Skip to content
Kesavaraja67Public

About

Autonomous AI agent that detects breaking dependency changes via Tree-Sitter AST analysis, synthesizes verified patches, and opens human-reviewed PRs.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

17 stars

Watchers

0 watching

Forks

Repository files navigation

Telex Logo



T E L E X

Automated dependency self-healing and 3D architectural cartography for production codebases.

Watches package registries · Maps 3D codebase topology · Isolates AST call sites · Synthesizes verified patches · Never auto-merges without human review.


CI Status Test Coverage Ruff Code Style: Black 3D Repo Atlas License: MIT



1. Problem  •  2. What Telex Does  •  3. Target Audience  •  4. Real Example  •  5. Verification  •  6. Quickstart  •  7. Architecture  •  08. Risk & Review


Telex Automated Agent Avatar



TELEX AGENT CORE · Industrial Code Repair Engine & Dependency Reactor




Product Launch Film & Architectural Walkthrough

Watch the 75-second architectural walkthrough of Telex in action. This film demonstrates 24/7 package registry surveillance, 3D Repo Atlas topology rendering, Tree-Sitter AST call-site isolation, LLM patch synthesis, and ephemeral CI test sandboxes with ElevenLabs Liam voice narration:


Telex-new-video.mp4

Launch Video: Telex-new-video.mp4 (1080p HD · 75.0s · ElevenLabs Liam Narration)


▲ Back to Top



01. What Problem Does Telex Solve?

Every production application relies on dozens or hundreds of open-source packages. When an upstream dependency publishes a major or breaking release, existing dependency bots (Dependabot, Renovate) only bump the version string in package.json or pyproject.toml and open a pull request.

The outcome is always one of two failure modes:

  1. The PR immediately breaks CI: The test suite or typechecker fails because function signatures changed or methods were removed. A software engineer must stop feature work, read third-party changelogs, search the repository for all affected call sites, rewrite the code, and re-test.
  2. The PR quietly passes CI but ships latent defects: If test coverage doesn't assert the changed behavior, breaking semantic changes slip straight into production builds unnoticed.

Teams either freeze versions indefinitely—accumulating technical debt and security vulnerabilities—or spend hundreds of engineering hours manually performing mechanical dependency migrations.


▲ Back to Top



02. What Does Telex Actually Do?

Telex turns breaking dependency updates into verified, ready-to-merge GitHub pull requests that rewrite your call sites automatically.

[Upstream Registry] ──(24/7 Poll)──> [Changelog Parser]
                                             │
                                   (Breaking Symbol List)
                                             │
                                             ▼
[Your Repository] ──(Tree-Sitter AST)──> [Exact Call Sites]
                                             │
                                    (Call-Site Context)
                                             │
                                             ▼
                                     [LLM Patch Engine]
                                             │
                                       (Unified Diff)
                                             │
                                             ▼
                                   [Sandbox CI Verification]
                                             │
                                    (Passing Test Proof)
                                             │
                                             ▼
                                 [Human-Reviewed GitHub PR]

The 6-Stage Automated Lifecycle:

  1. Registry Surveillance: Continuously polls npm and PyPI registries to detect published versions and fetch release metadata.
  2. Breaking Change Extraction: Analyzes release notes and changelogs to identify removed, renamed, signature-changed, or behavior-changed API symbols with confidence ratings.
  3. AST Precision Scanning: Uses Tree-Sitter to parse your repository's ASTs and locate exact call sites, imported identifiers, and member expressions—without regex false positives.
  4. Targeted Patch Synthesis: Prompts enterprise LLMs (Gemini, Claude, GPT-4o, Groq) with the exact call site and migration contract to synthesize minimal, surgical unified diffs.
  5. Ephemeral Sandbox Verification: Validates candidate patches in isolated GitHub Actions environments against your repository's actual test suites and typecheckers.
  6. Human-Reviewed Pull Requests: Opens a single comprehensive PR per dependency upgrade containing full verification receipts, test logs, and risk classifications.

Strict Non-Negotiable Contract: Telex never automatically merges code. Every patch requires human engineering review and approval before entering production.

