map-inspect-packages

name	map-inspect-packages
description	Use whenever the user is working with Inspect AI or any of its ecosystem packages (Evals, Flow, Scout, Viz, SWE, Harbor, Sandboxes). Maps concerns to packages and points at each package's docs before calling its API.

Inspect AI Ecosystem

Inspect AI is an open-source Python framework for LLM evaluations. The core (inspect_ai) was originally developed by the UK AI Security Institute (AISI) and is now primarily maintained by Meridian Labs. inspect_evals is co-maintained by AISI and outside contributors; the remaining ecosystem packages (Flow, Scout, Viz, SWE, Harbor) are maintained by Meridian Labs. Each package owns a distinct concern. Pick the right one before writing code; fetch the package's llms.txt (or docs root) before calling its API or writing more than a few lines against it.

Defer to a more specific skill if one exists. If another installed skill's description more specifically matches the request (e.g. one dedicated to log reading, log mutation, eval writing, etc.), use that one. This skill covers ecosystem routing only.

Package	Owns	Docs
Inspect AI	Core eval framework: Task / Solver / Scorer / datasets / models and providers / agents / tools / sandbox API / log and analysis APIs (inspect_ai.log, inspect_ai.analysis) / eval runners (eval(), eval_set()).	https://inspect.aisi.org.uk/llms.txt
Inspect Evals	Community catalog of 130+ pre-built benchmark evals (GAIA, SWE-Bench, Cybench, MMLU, GPQA, …). Use these before writing your own.	https://ukgovernmentbeis.github.io/inspect_evals/llms.txt
Inspect Flow	Declarative workflow orchestration: run many tasks × models × params at scale, with sweeps, defaults, and post-eval steps (FlowSpec, FlowTask).	https://meridianlabs-ai.github.io/inspect_flow/llms.txt
Inspect Scout	In-depth analysis of AI agent transcripts via LLM, grep/pattern, or multi-agent scanners. Detects issues like refusals, misconfiguration, or evaluation awareness. Imports natively from Inspect logs, plus Arize Phoenix, LangSmith, Logfire, MLFlow, W&B Weave, Claude Code sessions, and arbitrary Arrow/Parquet/custom Python sources.	https://meridianlabs-ai.github.io/inspect_scout/llms.txt
Inspect Viz	Data visualization for Inspect log dataframes: bar charts, line and scatter plots, scorer & sample heatmaps, radar plots, with interactive filtering and tables.	https://meridianlabs-ai.github.io/inspect_viz/llms.txt
Inspect SWE	Software-engineering agent scaffolds (Claude Code, Codex CLI) usable as Inspect solvers.	https://meridianlabs-ai.github.io/inspect_swe/llms.txt
Sandboxes	Isolated environments for sample execution. Docker is built into Inspect AI for local execution. Beyond that, several external sandbox packages exist: managed cloud (e.g. Inspect Sandboxes with Daytona/Modal) and self-hosted (K8s, Proxmox, EC2, …).	Inspect Sandboxes · k8s · proxmox · ec2
Inspect Harbor	Bridge to Harbor containerized agent benchmarks (e.g. Terminal Bench) as Inspect tasks.	https://meridianlabs-ai.github.io/inspect_harbor/llms.txt

The table above is curated for the most-used packages. For the comprehensive ecosystem catalog (25+ extensions including community projects from METR, Redwood, Transluce, Vector Institute, and more), fetch Inspect Extensions.

Picking the right package

• "Write/run an eval against a model" → Inspect AI. First check Inspect Evals, the Inspect AI evals catalog (cross-ecosystem index across 12 categories), and Inspect Harbor's registry to see if the benchmark already exists.
• "Simple sweep across models or tasks (one-off)" → Inspect AI's eval_set() (built-in, supports multi-model and multi-task runs directly).
• "Reusable / declarative orchestration: config-driven runs, post-eval workflows, or re-running the same setup later" → Inspect Flow (rather than a hand-rolled Python wrapper around eval_set()).
• "Identify patterns across many samples, or analyze long agent trajectories" → Inspect Scout.
• "Plot or dashboard eval results" → Inspect Viz.
• "Evaluate a coding agent" → Inspect SWE for the agent, Inspect Evals (SWE-Bench, etc.) or Inspect Harbor for the benchmark.
• "Sample needs an isolated environment" → Inspect AI sandbox API. Default to the built-in Docker sandbox for development and small/test runs; reach for an external sandbox package (managed cloud or self-hosted K8s/Proxmox/EC2) for large-scale runs, resource-intensive tasks, or GPU needs.
• "Read or inspect a single eval log or a handful of samples" → Inspect AI (inspect_ai.log for one log; inspect_ai.analysis for cross-log dataframes). Use this rather than Scout when you're looking at one log or a small set.
• "Add a custom scorer, solver, tool, or dataset" → Inspect AI core.

Before calling the API

ALWAYS fetch and consult the relevant llms.txt (or docs root) before answering any question or writing any code about these packages, even when you believe you know the API. The descriptions above are routing hints, not specs; the docs are the source of truth and change frequently. Prior-knowledge answers drift quickly and may reference renamed, moved, or removed APIs.

Keep packages aligned with the docs. Check the user's installed version (pip show <package>) against the latest on PyPI. If it's older, propose upgrading (pip install -U <package>) so the features in the docs are actually available at runtime. If the user can't upgrade (pinned deps, monorepo constraints), flag the version gap.

Tip: markdown variants for agent-readable docs. HTML versions of these docs sites are JS-rendered and may return empty content to fetch tools; use the markdown variant instead.

• inspect.aisi.org.uk pages: append .html.md (e.g. https://inspect.aisi.org.uk/agents.html.md).
• ukgovernmentbeis.github.io/inspect_evals pages: append .llms.md (e.g. https://ukgovernmentbeis.github.io/inspect_evals/evals/gpqa/index.llms.md).

---

Last reviewed: 2026-06-15 against inspect_ai v0.3.235