inspect-scout

name	inspect-scout
description	Analyze AI agent transcripts using Inspect Scout scanners, grep patterns, and LLM-based analysis
user-invocable	false

Inspect Scout Reference

Inspect Scout is a library for analyzing AI agent transcripts with parallel processing and visualization capabilities. Use this when analyzing evaluation logs for patterns, behaviors, or specific events.

Installation: pip install inspect-scout

Source: https://github.com/meridianlabs-ai/inspect_scout

---

Core Concepts

Transcripts

A Transcript represents an LLM conversation to analyze (e.g., an agent rollout or sample from an Inspect eval). It contains:

class Transcript(TranscriptInfo):
    # Identity
    transcript_id: str          # Globally unique ID (e.g., sample uuid)
    source_type: str | None     # Type of source (e.g., "eval_log")
    source_id: str | None       # Source ID (e.g., eval_id)
    source_uri: str | None      # URI for source data (e.g., log file path)

    # Task info
    task_set: str | None        # Benchmark name
    task_id: str | None         # Dataset sample ID
    task_repeat: int | None     # Epoch number

    # Agent/model info
    agent: str | None           # Agent used
    model: str | None           # Main model
    model_options: dict | None  # Generation options

    # Results
    score: JsonValue | None     # Task score
    success: bool | None        # Success/failure boolean
    error: str | None           # Error message if failed
    limit: str | None           # Limit that caused exit (tokens, messages, etc.)

    # Metrics
    message_count: int | None
    total_time: float | None
    total_tokens: int | None

    # Content
    messages: list[ChatMessage]  # Conversation messages
    events: list[Event]          # Events from transcript
    metadata: dict[str, Any]     # Additional metadata

Scanners

Scanners are functions that analyze transcript content and return Result objects. Decorated with @scanner to specify what content to scan.

from inspect_scout import scanner, Scanner, Result

@scanner(messages="all")  # Scan all messages
def my_scanner() -> Scanner[Transcript]:
    async def scan(transcript: Transcript) -> Result:
        # Analyze transcript
        return Result(value=True, explanation="Found pattern")
    return scan

@scanner(messages=["assistant"])  # Only assistant messages
def assistant_scanner() -> Scanner[Transcript]:
    ...

@scanner(events=["tool"])  # Only tool events
def tool_scanner() -> Scanner[ToolEvent]:
    ...

@scanner(messages=["user", "assistant"], events=["tool", "error"])
def combined_scanner() -> Scanner[Transcript]:
    ...

Results

class Result(BaseModel):
    uuid: str | None            # Unique identifier
    value: JsonValue            # The scan result value
    answer: str | None          # Answer extracted from model output
    explanation: str | None     # Explanation of result
    metadata: dict | None       # Additional metadata
    references: list[Reference] # References to messages/events
    label: str | None           # Label for multi-result scanners
    type: str | None            # Type designation for value

---

Built-in Scanners

llm_scanner - LLM-based Analysis

Uses an LLM to analyze transcripts based on a question.

from inspect_scout import scanner, llm_scanner, Scanner
from inspect_scout._transcript.types import Transcript

@scanner(messages="all")
def did_agent_succeed() -> Scanner[Transcript]:
    return llm_scanner(
        question="Did the agent successfully complete its task?",
        answer="boolean"
    )

@scanner(messages="all")
def classify_failure() -> Scanner[Transcript]:
    return llm_scanner(
        question="Why did the agent fail?",
        answer=["gave_up", "hit_limit", "wrong_approach", "environment_error", "other"]
    )

@scanner(messages="all")
def rate_performance() -> Scanner[Transcript]:
    return llm_scanner(
        question="Rate the agent's performance from 1-10",
        answer="numeric"
    )

Answer types:

• "boolean" - True/False
• "numeric" - Number
• "string" - Free text
• list[str] - Single choice from labels
• AnswerMultiLabel - Multi-label classification
• AnswerStructured - Custom structured output

Additional options:

llm_scanner(
    question="...",
    answer="boolean",
    model="openai/gpt-4o",          # Model to use
    template="...",                  # Custom prompt template
    template_variables={...},        # Extra template vars
    preprocessor=...,                # Message preprocessing
    retry_refusals=3,                # Retry on refusals
    name="my_scanner"                # Scanner name
)

grep_scanner - Pattern Matching

Fast pattern-based scanning without LLM calls.

from inspect_scout import scanner, grep_scanner, Scanner
from inspect_scout._transcript.types import Transcript

@scanner(messages=["assistant"])
def find_secrets() -> Scanner[Transcript]:
    return grep_scanner(["password", "secret", "token", "api_key"])

@scanner(messages="all")
def find_urls() -> Scanner[Transcript]:
    return grep_scanner(r"https?://\S+", regex=True)

@scanner(messages="all", events=["tool"])
def categorize_errors() -> Scanner[Transcript]:
    return grep_scanner({
        "permission_denied": ["permission denied", "access denied"],
        "not_found": ["not found", "no such file"],
        "timeout": ["timeout", "timed out"],
    })

Options:

grep_scanner(
    pattern,                    # str, list[str], or dict[str, str|list[str]]
    regex=False,                # Treat as regex
    ignore_case=True,           # Case insensitive
    word_boundary=False,        # Match whole words only
    name="my_grep"              # Scanner name
)

Returns:

• Single pattern/list: Result with value=count, explanation=context snippets
• Dict patterns: list[Result] with one per label