08. Risk Classification & Human-Review

  1. Mechanical vs. Semantic Risk Classification:
    • Upstream breaking changes are analyzed by Gemini 2.0 Flash to extract structured change types (removed, renamed, signature_change, deprecated, behavior_change) along with extraction confidence.
    • A change is classified as a semantic risk if it is a behavior_change (always flagged—passing tests do not guarantee old behavior is preserved), or if it is a signature_change or deprecated with classifier confidence < 75%. Changes of type removed or renamed are treated as mechanical.
    • When a semantic risk is identified, Telex tags the PR title with [semantic-risk] and flags in the PR body: ⚠️ Possible semantic/behavior change — passing tests do not guarantee old behavior is preserved.
  2. Fail-Closed Human Review Gate:
    • Whenever a patch has weak evidence—such as missing test coverage on the affected symbol, failing CI checks, or unverified semantic risks—Telex attaches the GitHub label needs-human-review and highlights the review requirement in the PR description.
  3. Supply-Chain Defense (Install Scripts Blocked by Default):
    • Untrusted dependency upgrades can execute malicious arbitrary lifecycle hooks (postinstall, preinstall). During isolated CI sandbox verification, Telex defaults to npm ci --ignore-scripts, pnpm install --frozen-lockfile --ignore-scripts, yarn install --frozen-lockfile --ignore-scripts, and disables Python install scripts unless repository maintainers explicitly toggle allow_install_scripts: true.

▲ Back to Top



03. Who Is Telex For?

Telex is built for:

  • Fast-Moving Product Teams: Avoid losing days of sprint velocity every time a critical framework or library (e.g., Pydantic v1→v2, Axios v0→v1, Next.js, FastAPI) introduces breaking changes.
  • Platform & Infrastructure Teams: Keep fleet-wide dependencies modern across dozens of microservices without manual patch coordination.
  • Maintainers of Mission-Critical Services: Protect systems with strict, evidence-bound verification where passing tests on the exact commit SHA are required before any code is approved.

Supported Ecosystems & Languages:

  • Automated Dependency Repair (4 Languages):
    • Node / TypeScript: .ts, .mts, .cts (package.json)
    • TSX: .tsx
    • JavaScript: .js, .mjs, .cjs
    • Python: .py (pyproject.toml, requirements.txt, setup.py)
  • Polyglot AST Call-Site Scanning & Cartography (11 Dialects):
    • Go: .go (go.mod, go.sum)
    • Rust: .rs (Cargo.toml, Cargo.lock)
    • Java: .java (pom.xml, build.gradle)
    • Ruby: .rb (Gemfile)
    • C# / .NET: .cs (.csproj, .sln)
  • Package Registry Surveillance:
    • Live Continuous Polling: npm, PyPI
    • Registry Hooks & Offline Manifests: crates.io, Go Proxy, Maven, RubyGems, NuGet
  • Repo Atlas 3D Visualization & Cartography:
    • Full import-graph cartography supports 21 languages including Go, Rust, Java, C/C++, Ruby, PHP, and more.

▲ Back to Top



04. Concrete Before/After Example

Scenario: Upgrading axios (v0.x → v1.x)

In Axios v1.x, the legacy CancelToken factory and source API was deprecated (while remaining backward-compatible) in favor of the standard Web API AbortController as an optional modernization.

Upstream Change Extracted by Telex:

{
  "change_type": "deprecated",
  "symbol_old": "axios.CancelToken.source",
  "symbol_new": "new AbortController()",
  "description": "CancelToken.source() deprecated in favor of native AbortController signal."
}

1. Target Repository Code (src/services/apiClient.ts):

import axios from 'axios';

export async function fetchUserData(userId: string) {
  const cancelSource = axios.CancelToken.source();
  const request = axios.get(`/users/${userId}`, {
    cancelToken: cancelSource.token,
  });
  return { request, cancel: () => cancelSource.cancel('User aborted') };
}

2. Tree-Sitter AST Call Site Detection:

Telex pinpoints lines 4–7 without string matching or false positives in comments:

Matched: CancelToken.source() at src/services/apiClient.ts:4:24
Matched: cancelToken property at src/services/apiClient.ts:6:5

3. Surgical Unified Diff Synthesized and Verified by Telex:

--- a/src/services/apiClient.ts
+++ b/src/services/apiClient.ts
@@ -3,6 +3,6 @@
 export async function fetchUserData(userId: string) {
-  const cancelSource = axios.CancelToken.source();
+  const controller = new AbortController();
   const request = axios.get(`/users/${userId}`, {
-    cancelToken: cancelSource.token,
+    signal: controller.signal,
   });
-  return { request, cancel: () => cancelSource.cancel('User aborted') };
+  return { request, cancel: () => controller.abort('User aborted') };
 }

4. The Resulting Pull Request:

Telex tests this diff against npm test and tsc --noEmit in an ephemeral sandbox. When all tests pass green, it opens a PR with the diff, test receipts, and a complete migration breakdown.


