LARS b5d6aa89bb feat(auto-think): Automatic thinking level classification

Adds a new auto-think feature that automatically classifies incoming messages
and selects an appropriate thinking level using heuristics, without requiring
explicit /think directives.

Features:
- Heuristic-based classification (no API calls, <1ms latency)
- High complexity: debug, security, architecture, large code blocks
- Medium complexity: how-to, implementation, analysis, step-by-step
- Low complexity: simple lookups, definitions, translations
- Configurable floor/ceiling levels
- Custom pattern rules support
- User directives still override auto-classification

Config:
  agents:
    defaults:
      autoThink:
        enabled: true
        floor: 'off'
        ceiling: 'high'
        rules:
          - match: 'newsletter'
            level: 'medium'

Files changed:
- src/auto-reply/auto-think.ts: Classification logic
- src/auto-reply/auto-think.test.ts: Unit tests (25 tests)
- src/auto-reply/reply/get-reply-run.ts: Integration point
- src/config/zod-schema.agent-defaults.ts: Config schema
- src/config/types.agent-defaults.ts: TypeScript types
- docs/auto-think.md: Documentation

2026-01-28 03:47:01 +00:00

3.4 KiB

Raw Blame History

summary

read_when

Auto-think: Automatic thinking level classification based on message content

You want to enable automatic thinking level selection

You want to configure auto-think heuristics or rules

Auto-Think

Auto-think automatically classifies incoming messages and selects an appropriate thinking level without requiring explicit /think directives from users.

Overview

When enabled, auto-think analyzes each incoming message using heuristics to determine complexity:

High complexity: Debug requests, security reviews, architecture discussions, large code blocks
Medium complexity: How-to questions, implementation requests, step-by-step guides, comparisons
Low complexity: Simple lookups, definitions, format/translation requests
Minimal/Off: Short messages, greetings, unclassified content

Configuration

Enable auto-think in your agent config:

agents:
  defaults:
    autoThink:
      enabled: true

Full options

agents:
  defaults:
    autoThink:
      enabled: true
      floor: "off"        # Never go below this level
      ceiling: "high"     # Never go above this level
      rules:              # Custom patterns (optional)
        - match: "newsletter"
          level: "medium"
        - match: "security|audit"
          level: "high"

Config reference

Option	Type	Default	Description
`enabled`	boolean	`false`	Enable auto-think classification
`floor`	ThinkLevel	`"off"`	Minimum thinking level
`ceiling`	ThinkLevel	`"high"`	Maximum thinking level
`rules`	Array	`[]`	Custom pattern rules

Custom rules

Rules are evaluated in order; the first match wins. Each rule has:

match: A regex pattern (case-insensitive) or plain string
level: The thinking level to use when matched

Example:

rules:
  - match: "urgent|asap"
    level: "high"
  - match: "quick question"
    level: "off"

Override behavior

Auto-think respects the existing directive precedence:

Inline directive (/t high in the message) — always wins
Session sticky (previous /think:high directive-only message)
Auto-think classification — when enabled and no directive present
Default level (thinkingDefault config)

Users can always override auto-think with explicit /think directives.

Built-in heuristics

High complexity signals

Debug/debugging requests
Error messages, stack traces, exceptions
Security, vulnerability, audit keywords
Architecture, design pattern discussions
Refactoring, optimization requests
Code blocks > 500 characters

Medium complexity signals

"How do/would/should/can" questions
"Explain", "analyze", "compare" requests
Step-by-step guides
Implementation/build/create requests
Planning, strategy discussions
Code blocks > 100 characters
Messages > 2000 characters

Low complexity signals

Simple "What is X?" questions
Translation/conversion requests
List/enumerate requests
Definition lookups

Fallbacks

Very short messages (< 50 chars): off
Unclassified messages: minimal

Performance

Auto-think uses pure regex heuristics with no additional API calls. Classification adds negligible latency (< 1ms).

Thinking levels — Manual thinking control via /think directives
Token use — Cost implications of thinking levels

3.4 KiB Raw Blame History