---

CLI Usage

Running Scans

# Basic scan
scout scan scanner.py -T ./logs --model openai/gpt-4o

# With transcript filtering
scout scan scanner.py -T ./logs -F "task_set='cybench'" --model openai/gpt-4o

# Multiple scanners
scout scan scanners/ -T ./logs --model openai/gpt-4o

# Specific scanner from file
scout scan scanner.py::my_scanner -T ./logs --model openai/gpt-4o

Viewing Results

# Launch Scout View UI
scout view

# View specific scan results
scout view --scans ./scans/scan_id=abc123

# View remote results
scout view --scans s3://my-bucket/scans

# List scans
scout scan list

Managing Scans

# Resume interrupted scan
scout scan resume "scans/scan_id=iGEYSF6N7J3AoxzQmGgrZs"

# Mark incomplete scan as complete
scout scan complete "scan_id"

Parallelism Options

scout scan scanner.py -T ./logs \
    --max-transcripts 25 \    # Concurrent transcripts (default: 25)
    --max-connections 10 \    # Concurrent API requests
    --max-processes 4         # Parsing/scanning processes (default: 4)

---

Python API

Loading Transcripts

from inspect_scout import transcripts_from, columns as c

# From local eval logs
transcripts = transcripts_from("./logs")

# From S3
transcripts = transcripts_from("s3://bucket/logs")

# With filtering
transcripts = transcripts_from("./logs")
transcripts = transcripts.where(c.task_set == "cybench")
transcripts = transcripts.where(c.model == "gpt-4")
transcripts = transcripts.where("score > 0.5")  # SQL-style
transcripts = transcripts.limit(100)
transcripts = transcripts.shuffle(seed=42)

Running Scans Programmatically

from inspect_scout import scan, ScanJob, scanjob

# Single scanner
await scan(
    my_scanner(),
    transcripts="./logs",
    model="openai/gpt-4o"
)

# Multiple scanners via ScanJob
@scanjob
def analysis_job() -> ScanJob:
    return ScanJob(
        scanners=[
            did_agent_succeed(),
            classify_failure(),
            find_secrets(),
        ],
        transcripts="./logs",
        model="openai/gpt-4o"
    )

Accessing Results

from inspect_scout import scan_results_df, scan_status

# Check scan status
status = scan_status("scans/scan_id=abc123")
print(status.complete)  # True/False
print(status.summary)   # Counts, errors, etc.

# Get results as DataFrames
results = scan_results_df("scans/scan_id=abc123")

# Access by scanner name
df = results.scanners["my_scanner"]

# Available columns in result DataFrame:
# - transcript_id, source_id, source_uri
# - task_set, task_id, model, agent
# - score, success, error, limit
# - value, value_type, answer, explanation
# - message_references, event_references
# - scan_error, scan_total_tokens

# Row granularity options
results = scan_results_df("scans/...", rows="results")      # One row per result
results = scan_results_df("scans/...", rows="transcripts")  # One row per transcript

DuckDB Integration

from inspect_scout import scan_results_db

db = scan_results_db("scans/scan_id=abc123")
# Use db.connection for SQL queries

---

Project Configuration

Create scout.yaml for reusable configuration:

transcripts: ./logs
model: openai/gpt-4o

# Or with filters
transcripts:
  path: s3://bucket/logs
  where: "task_set = 'cybench'"

Then run: scout scan scanner.py (uses config automatically)

---

Helper Functions

messages_as_str

Convert messages to a string for LLM analysis:

from inspect_scout._scanner.extract import messages_as_str, MessagesPreprocessor

# Basic usage (in a scanner)
messages_str, extract_refs = await messages_as_str(transcript)

# With preprocessing
messages_str, extract_refs = await messages_as_str(
    transcript,
    preprocessor=MessagesPreprocessor(
        exclude_system=True,      # Remove system messages
        exclude_reasoning=True,   # Remove reasoning tokens
        exclude_tool_calls=True,  # Remove tool calls
    ),
    include_ids=True  # Include message IDs for references
)

---

Validation

Validate scanner results against ground truth:

# validation.csv
# id,target
# sample_123,true
# sample_456,false

from inspect_scout import scan

await scan(
    my_scanner(),
    transcripts="./logs",
    validation="validation.csv"
)

Results include validation_target, validation_result, and validation_predicate columns.

---

Common Patterns

Behavioral Analysis Scanner

@scanner(messages="all", events=["tool"])
def analyze_strategy() -> Scanner[Transcript]:
    return llm_scanner(
        question="""Analyze the agent's strategy:
        1. What approach did it take?
        2. Did it adapt when encountering obstacles?
        3. What tools did it use most?""",
        answer="string"
    )

Error Detection Scanner

@scanner(messages="all", events=["tool", "error"])
def detect_errors() -> Scanner[Transcript]:
    return grep_scanner({
        "command_errors": ["command not found", "permission denied", "no such file"],
        "api_errors": ["rate limit", "timeout", "connection refused"],
        "logic_errors": ["assertion failed", "unexpected", "invalid"],
    })

Success Criteria Scanner

@scanner(messages="all")
def verify_success() -> Scanner[Transcript]:
    return llm_scanner(
        question="Did the agent retrieve the flag and output it correctly?",
        answer="boolean"
    )

Multi-label Classification

from inspect_scout._llm_scanner import AnswerMultiLabel

@scanner(messages="all")
def classify_behaviors() -> Scanner[Transcript]:
    return llm_scanner(
        question="What behaviors did the agent exhibit?",
        answer=AnswerMultiLabel(labels=[
            "exploration",
            "exploitation",
            "backtracking",
            "tool_chaining",
            "giving_up"
        ])
    )