▲ Back to Top



05. How Verification Works

Telex enforces evidence-bound verification. A patch is never labeled verified merely because an LLM produced syntactically plausible code.

[Synthesized Diff]
       │
       ▼
[Pre-Flight Git Apply] ──(Fails)──> [Abort & Reject Patch]
       │
    (Passes)
       │
       ▼
[Ephemeral GitHub Actions Branch]
 (telex/validate-<patch_id>)
       │
       ▼
[Execute Project Test Suite]
 - Node: npm test / pnpm test / yarn test / tsc
 - Python: pytest / python -m unittest / mypy
 - Install scripts blocked by default (--ignore-scripts)
       │
       ▼
[Check Run Evaluation]
 - Must pass specifically on verification commit SHA
 - Skipped or neutral status = FAILS CLOSED (Not verified)
       │
       ▼
[Semantic Risk Classifier]
 - Marks behavioral changes with [semantic-risk]
 - Flags "needs-human-review" label when coverage is uncertain
       │
       ▼
[Delete Temporary Branch & Deliver PR]

Key Verification Invariants:

  1. Isolated Verification Branches: Candidate patches are tested on temporary branches (telex/validate-<id>), never directly on your default branch.
  2. Lifecycle Script Blocking: Package installation in verification sandboxes uses --ignore-scripts by default to prevent untrusted packages from executing arbitrary postinstall code.
  3. Commit SHA Binding: Validation receipts are cryptographically tied to the exact base_sha and commit_sha. If your default branch drifts before PR creation, Telex marks validation stale and revalidates.
  4. Fail-Closed Gate: If CI check runs are skipped, canceled, timed out, or missing, the patch is marked unverified.

▲ Back to Top



06. How to Try It in 5 Minutes

Prerequisites

  • Python: 3.11+
  • Node.js: 20+ (npm 10+)
  • PostgreSQL: 15+ (local or hosted via Neon / Supabase)
  • Git

1. Clone the Repository

git clone https://github.com/Kesavaraja67/telex.git
cd telex

2. Configure Backend (apps/api)

cd apps/api

# Create & activate virtual environment
python -m venv venv
# Windows (PowerShell):
.\venv\Scripts\Activate.ps1
# macOS / Linux:
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Configure environment variables
cp .env.example .env

Generate an encryption key for credentials:

python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"

Paste the generated key as TELEX_ENCRYPTION_KEY in apps/api/.env.

Apply database migrations:

alembic upgrade head

Start the FastAPI application and worker:

uvicorn main:app --reload --port 8000

API runs at http://localhost:8000. Health check available at http://localhost:8000/health.


3. Configure Frontend (apps/web)

In a second terminal window:

cd apps/web

# Install frontend dependencies
npm install

# Configure environment variables
cp .env.example .env.local

# Launch Next.js development server
npm run dev

Open your browser to http://localhost:3000.


4. Run the Test Suite

Verify your installation by running the backend test suite:

cd apps/api
pytest -v

▲ Back to Top



07. Architecture & Deeper Details

Comparison Matrix

Capability Naive Dependabot / Renovate Grep / Regex Scripts Telex Automated System
Upstream Registry Interception Periodic manifest scan None Continuous 24/7 npm & PyPI polling + breaking symbol extraction
Codebase Cartography None Flat directory tree Interactive 3D Repo Atlas with catenary wire physics
Blast Radius Mapping Blind version bump High false-positive grep Real-time 3D failure propagation & caller cascade
Call Site Resolution None Broken by strings & comments Tree-Sitter AST parser (TypeScript, TSX, JavaScript, Python)
Patch Generation Bumps version number only None Minimal surgical unified diff targeting exact call sites
CI Verification Fails in downstream CI Untested Pre-commit ephemeral sandbox running your actual test suite
Multi-Tenancy Fair Queue Basic sequential queue N/A PostgreSQL SKIP LOCKED with per-tenant fairness caps
LLM Flexibility Fixed proprietary bot None 10 LLM providers (BYOK Fernet-encrypted or hosted Gemini)
Human Review Governance Often configured to auto-merge Manual Zero auto-merge policy; comprehensive PR validation receipt

3D Repo Atlas Cartography & Blast Radius

Located at /dashboard/atlas, Repo Atlas is Telex's 3D spatial cartography engine. It turns complex AST relationships and import dependencies into an interactive spatial reactor room.


