Agentlas SEI

Find the gap between code that passes and products that work.

Agentlas SEI is a local-first assurance agent for existing software projects. It maps the product from large user journeys down to small observable steps, connects those steps to code and evidence, and shows where the team knows, assumes, or cannot yet prove that the product works.

It runs against a local project folder. GitHub is optional. Source code and interview answers stay on your machine by default.

Install and run with one prompt

Give this prompt to Codex, Claude Code, Gemini CLI, Antigravity, or another coding agent:

Install https://github.com/agentlas-ai/agentlas-sei and run SEI on this project.
Interview me first, map the large, middle, and small user flows, inspect evidence
gaps, then open the dashboard in English.

Or run it directly:

git clone https://github.com/agentlas-ai/agentlas-sei.git
cd agentlas-sei
./scripts/install.sh
sei /path/to/project

The first run interviews you about the product, maps the project, performs a deterministic inspection, writes local assurance state, and opens a tokenized dashboard on 127.0.0.1.

The problem

Most engineering tools answer narrow questions:

A normal check can show	It usually cannot prove
a test passed	the user completed the intended journey
no exception was thrown	a silent fallback did not hide failure
a function behaves correctly	the complete cross-screen flow works
documentation exists	the code still matches the documented intent
a metric changed	the change produced the right user outcome

This leaves a second defect surface beyond code bugs: stale assumptions, unowned decisions, missing observability, contradictory product intent, and flows that nobody can reconstruct end to end. SEI treats those as cognitive, intent, evidence, and governance debt—not as vague documentation problems.

How SEI works

flowchart LR
    A["Product interview"] --> B["Large user journeys"]
    B --> C["Middle processes"]
    C --> D["Small observable steps"]
    D --> E["Entry and exit code links"]
    E --> F["Evidence and contradiction checks"]
    F --> G["Local assurance dashboard"]

Loading

SEI compares five layers:

1. what the product intends;
2. what maintainers believe;
3. what the code implements;
4. what runtime evidence shows;
5. what users actually achieve.

The result is not a magical “all bugs found” claim. It is a reviewable map of claims, flows, evidence gaps, and issue candidates. Missing links remain marked as unknown instead of being guessed.

What you get

• a built-in product-intent interview;
• large → middle → small user-flow maps;
• entry-code and exit-code slots for every small flow;
• bounded project and code maps;
• deterministic technical and cognitive debt candidates;
• contradiction and evidence-gap detection;
• an append-only local ledger for findings, interviews, and change reviews;
• a multilingual local dashboard in Korean, English, Japanese, and Chinese;
• digest-bound review gates for high-risk commits and pushes;
• an optional, privacy-bounded OpenAI-compatible LLM pass.

SEI is read-only by default. v0.3.0 does not mutate application code, run production experiments, approve repairs, or deploy.

Dashboard

Open the dashboard in English:

sei dashboard /path/to/project --lang en

Available languages:

auto | ko | en | ja | zh

The dashboard presents the issue list and the selected issue side by side, then connects it to the relevant large, middle, and small flow. Verified relative code references can open in VS Code. Unverified links stay visibly unresolved.

The server binds only to 127.0.0.1, uses a random URL token, and serves no project files.

Installation

Python 3.11 or newer is required.

Shell

git clone https://github.com/agentlas-ai/agentlas-sei.git
cd agentlas-sei
./scripts/install.sh
sei /path/to/project

Codex

codex plugin marketplace add https://github.com/agentlas-ai/agentlas-sei.git
codex plugin add agentlas-sei@agentlas-sei

Invoke it with:

$sei

Claude Code, Gemini CLI, and Antigravity

The installer adds the local host adapters. Invoke them with:

/sei

Development install

python3 -m venv .venv
.venv/bin/python -m pip install -e .

Core commands

Goal	Command
Interview, inspect, and open the dashboard	sei /path/to/project
Open only the dashboard	sei dashboard /path/to/project --lang en
Create local SEI state	sei init /path/to/project
Refresh project and code maps	sei map /path/to/project
Run the product interview	sei interview /path/to/project
Inspect without an LLM	sei inspect /path/to/project
Inspect with an optional LLM	sei inspect /path/to/project --llm
Check state and privacy invariants	sei validate /path/to/project
Audit SEI itself	sei self-audit /path/to/agentlas-sei
Install commit and push gates	sei hooks install /path/to/project
Remove managed gates	sei hooks uninstall /path/to/project

Use sei <command> --help for command-specific options.

Local assurance state

SEI writes its working state inside the inspected project:

.sei/
├── config.json
├── boundary.json
├── status.json
├── maps/
│   ├── project-map.json
│   └── code-map.json
├── memory/interviews.jsonl
├── registry/
│   ├── claims.jsonl
│   └── flows.jsonl
├── evidence/evidence.jsonl
├── findings/findings.jsonl
├── decisions/change-reviews.jsonl
└── reports/latest-inspection.md

.sei/ is local assurance state and should not be published by default.

Risk-based change review

SEI does not interrupt every commit. Its managed hooks request a review when an exact change includes signals such as:

• broad or unusually large change surfaces;
• authentication, payment, schema, migration, release, or permission paths;
• fallback, retry, exception, rollback, or silent-failure behavior;
• new TODO, FIXME, HACK, or temporary markers;
• source changes without corresponding test or knowledge changes.

Install the hooks:

sei hooks install /path/to/project

When a gate requests review:

sei review /path/to/project \
  --stage pre-commit \
  --reviewer maintainer-name

The review is bound to the exact change digest. If the diff changes, the receipt becomes invalid. Existing hooks are never overwritten unless --force is supplied; forced installation creates a backup first.

Optional LLM

The deterministic inspection works without an LLM. LLM use is explicit:

export SEI_LLM_BASE_URL=http://localhost:11434/v1
export SEI_LLM_MODEL=your-model
sei inspect /path/to/project --llm

For a hosted OpenAI-compatible endpoint, also set:

export SEI_LLM_API_KEY=...

The LLM receives bounded counts, risk signals, claim IDs and states, observables, and finding summaries. Source code, file paths, and interview answers are excluded. The model may propose competing hypotheses; it cannot confirm a defect or approve a change.

Privacy and authority boundaries

• local folder first; no GitHub connection is required;
• no network call unless --llm is explicitly supplied;
• no symlink traversal;
• bounded file and byte scan budgets;
• known secret filenames are excluded;
• source content is analyzed in memory and is not copied into the ledger;
• interview answers are never sent to the LLM;
• the dashboard serves no source files;
• mutation and deployment remain outside the v0.3.0 authority boundary.

Current scope

v0.3.0 is an alpha production foundation. The local CLI, interview, maps, deterministic inspection, dashboard, validation contracts, and risk gates are implemented. Runtime telemetry adapters, outcome adapters, controlled experiments, and repair execution remain future work.

Apache-2.0. Maintained by Agentlas — appbridge@appbridge.co.kr.