Skip to content
/

Continue

continue · continuedev/continue · ★ 33k · last commit 2026-05-26

Primitive shape 12 total
Skills 7 Subagents 5
00

Summary

Continue — Summary

Continue has pivoted from a VS Code extension to a CI-first AI code review platform. The cn CLI runs agents on every pull request as GitHub status checks. Each agent is a markdown file in .continue/checks/ — green if the code looks good, red with a suggested diff if not. The .continue/ directory also holds agents (5 named agents for specialized reviews), rules (24 behavioral instruction files), and prompts. The VS Code extension remains available but is now secondary to the CI enforcement story. Continue's distinctive position: it moves AI code review from "optional chat assistant" to "enforceable CI check with a green/red badge."

Continue differs from all seed frameworks: it is not a development agent at all — it is a code review and CI enforcement layer. No seed framework plays this role. The closest seed is spec-kit (which has CI integration concepts) but spec-kit is about spec generation, not AI-as-PR-check. Continue's .continue/checks/ files are analogous to .github/workflows/ — they define automated quality gates that run on every commit.

01

Overview

Continue — Overview

Origin

Continue Dev, Inc. (continuedev) created the Continue VS Code extension as a coding assistant tool. The project has evolved significantly: what started as a chat-in-IDE assistant has pivoted to a CI-enforced AI review platform. The README now leads with the CI story: "Source-controlled AI checks, enforceable in CI."

Philosophy

From README:

"Source-controlled AI checks, enforceable in CI. Powered by the open-source Continue CLI"

The key insight: AI code review should be:

  1. Source-controlled: checks live in .continue/checks/ as markdown files, versioned in git
  2. Enforceable: checks run as GitHub status checks; PRs can be blocked on failed AI review
  3. Transparent: the check definitions are readable markdown, not opaque ML models

The Pivot

The project description changed from "AI coding assistant" to "Source-controlled AI checks." This is a major repositioning — from a developer productivity tool to a code quality enforcement layer.

Key Design Principles

  1. Checks as markdown: .continue/checks/*.md files define what the agent checks for. Simple YAML frontmatter + natural language instructions.
  2. CI integration: GitHub Actions integration; checks appear as status checks on PRs
  3. Green/red with diff: Failed checks produce a suggested diff, not just a comment
  4. Rules as behavioral constraints: .continue/rules/*.md define how the agent behaves (24 rule files for project-specific conventions)

VS Code Extension

Still available but now positioned as secondary. The cn CLI is the primary tool. The extension connects to the same agent engine.

02

Architecture

Continue — Architecture

Distribution

  • CLI: cn binary
    • macOS/Linux: curl -fsSL .../install.sh | bash
    • Windows: PowerShell install script
    • npm: npm i -g @continuedev/cli
  • VS Code Extension: Available in Marketplace
  • CI: GitHub Actions integration (runs cn in CI)

Install

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/continuedev/continue/main/extensions/cli/scripts/install.sh | bash

# Windows
irm https://raw.githubusercontent.com/continuedev/continue/main/extensions/cli/scripts/install.ps1 | iex

# npm
npm i -g @continuedev/cli

Directory Layout

.continue/
  checks/           # CI enforcement checks (7 defined in this repo)
    anti-slop.md
    react-best-practices.md
    security-audit.md
    setup-scripts.md
    stale-comments.md
    update-agents-md.md
    update-continue-docs.md
  agents/           # Named agents (5 defined in this repo)
    breaking-change-detector.md
    dependency-security-review.md
    error-message-quality.md
    input-validation.md
    test-coverage.md
  rules/            # Behavioral rules (24 defined in this repo)
    anti-slop-rule.md
    colors.md
    continue-specificity.md
    css-units.md
    ...
  prompts/          # Prompt templates
  environment.json  # Environment configuration
.claude/
  skills/           # (for Claude Code users)
core/               # TypeScript agent core
gui/                # VS Code webview UI
extensions/
  vscode/           # VS Code extension source
  cli/              # cn CLI source

Check File Format

---
name: Security Review
description: Review PR for basic security vulnerabilities
---
Review this PR and check that:
  - No secrets or API keys are hardcoded
  - All new API endpoints have input validation
  - Error responses use the standard error format

Required Runtime

  • Node.js >= 20 (for npm install)
  • No runtime for binary install

Target AI Tools

The cn CLI is provider-agnostic. Configuration in .continue/environment.json. Supports multiple LLM providers.