Telex Repo Atlas - 3D Macro Topology

Figure 1: Full-repository 3D architectural topology with polar force equilibrium and depth stratification.


  • Depth Stratification: $Y = -\text{depth} \times 4.8$, creating clean vertical parallax across directory tiers.
  • Deterministic Polar Layout: Ring coordinates with 2D force collision repulsion prevent label overlap.
  • Catenary Conduit Physics: Imports render as dielectric cables with gravitational sag and tension. Healthy cables render teal (#14B8A6); breaking connections pulse crimson (#F43F5E).
  • Live Incident Stream (SSE): Streams real-time breakage pulses to highlight impacted caller nodes over Server-Sent Events.
  • In-Canvas Source Inspection: Clicking any 3D node opens a slide-in code viewer displaying line-level AST references and commit metadata.
  • Incremental Update Engine: Webhook pushes trigger fine-grained graph updates (atlas_nodes / atlas_edges) fetching only changed files, with $O(1)$ reverse-dependency lookups (ix_atlas_edges_target).
  • Precision Camera Rig: Raycast cursor zoom, camera screen-plane panning (Right-drag, Space+drag, Middle-drag), and dynamic fog calibrated to graph bounding radius.

Telex Repo Atlas - In-Canvas Card Inspection

Figure 2: Slide-in AST card inspector showing line-level import references and breakage state.


Evidence-Based Run Analysis & Physical Instrumentation

Telex replaces arbitrary AI ratings with deterministic, fact-grounded resilience scoring. Every score is mathematically derived from verified AST facts and repository telemetry—the LLM writes prose explanations constrained strictly to computed facts and never invents a number.

  • Deterministic Sub-Scores (services/analysis_weights.py):
    • Structure (30%): Tarjan SCC circular cycle detection, longest DAG path depth, fan-in/fan-out hub files, orphan file ratios.
    • Dependencies (30%): Breaking change blast radius, packages behind major versions, unpatched symbols.
    • Change Safety (25%): Test file detection, automated CI workflow presence, churn on high-centrality files.
    • Verification (15%): Historical patch merge rate, sandbox CI pass rate, open review queues.
  • Physical Analog Dial Instrument (RunAnalysisCard.tsx):
    • Skeuomorphic instrument cluster with machined aluminum bezel and recessed meter face.
    • Rotating mechanical galvanometer needle calibrated from $-130^\circ$ (Score 0) to $+130^\circ$ (Score 100) with spring inertia.
    • 4 horizontal sub-score gauges with explicit "NOT MEASURED" states for unavailable signals.
    • Delta score badge vs prior runs and interactive findings with deep-links directly focusing nodes in 3D Repo Atlas.

Telemetry & Repository Fleet Management

Telex monitors multiple repositories across organizations with centralized policy management, automated telemetry summaries, and historical patch verification metrics.


Telex Multi-Repo Fleet Overview

Figure 3: Multi-repository fleet overview showing active package tracking, verification gates, and patch status.


Telex AST Code Scanner Radar

Figure 4: AST Code Scanner radar view displaying breaking symbol analysis and provider diff synthesis.


LLM Repair Providers & BYOK Architecture

Telex supports 10 enterprise LLM providers. You can bring your own API key (BYOK) or leverage the built-in Gemini platform service:

Provider Default Production Model Supported Context Encryption at Rest
Google Gemini (Platform Default) gemini-2.5-flash 1,000,000 tokens Managed / Hosted
Anthropic Claude claude-sonnet-4-5 200,000 tokens AES-128 Fernet
OpenAI gpt-4o-mini 128,000 tokens AES-128 Fernet
Mistral AI mistral-small-latest 128,000 tokens AES-128 Fernet
Groq (Llama 3.3) llama-3.3-70b-versatile 128,000 tokens AES-128 Fernet
DeepSeek deepseek-chat 64,000 tokens AES-128 Fernet
xAI Grok grok-3-mini 128,000 tokens AES-128 Fernet
Cohere command-r-plus-08-2024 128,000 tokens AES-128 Fernet
Together AI llama-3.3-70B-Instruct-Turbo 128,000 tokens AES-128 Fernet
Nvidia Nemotron llama-3.1-nemotron-70b-instruct 128,000 tokens AES-128 Fernet

