celery import CeleryBackend, RuntimeRecipe, register_cycle_task

register_cycle_task(celery_app)

recipe = RuntimeRecipe(
    settings_file="local_settings.py",
    backend="moonshot",
    model="kimi-k2.5",
    workspace="./workspace",
)
backend = CeleryBackend(celery_app=app, state_store=store, runtime_recipe=recipe)
runtime = AgentRuntime(llm_client=llm, tool_registry=registry, execution_backend=backend)

Install celery extras: uv sync --extra celery.

Cancellation and Streaming

from vv_agent.runtime import CancellationToken, ExecutionContext

# Cancel from another thread
token = CancellationToken()
ctx = ExecutionContext(cancellation_token=token)
result = runtime.run(task, ctx=ctx)

# Stream LLM output token by token
ctx = ExecutionContext(stream_callback=lambda text: print(text, end=""))
result = runtime.run(task, ctx=ctx)

Runtime Log Payloads

tool_result runtime events carry full tool output in content and any structured tool payload in metadata (no implicit truncation of content). content_preview and assistant_preview are still emitted for UI convenience.

If you need shorter previews for logs/transport, configure an explicit preview limit:

from vv_agent.sdk import AgentSDKOptions

options = AgentSDKOptions(
    settings_file="local_settings.py",
    default_backend="moonshot",
    log_preview_chars=220,  # optional: enable preview truncation explicitly
)

Workspace Backends

Workspace file I/O is delegated to a pluggable WorkspaceBackend protocol. All built-in file tools (read_file, write_file, list_files, etc.) go through this abstraction.

list_files includes built-in safety defaults for large workspaces:

• Returns at most 500 paths per call by default (max_results can tune this, with hard cap).
• Uses ripgrep (rg) for fast local traversal when available, with automatic fallback to Python walk.
• workspace_grep also uses rg for local workspaces (with Python fallback), defaults to smart-case matching (lowercase patterns are case-insensitive; patterns with uppercase stay case-sensitive), and skips hidden/common dependency roots unless explicitly included.
• workspace_grep returns model-facing grep text in ToolExecutionResult.content, while structured matches/counts live in ToolExecutionResult.metadata.
• When listing from workspace root, common dependency/cache roots (for example node_modules, .venv, .git) are summarized instead of expanded.
• You can still inspect those paths explicitly by setting path to that directory (or by setting include_ignored=true).
• Supports scan_limit to stop early on very large trees; when triggered, response sets count_is_estimate=true.

Backend	Use case
LocalWorkspaceBackend	Default. Reads/writes to a local directory with path-escape protection.
MemoryWorkspaceBackend	Pure in-memory dict storage. Great for testing and sandboxed runs.
S3WorkspaceBackend	S3-compatible object storage (AWS S3, Aliyun OSS, MinIO, Cloudflare R2).

from vv_agent.workspace import LocalWorkspaceBackend, MemoryWorkspaceBackend

# Explicit local backend
runtime = AgentRuntime(
    llm_client=llm,
    tool_registry=registry,
    workspace_backend=LocalWorkspaceBackend(Path("./workspace")),
)

# In-memory backend for testing
runtime = AgentRuntime(
    llm_client=llm,
    tool_registry=registry,
    workspace_backend=MemoryWorkspaceBackend(),
)

S3WorkspaceBackend

Install the optional S3 dependency: uv pip install 'vv-agent[s3]'.

from vv_agent.workspace import S3WorkspaceBackend

backend = S3WorkspaceBackend(
    bucket="my-bucket",
    prefix="agent-workspace",
    endpoint_url="https://oss-cn-hangzhou.aliyuncs.com",  # or None for AWS
    aws_access_key_id="...",
    aws_secret_access_key="...",
    addressing_style="virtual",  # "path" for MinIO
)

Custom Backend

Implement the WorkspaceBackend protocol (8 methods) to plug in any storage:

from vv_agent.workspace import WorkspaceBackend

class MyBackend:
    def list_files(self, base: str, glob: str) -> list[str]: ...
    def read_text(self, path: str) -> str: ...
    def read_bytes(self, path: str) -> bytes: ...
    def write_text(self, path: str, content: str, *, append: bool = False) -> int: ...
    def file_info(self, path: str) -> FileInfo | None: ...
    def exists(self, path: str) -> bool: ...
    def is_file(self, path: str) -> bool: ...
    def mkdir(self, path: str) -> None: ...