03

Components

Continue — Components

Checks (.continue/checks/)

CI-enforcement agents that run on every PR as GitHub status checks:

Check Purpose
anti-slop.md Fix AI-generated code patterns (verbose comments, over-abstraction, etc.)
react-best-practices.md Enforce React coding standards
security-audit.md Security vulnerability review
setup-scripts.md Validate setup script quality
stale-comments.md Remove stale/outdated comments
update-agents-md.md Keep AGENTS.md up to date
update-continue-docs.md Keep documentation current

Check format: YAML frontmatter (name, description) + natural language instructions.

Agents (.continue/agents/)

Named specialist agents for focused reviews:

Agent Purpose
breaking-change-detector.md Detect breaking API changes
dependency-security-review.md Audit dependency vulnerabilities
error-message-quality.md Review quality of error messages
input-validation.md Check input validation completeness
test-coverage.md Assess test coverage adequacy

Rules (.continue/rules/)

Behavioral constraints (24 files in this repo):

Rule Purpose
anti-slop-rule.md Anti-AI-slop patterns to avoid
colors.md Color system standards
continue-specificity.md Continue-specific patterns
css-units.md CSS unit conventions
documentation-standards.md Doc formatting standards
no-any-types.md TypeScript strict typing
overeager.md Prevent scope creep
personality.md Agent tone/style
programming-principles.md Core coding principles
pure-function-unit-tests.md Testing patterns
typescript-enum-usage.md TypeScript conventions
unit-testing-rules.md Testing standards
... (24 total) ...

CLI (cn)

The cn binary:

  • Runs checks on PRs (cn check <check-name>)
  • Lists available checks
  • Reports green/red with suggested diff

VS Code Extension

Chat panel for interactive AI assistance. Now secondary to CI use case.

Skills (.claude/skills/)

cn-check.zip and cn-check/ — skill for running Continue checks from Claude Code. Unique cross-tool integration.

05

Prompts

Continue — Prompts

Prompt 1: Security Audit Check

Source: .continue/checks/security-audit.md

Technique: Task-scoped review agent with explicit scope constraint ("do NOT look into anything unless it was changed in the pull request")

---
name: Security Audit
description: Security Audit
---

Please audit this pull request for any security vulnerabilities that were introduced. In particular you should look for new sources of potential prompt injections, or other vulnerabilities caused by the fact that this extension / CLI run locally on a user's machine. Do NOT look into anything unless it was changed in the pull request. When you are done, please make the required changes.

Note: The prompt is highly specific to Continue's own security concerns (prompt injection in an AI tool that runs locally).


Prompt 2: Anti-Slop Check

Source: .continue/checks/anti-slop.md

Technique: Anti-pattern enumeration — 10 specific patterns to detect and remove. Each pattern is concrete enough to be reliably detected.

---
name: Anti-slop
description: Fix AI Slop
---

I want to follow the **Anti AI-slop rule**: clean up any AI-generated code patterns that harm readability and maintainability...

**What qualifies as AI slop in code:**

1. **Overly verbose comments** - Comments that restate exactly what the code does (e.g., `// increment counter by 1` above `counter++`)
2. **Excessive defensive programming** - Unnecessary null checks, try-catches, or validations that clutter the logic
3. **Redundant type annotations** - Type declarations already inferred by the compiler
4. **Boilerplate explosion** - Separate classes/functions for trivial operations that could be a simple expression
5. **Over-abstraction** - Interfaces with single implementations, factories that create one thing
6. **Verbose variable names that obscure intent** - e.g., `currentUserAuthenticationStatusBoolean` instead of `isAuthenticated`
7. **Unnecessary intermediate variables** - Variables used exactly once on the next line
8. **Repetitive error handling** - Copy-pasted try-catch blocks that could be consolidated
9. **Filler documentation** - JSDoc that adds no information beyond the function signature
10. **"Just in case" code** - Unused parameters, dead code paths, or features for hypothetical future needs

**The goal:** Code should be concise, readable, and no more complex than necessary. Remove ceremony, not functionality.

Prompt 3: Agent Definition (input-validation.md)

Source: .continue/agents/input-validation.md

Technique: Named specialist agent with focused domain.

