Architec

Incremental architecture review for AI-assisted codebases.

English | 中文

Architec is an advisory architecture analysis CLI. It helps answer one practical question:

Will this change make the codebase harder to maintain?

It reviews current changes by default, asks an LLM to interpret compact selected-scope evidence, and reports architecture risks such as duplicated logic, shadow implementations, unclear boundaries, stale structure, topology pressure, and risky hotspots.

Architec does not make merge decisions and does not edit code. It gives structured advice for humans and coding agents to review.

Why Architec

LLM-assisted development can move quickly, but architecture can drift quietly. Architec is designed to catch the kinds of issues that accumulate over time:

• repeated implementations and "same idea twice" code;
• compatibility paths that blur into canonical implementations;
• changed files crossing intended module boundaries;
• stale cleanup/archive candidates;
• high-risk work landing in churn-heavy areas;
• full-project topology pressure that is easy to miss during local edits.

The default workflow is incremental-first:

archi

Use full review when you want the whole-project baseline:

archi --full

How It Fits Together

Architec is the review layer. It uses two companion components:

Component	Command / package	Role
Architec	archi / architec	Runs architecture review, calls the LLM through llmgateway, writes advisory results under .architec/.
Hippo	hippo / hippocampus	Builds structural project snapshots under .hippocampus/: file manifests, code signatures, repository indexes, structure prompts, and metrics.
llmgateway	llmgateway	Owns provider credentials, base URLs, API style, model names, and strong/weak model routing.

source tree + git changes
        |
        v
Hippo structural snapshot  ->  .hippocampus/
        |
        v
Architec evidence builder  ->  selected-scope or full-project context
        |
        v
llmgateway LLM call        ->  strong / weak model tiers
        |
        v
Architec review output     ->  .architec/

Day-to-day archi runs still use the LLM, but they avoid refreshing the whole Hippo snapshot unless requested. archi --full uses the Hippo snapshot more heavily, and archi --refresh-from-hippo --full refreshes it before review.

How It Works

Architec combines deterministic code signals with LLM interpretation. The deterministic layer keeps the review grounded in concrete evidence; the LLM layer turns that evidence into readable architecture advice.

1. Select scope

   • archi reads the current git changes and focuses on changed files.
   • archi --full reviews the whole project.

2. Read structural context

   • Hippo produces .hippocampus/ snapshots: file manifests, code signatures, repository indexes, metrics, and structure prompts.
   • Architec checks whether that snapshot is present, stale, or unknown.

3. Build architecture evidence

   • Architec runs static scanners for duplicated logic, shadow implementations, import-boundary pressure, cleanup/archive candidates, hotspots, topology pressure, and snapshot freshness.
   • Incremental review keeps selected-change concerns separate from broader project context so small diffs are not drowned by global noise.

4. Ask the LLM for interpretation

   • Architec sends compact evidence to llmgateway.
   • llmgateway chooses the configured strong or weak model tier and owns all provider credentials.

5. Write advisory output

   • Architec ranks concerns, keeps raw artifacts for inspection, and writes human-readable plus machine-readable output under .architec/.
   • The result is advice, not an automatic merge decision or proof of runtime correctness.

Install

Architec requires Python 3.11+.

Recommended install from GitHub:

python3 -m pip install --user "architec @ git+https://github.com/SeemSeam/architec.git"

This installs:

• archi, the Architec CLI;
• llmgateway, the LLM provider gateway;
• hippocampus, the package that provides Hippo structural snapshots.

llmgateway and hippocampus are installed from the SeemSeam GitHub repositories, so no separate package index setup is required.

Local development from this repository:

python3 -m pip install -e .

Configure LLM Access

Architec gets all LLM access through llmgateway. Configure provider credentials and model tiers in:

~/.llmgateway/config.yaml

Minimal example:

version: 1
provider:
  provider_type: openai
  api_style: openai_responses
  base_url: https://your-llm-endpoint/v1
  api_key: sk-...
settings:
  strong_model: your-strong-model
  weak_model: your-fast-model

Architec consumes the configured strong_model and weak_model tiers. It does not store model-provider credentials itself.

Check the installation and LLM route:

archi --check .

If the check reports missing LLM configuration, update ~/.llmgateway/config.yaml.

Quick Start

Review the current selected changes:

archi

Run whole-project architecture review:

archi --full

Save JSON output:

archi --out review.json
archi --full --out full-review.json

Refresh Hippo inputs before full review:

archi --refresh-from-hippo --full

Command Summary

Command	Purpose
archi	Incremental LLM architecture review for current selected changes.
archi --full	Full-project LLM architecture review.
archi --out review.json	Save incremental review JSON.
archi --full --out full-review.json	Save full-review JSON.
archi --refresh-from-hippo --full	Refresh Hippo structural inputs, then run full review.
archi --check .	Validate Hippo bundle state and llmgateway configuration.

Advanced compatibility flags and older subcommands may still be accepted for existing automation, but new usage should prefer the commands above.

What Architec Reports

Architec reports advisory concerns and signals, including:

• Duplication: repeated logic and suspicious near-duplicates.
• Shadow implementations: second implementations of similar behavior.
• Architecture contracts: import-boundary or dependency-direction pressure.
• Cleanup/archive candidates: stale or legacy-looking code and docs.
• Hotspots: churn-heavy or structurally risky areas.
• Topology pressure: flat or confusing project structure.
• Snapshot freshness: missing, stale, or unknown Hippo context.
• Risk context: optional external facts attached to existing concerns.

The output is advisory. It is not a pass/fail result and is not proof of runtime correctness.

Outputs

Architec writes generated files under .architec/:

.architec/
  architec-analysis.json
  architec-summary.md
  architec-viz.html
  code-review-concerns.json
  code-review-discovery.json
  review-events.jsonl
  cache/

Hippo writes structural inputs under .hippocampus/.

Start with .architec/architec-summary.md for the human-readable report, then open .architec/architec-analysis.json for exact scores, concerns, signals, and artifact paths.

Agent Command Compatibility

The commands above describe the current public workflow. Some older installed archi binaries may still show the previous command shape, where full review is archi . and incremental review is archi --diff ..

Agents and automation should inspect the local binary before choosing commands:

archi --help

Help output	Incremental review	Full review
Includes --full	archi	archi --full
Lacks --full but includes --diff	archi --diff .	archi .

No Login Required

Architecture analysis does not require archi login.

Account commands such as archi login, archi whoami --json, and archi devices --json may exist for diagnostics, but they are not part of normal Architec analysis.

Development

Run tests:

PYTHONPATH=src python3 -m pytest -q

Run Architec from this checkout:

PYTHONPATH=src python3 -m architec
PYTHONPATH=src python3 -m architec --full

More Documentation

• Usage manual
• Architecture stability notes
• Evidence model