Modules

Module	Description
vv_agent.runtime.AgentRuntime	Top-level state machine (completed / wait_user / max_cycles / failed)
vv_agent.runtime.CycleRunner	Single LLM turn and cycle record construction
vv_agent.runtime.ToolCallRunner	Tool execution with directive convergence
vv_agent.runtime.RuntimeHookManager	Hook dispatch (before/after LLM, tool call, memory compact)
vv_agent.runtime.StateStore	Checkpoint persistence protocol (InMemoryStateStore / SqliteStateStore / RedisStateStore)
vv_agent.memory.MemoryManager	Context compression when history exceeds threshold
vv_agent.workspace	Pluggable file storage: LocalWorkspaceBackend, MemoryWorkspaceBackend, S3WorkspaceBackend
vv_agent.tools	Built-in tools: workspace I/O, todo, bash, image, sub-agents, skills
vv_agent.sdk	High-level SDK: AgentSDKClient, AgentSession, AgentResourceLoader
vv_agent.skills	Agent Skills support (SKILL.md parsing, validation, unified normalization, prompt rendering with budget management, activate_skill tool)
vv_agent.llm.VVLlmClient	Unified LLM interface via vv-llm (endpoint rotation, retry, streaming)
vv_agent.config	Model/endpoint/key resolution from local_settings.py

Memory Compaction

MemoryManager now measures context size in tokens and compacts history when a model-derived auto-compaction threshold is exceeded.

• Task-level knobs:
  • memory_compact_threshold (default 128000, legacy fallback only when token counting is unavailable)
  • memory_threshold_percentage (warning threshold percentage, default 90)
• SDK mapping:
  • AgentDefinition.memory_compact_threshold
  • AgentDefinition.memory_threshold_percentage
  • AgentSDKClient.prepare_task(...) forwards both values to AgentTask.
• Token budget model:
  • effective_context_window = model_context_window - reserved_output_tokens
  • autocompact_threshold = effective_context_window - autocompact_buffer_tokens
  • Defaults come from vv-llm model metadata when available, otherwise fall back to 200000 / 16000 / 13000
• Effective-length strategy (backend-aligned):
  • If previous cycle token usage exists:
    • effective_length = previous_prompt_tokens + token_count(recent_tool_messages)
  • Otherwise fallback to:
    • vv_llm.chat_clients.utils.get_message_token_counts(...)
    • If tokenizer resolution fails, use a local CJK-aware estimate
• Compaction pipeline:
  1. Preemptive microcompact: clear old large tool results when usage crosses microcompact_trigger_ratio
  2. Session Memory extraction: persist key facts before full summarization so they survive later compactions
  3. Structural cleanup (stale tool calls, orphan tool messages, assistant-no-tool collapse, old tool result artifactization)
  4. If still over threshold, generate a compressed memory summary that preserves original user messages, file operations, current work state, and resolved errors
  5. If the provider still returns prompt-too-long, retry with forced compaction once, then progressively stronger emergency tail-dropping
  6. After full compaction, re-inject relevant workspace files into <Post-Compaction File Context> under a bounded token budget
• Session Memory behavior:
  • Stored in workspace/.memory/session/session_memory.json by default
  • Injected into the first system message on every cycle as <Session Memory>
  • Extraction reuses the configured memory summary backend/model
  • Full compaction resets transcript tracking but preserves persisted memory entries
  • Sub-tasks disable Session Memory by default to avoid parent/child memory-file contamination

Runtime metadata keys

Pass these via AgentTask.metadata:

• memory_keep_recent_messages
• model_context_window
• reserved_output_tokens
• autocompact_buffer_tokens
• microcompact_trigger_ratio
• microcompact_keep_recent_cycles
• microcompact_min_result_length
• microcompact_compactable_tools
• include_memory_warning
• session_memory_enabled / enable_session_memory
• session_memory_min_tokens
• session_memory_max_tokens
• session_memory_min_text_messages
• session_memory_storage_dir
• tool_result_compact_threshold
• tool_result_keep_last
• tool_result_excerpt_head
• tool_result_excerpt_tail
• tool_calls_keep_last
• assistant_no_tool_keep_last
• tool_result_artifact_dir
• summary_event_limit