(Content not fully visible, but format follows YAML frontmatter + specialized instructions for checking input validation completeness across a PR diff.)


Prompt 4: Anti-Slop Rule (rules file)

Source: .continue/rules/anti-slop-rule.md

Technique: Persistent behavioral rule injected into all agent runs — the rule file version of the anti-slop check is injected as a constraint, not just a one-time check.


Prompting Techniques Used

  1. PR-diff scoping — all checks are scoped to "what changed in this PR" not "the whole codebase"
  2. Anti-pattern enumeration — exhaustive lists of specific patterns to detect
  3. Domain-specific agents — each agent/check has a single responsibility
  4. Make-required-changes pattern — checks don't just report; they "make the required changes" (action not just comment)
  5. Persistent rule injection — rules in .continue/rules/ are injected as constraints across all runs
  6. Markdown-as-CI — check files are version-controlled quality gates, not chatbot prompts
09

Uniqueness

Continue — Uniqueness and Positioning

Differs from Seeds

Continue is the only framework in both this batch and the seed frameworks that positions AI as a CI enforcement layer rather than a development assistant. No seed framework (superpowers, spec-kit, BMAD-method, openspec, taskmaster-ai, etc.) runs as a GitHub status check. The closest conceptual analog would be spec-kit's post-commit hooks, but spec-kit hooks run locally on the developer's machine while Continue checks run in CI and block merges. Continue's .continue/checks/*.md format is the first in this corpus to define AI behavior as version-controlled CI checks with green/red enforcement semantics.

Key Differentiators

  1. AI as CI gate: Checks run as GitHub status checks. Failed checks can block PRs from merging. No other framework in this corpus has this enforcement model.

  2. Anti-slop check: The most concrete "AI slop" detection framework found in this research — 10 specific patterns enumerated and detectable. This is notable because most frameworks suffer from AI slop in their own output; Continue explicitly checks for and removes it.

  3. Separation of checks/agents/rules: Three distinct file types with different roles:

    • Checks: CI enforcement (pass/fail with diff)
    • Agents: specialist reviewers (invoked by name)
    • Rules: behavioral constraints (always active)
  4. Stateless CI design: By design, no session state — each run is independent. This contrasts with all other frameworks that maintain conversational state.

  5. Claude Code cross-tool skill: The .claude/skills/cn-check integration bridges Continue's CI system into Claude Code sessions. This is the only framework in this batch that ships an integration skill for another tool in the same batch (Claude Code / superpowers territory).

  6. Anti-pattern enumeration style: The anti-slop check's 10-item numbered list is an unusually concrete prompting technique. Most prompts describe what to do; this one describes what NOT to do with specific named examples.

Observable Failure Modes

  1. Pivot confusion: The README description ("Source-controlled AI checks") doesn't match what most users expect from "Continue" (the VS Code extension)
  2. No interactive loop in CI: Failed checks require a new commit to re-run; slow feedback loop for minor issues
  3. Context limit on large PRs: The CI model doesn't handle large PRs well — the full diff + files may exceed context
  4. Rules files management: 24 rules files in this repo already; managing rule conflicts and priorities is not addressed
  5. VS Code extension is now secondary: Existing users of the VS Code extension may be confused by the pivot
04

Workflow

Continue — Workflow

CI Check Workflow (Primary)

  1. PR Opened — developer opens pull request
  2. GitHub Actions triggeredcn CLI runs in CI context
  3. Check execution — each .continue/checks/*.md file runs as a separate check
  4. Agent review — agent reads PR diff + relevant files; applies check instructions
  5. Pass / Fail — green badge if check passes; red badge with suggested diff if fails
  6. Developer response — developer can apply suggested diff or iterate

Interactive (Chat) Workflow

  1. VS Code / CLI — developer opens Continue chat
  2. Request — types request
  3. Response — agent responds with code suggestions or review

Check Development Workflow

  1. Add .continue/checks/my-check.md to project
  2. Define YAML frontmatter (name, description)
  3. Write natural language check instructions
  4. Commit to repo
  5. On next PR, check automatically runs in CI

Phases + Artifacts Table

Phase Artifact
CI check run GitHub status check (green/red)
Failed check Suggested diff
Agent review Review comment

Approval Gates

Gate Type Notes
PR check status yes-no GitHub branch protection can block merges on failed checks
No user-in-the-loop gates N/A CI runs autonomously

Continue is unique in this batch: there is NO interactive approval gate during the agent's work. The agent runs autonomously in CI and produces output. The "approval gate" is the PR review process itself.

