案例 / Scheduled Agent

用 90 行 TOML 构建 Standup Bot

一个无需你盯着也能运行的 autonomous agent，只使用 shell + git，并能跨 sessions 记忆。

standup_bot.toml

问题

Standup 很烦。你会忘记昨天做了什么。git history 知道，但你每天早上没时间回顾。即使写 standup notes，也常常泛泛而谈、没什么帮助。

如果 agent 能每天在同一时间自动替你完成这些，会怎样？

Config

整个 agent 定义在 standup_bot.toml 中 - 90 行 TOML。没有 Python。没有 JavaScript。除 QueryMT 本身外没有 runtime dependencies。

Agent definition

[agent]
provider = "anthropic"
model = "claude-sonnet-4-5-20250929"
assume_mutating = false
# No mutating tools — this agent is strictly read-only
mutating_tools = []
Read-only by design

assume_mutating = false and an empty mutating_tools list means this agent can never write to disk or execute mutating commands. It reads git history and writes to its knowledge store - nothing else.

Tool selection

[agent]
tools = [
  # Git & codebase analysis (read-only)
  "read_tool", "index", "glob", "search_text", "ls", "shell",

  # Knowledge lifecycle — the magic
  "knowledge_ingest",      # Store findings
  "knowledge_query",       # Recall past standups
  "knowledge_consolidate", # Merge daily → weekly
  "knowledge_list_unconsolidated",
  "knowledge_stats",

  # Task tracking
  "create_task", "todowrite", "todoread",
]
Knowledge lifecycle tools

Three tools handle persistent memory: knowledge_ingest stores structured findings, knowledge_query retrieves them, and knowledge_consolidate merges related entries into higher-level summaries. The agent remembers across sessions automatically.

Three-layer context compaction

# Layer 1: Tool output truncation
[agent.execution.tool_output]
max_lines = 1000
max_bytes = 25600

# Layer 2: Pruning after every turn
[agent.execution.pruning]
protect_tokens = 20000

# Layer 3: AI summary on context overflow
[agent.execution.compaction]
auto = true

Git log output can be huge. These three layers keep the context window manageable: truncate tool output at 1000 lines, prune old messages while protecting recent ones, and automatically summarize the conversation when it grows too large.

Middleware: execution limits

[[middleware]]
type = "limits"
max_steps = 30    # Max tool calls per cycle
max_turns = 15    # Max conversation turns

[[middleware]]
type = "context"
warn_at_percent = 80   # Warn when 80% of context used
compact_at_percent = 90 # Auto-compact at 90%
fallback_max_tokens = 128000
Cost-bounded execution

With max_steps = 30 and max_turns = 15 each scheduled cycle has a hard ceiling on cost. The agent can't run away. Context management ensures it never hits token limits unexpectedly.

设置 Schedule

启动 dashboard 后，通过 UI 创建 schedule：

  1. Open Session -> Schedules -> Create Schedule
  2. Prompt: "Run a standup cycle: analyze recent git activity, ingest findings, and report."
  3. Trigger: Interval, 86400 seconds (every 24 hours)
  4. Max runs: 30 (one month of daily standups)
  5. Max steps: 20
Automatic expiry

After 30 runs (one month), the schedule expires automatically. No orphaned processes. No cleanup needed.

会发生什么

每天同一时间，agent 会 autonomously 运行这个 cycle：

1. Git analysis Runs git log for the last 24 hours. Reads changed files, commit messages, and diff patterns.
→
2. Pattern extraction Identifies what areas were worked on, what issues appeared, what patterns emerged.
→
3. Knowledge ingest Stores structured findings: "API layer saw heavy refactoring, 3 new endpoints added, test coverage dropped."
→
4. Standup report Generates the day's standup: what was done, what's in progress, what's blocked.

Over time, the agent builds a knowledge graph of your project's history. Weekly consolidation merges daily standups into weekly summaries. Query past standups anytime:

# Ask the agent in the dashboard:
"What were the main themes from last week's standups?"

关键功能

  • Interval scheduling — runs on a fixed cadence, no user interaction needed
  • Knowledge lifecycle — ingest daily, consolidate weekly, query anytime
  • Autonomous execution — the bot does its job without you watching
  • Cost-bounded — max_steps and max_turns prevent runaway costs
  • Three-layer compaction — handles large git outputs gracefully
  • Automatic schedule expiry — no orphaned processes after the trial period

自己试试

# Clone and run
git clone https://github.com/querymt/querymt.git && cd querymt
cargo run --example qmtcode --features dashboard -- \
  confs/standup_bot.toml --dashboard

# In the dashboard, create a schedule:
#   Session -> Schedules -> Create Schedule
#   Prompt: "Run a standup cycle: analyze recent git activity, ingest findings, and report."
#   Trigger: Interval, 86400s (daily)
#   Max runs: 30