Memory summary model selection priority

Priority is strict:

1. AgentTask.metadata
   • memory_summary_backend / memory_summary_model
   • aliases: compress_memory_summary_backend / compress_memory_summary_model
   • aliases: memory_compress_backend / memory_compress_model
2. local_settings.py constants
   • DEFAULT_USER_MEMORY_SUMMARIZE_BACKEND / DEFAULT_USER_MEMORY_SUMMARIZE_MODEL
   • aliases: DEFAULT_MEMORY_SUMMARIZE_BACKEND / DEFAULT_MEMORY_SUMMARIZE_MODEL
   • aliases: VV_AGENT_MEMORY_SUMMARY_BACKEND / VV_AGENT_MEMORY_SUMMARY_MODEL
3. Fallback
   • runtime default_backend + current task model

Built-in Tools

list_files, file_info, read_file, write_file, file_str_replace, workspace_grep, compress_memory, todo_write, task_finish, ask_user, bash, read_image, create_sub_task, sub_task_status.

Custom tools can be registered via ToolRegistry.register().

The bash tool supports two background paths:

• Explicit background: pass run_in_background=true, receive a session_id immediately, then poll with check_background_command.
• Timeout handoff: if a foreground command reaches timeout, it is moved into a background session instead of failing immediately. The tool returns a session_id, and the session emits terminal background-command events when that process completes, fails, or times out.

Sub-agents

Configure named sub-agents on AgentTask.sub_agents. The parent agent delegates work via create_sub_task: use agent_id to select the target sub-agent, task_description for one task, tasks for batch mode, and wait_for_completion=false to start background sub-tasks. Each sub-agent gets its own runtime, model, and tool set. The default system prompt automatically injects the callable sub-agent list, including each agent_id and description, so the model can choose directly.

Each delegated sub-task now runs in a real AgentSession (session id defaults to the sub-task id). Tool payloads include session_id, and runtime events include stable identifiers (task_id / session_id) so host apps can subscribe, persist, and stream sub-task progress independently (including sub_agent_stream_delta token chunks).

Batch mode in create_sub_task dispatches valid sub-task items through the runtime execution backend's parallel_map, so synchronous batches run concurrently when the backend supports parallel execution.

Use sub_task_status to query sub-task states, inspect lightweight progress snapshots (detail_level=snapshot), or send follow-up messages to running/completed sub-tasks. When you run agents through AgentSDKClient.create_session(), the sub-task registry stays attached to that session, so later turns can still query background sub-tasks created earlier in the same session.

Before a completed sub-task is resumed, the runtime now sanitizes the saved session transcript: empty assistant turns, thinking-only turns, orphaned tool results, and unresolved tail tool calls are removed so the next follow-up prompt resumes from a coherent history.

Sub-task runtime metadata now includes task_id, session_id, and browser_scope_key for each sub-agent run, so session-scoped tools (for example, browser controllers) stay isolated across parallel sub-tasks.

Host apps can interrupt a currently running sub-agent by calling vv_agent.runtime.engine.steer_sub_agent_session(session_id=..., prompt=...).

When a sub-agent uses a different model from the parent, the runtime needs settings_file and default_backend to resolve the LLM client.

Examples

24 numbered examples in examples/. See examples/README.md for the full list.

uv run python examples/01_quick_start.py
uv run python examples/24_workspace_backends.py

Testing

uv run pytest                              # unit tests (no network)
uv run ruff check .                        # lint
uv run ty check                            # type check

V_AGENT_RUN_LIVE_TESTS=1 uv run pytest -m live   # integration tests (needs real LLM)

Environment variables for live tests:

Variable	Default	Description
V_AGENT_LOCAL_SETTINGS	local_settings.py	Settings file path
V_AGENT_LIVE_BACKEND	moonshot	LLM backend
V_AGENT_LIVE_MODEL	kimi-k2.5	Model name
V_AGENT_ENABLE_BASE64_KEY_DECODE	-	Set 1 to enable base64 API key decoding