Onboarding

From README: paste Help me write checks for this codebase: https://continue.dev/walkthrough into your coding agent of choice. The onboarding is itself a prompt that bootstraps the check system.

06

Memory Context

Continue — Memory and Context

State Storage

State Storage Scope
Check definitions .continue/checks/*.md Project, git-tracked
Agent definitions .continue/agents/*.md Project, git-tracked
Rules .continue/rules/*.md Project, git-tracked
Environment config .continue/environment.json Project

No Session State

Continue's CI model means there is no persistent session state between CI runs. Each PR check is a stateless invocation:

  1. Read PR diff
  2. Load relevant check instructions
  3. Run agent
  4. Report result
  5. Exit

Rules as Persistent Context

.continue/rules/*.md files are injected as persistent constraints into every agent run. This is the "memory" mechanism — not a database, but version-controlled behavioral constraints.

Cross-Session Handoff

None by design. Each CI run is independent. The .continue/ directory is the shared state.

Context Window Management

Not documented for the CI use case. Each check run likely includes:

  • PR diff
  • Relevant changed files
  • Check instructions
  • Active rules

For large PRs, this could exceed context limits. Not explicitly addressed in public docs.

07

Orchestration

Continue — Orchestration

Multi-Agent

The 5 named agents in .continue/agents/ are specialist reviewers. Whether they run in parallel or sequentially in CI is not explicitly documented.

Orchestration Pattern

Sequential (one check at a time in CI). Each check is a separate agent run.

Isolation

Each CI check run is isolated by the GitHub Actions environment — separate process, separate context.

Multi-Model

Not explicitly documented. Configuration in .continue/environment.json presumably sets the model.

Execution Mode

Event-driven: triggered by GitHub PR events (PR opened, PR updated).

Crash Recovery

Not applicable — stateless CI runs; a failed run can be re-triggered.

Context Compaction

Not documented.

Consensus Mechanism

None.

Prompt Chaining

Not used. Each check is independent.

No Interactive Loop

Unlike all other frameworks in this batch, Continue's CI mode has no interactive loop. The agent runs once per PR event, produces output, and exits.

08

Ui Cli Surface

Continue — UI and CLI Surface

CLI Binary

Name: cn
Install: npm i -g @continuedev/cli or curl install script
Is thin wrapper: No — own agent runtime

Usage

# Run a check
cn check security-review

# List available checks
cn list

# Run all checks
cn run

VS Code Extension

The Continue VS Code extension provides:

  • Chat panel for interactive AI assistance
  • Inline code completions
  • Context-aware suggestions
  • Connection to the same agent backend as cn

CI Integration

The primary "UI" is GitHub's status check panel:

  • Green checkmark: check passed
  • Red X: check failed, with suggested diff
  • Check name and description visible in PR

Observability

  • GitHub status checks panel (per PR)
  • CI logs (stdout from cn runs)
  • No dedicated audit log or dashboard

Cross-Tool

Continue ships a .claude/skills/cn-check.zip and cn-check/ directory — a Claude Code skill for running Continue checks from within Claude Code sessions. This is unique: it bridges the Continue CI system to the Claude Code agent system.

Related frameworks

same archetype · same primary tool · same memory type

Claude-Flow / Ruflo ★ 55k

Eliminates single-agent context limits and sequential bottlenecks by orchestrating fault-tolerant swarms of specialized AI agents…

Hermes Agent (NousResearch) ★ 168k

Self-improving personal AI agent with closed learning loop, 7 terminal backends, and messaging gateway — not tied to any AI…

OpenCode ★ 165k

Terminal-first AI coding agent with multi-model routing, native desktop app, and a typed .opencode/ configuration system for…

OpenHands ★ 75k

Open-source AI software development platform (open-source Devin alternative) with Docker sandbox isolation, 77.6% SWE-bench…

DeerFlow ★ 70k

Long-horizon superagent that researches, codes, and creates by orchestrating parallel sub-agents with isolated contexts in Docker…

oh-my-openagent (omo) ★ 60k

Multi-provider AI agent orchestration for OpenCode: escape vendor lock-in by routing Sisyphus (Claude/Kimi/GLM) and Hephaestus…