Security & Cryptographic Model

  • Fernet Secret Encryption: BYOK API keys are encrypted at rest using AES-128 Fernet symmetric encryption. Keys are decrypted only in-memory during LLM execution and are never logged or echoed back.
  • Strict Log Scrubbing: A root logging filter automatically scans and redacts API keys (sk-*, AIza*, sk-ant-*) and high-entropy secrets from standard output and disk logs.
  • HMAC-SHA256 Webhook Verification: All GitHub App webhook payloads are verified against your secret signature (X-Hub-Signature-256) before task execution.
  • HttpOnly Cross-Origin JWT Sessions: Authentication uses secure HttpOnly, SameSite cookies to protect tokens from cross-site scripting (XSS).
  • Automated CI Security Scanners: Automated dependency auditing via pip-audit, npm audit, and secret detection via gitleaks.

Repository Layout

telex/
├── apps/
│   ├── api/                           # Backend FastAPI & Background Worker
│   │   ├── alembic/                   # Database migrations & schema version history
│   │   ├── db/models.py               # SQLAlchemy models (Repo, Patch, Atlas, etc.)
│   │   ├── jobs/handlers/             # Automated queue handlers
│   │   │   ├── poll_registry.py       # Watches npm & PyPI registries
│   │   │   ├── extract_changes.py     # Parses breaking symbol changes
│   │   │   ├── scan_repo.py           # Tree-Sitter AST code scanner (TS/JS/Py)
│   │   │   ├── generate_patch.py      # LLM diff synthesis engine
│   │   │   ├── validate_patch.py      # Ephemeral sandbox runner
│   │   │   ├── build_atlas_graph.py   # Computes 3D structural graph
│   │   │   └── open_pr.py             # Delivers human-reviewed PR
│   │   ├── jobs/queue.py              # PostgreSQL SKIP LOCKED fair queue
│   │   ├── routers/                   # REST API routes (auth, repos, atlas, stats)
│   │   ├── services/                  # Tree-Sitter, GitHub API, Crypto, LLM
│   │   └── tests/                     # Pytest suite with benchmark fixtures
│   │
│   └── web/                           # Frontend Next.js 16 Web Application
│       ├── app/
│       │   ├── page.tsx               # High-contrast monochrome landing page
│       │   └── dashboard/             # Operator control dashboard
│       │       ├── page.tsx           # Fleet telemetry overview
│       │       ├── repos/             # Repo list & policy toggles
│       │       ├── atlas/             # Dedicated 3D Repo Atlas view
│       │       ├── settings/          # BYOK key management
│       │       └── activity/          # Live event audit feed
│       ├── components/
│       │   ├── atlas/                 # Three.js 3D visualizer & layout engine
│       │   └── marketing/             # Interactive 3D bot, hero, marquee
│       └── public/                    # High-res screenshots, logo, and video assets
│           ├── telex-man1.png         # Agent mascot avatar
│           ├── Telex-new-video.mp4    # 1080p 75s launch & pipeline walkthrough video
│           ├── video-poster.jpg       # Video poster frame
│           └── repo-atlas-*.png       # 3D architecture screenshots
│
├── ARCHITECTURE.md                    # Deep-dive system architecture specification
├── DEMO.md                            # 15-minute evaluator demonstration guide
├── DESIGN.md                          # Design system & aesthetic doctrine
├── CONTRIBUTING.md                    # Contributor guide and pull request rules
├── CODE_OF_CONDUCT.md                 # Contributor covenant
├── SPEC.md                            # Executable specification & contracts
└── LICENSE                            # MIT License

▲ Back to Top



License & Community

Telex is distributed under the MIT License.


Architecture Spec  •  Evaluator Guide  •  Design System  •  Contributing  •  Code of Conduct


Built with precision engineering for production reliability.


Back to Top

About

Autonomous AI agent that detects breaking dependency changes via Tree-Sitter AST analysis, synthesizes verified patches, and opens human-reviewed PRs.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

17 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages