2026 · ACTIVE · UNKNOWN

CodeRelay

Models can fail. Your task should not.

TypeScript
VS Code Extension API
Node.js
Multi-provider LLM

CodeRelay is a VS Code extension that runs a coding task across multiple AI providers and keeps the task alive when one of them fails. It is not a chat panel; it is the reliability layer underneath one, with an interface built to show you the difference.

When a provider rate-limits, times out, drops a stream halfway or rejects a key, CodeRelay works out what had already happened on disk, hands a reconstructed account of the work to a different model, and carries on.

Problem

A long coding task driven by an AI provider dies for reasons that have nothing to do with the task: a rate limit, a timeout, a stream that drops halfway, a rejected key. The work already written to disk survives; the thread of intent does not.

Question

When a model fails mid-task, what does the next model actually need to know to continue — and can that be reconstructed from what happened on disk rather than from a chat transcript?

Approach

Treat the task as a ledger, not a conversation. Every operation is recorded; when a provider fails, CodeRelay reconstructs an account of the work from the ledger and hands it to a different model, skipping edits that already landed. The interface is a projection of the ledger rather than a commentary on it, so what you read is exactly what a resumed task would act on.

System

Open any stage to see what it does.

TASK — The objective, status, the model actually doing the work, elapsed time, steps, files changed, and token cost where the catalog declares prices.

LEDGER — The durable record of every operation. Not a chat log — the thing a resumed task reads.

TIMELINE — A projection of the ledger: one card per operation rather than one row per entry, large output collapsed, every file a link into the editor.

PROVIDER FAILURE — Rate limit, timeout, dropped stream, rejected key. Detected rather than surfaced as a crash.

RECONSTRUCT — Works out what already happened on disk and builds an account of it for a model that was not present for any of it.

RELAY — Hands the reconstructed state to a different provider and carries on — without re-running an edit that already landed.

Demo

Four views in the VS Code Activity Bar drive the whole task lifecycle without the Command Palette: task state, a streaming execution timeline, a composer, and the file changes.

Result

A task survives a provider failure and continues on a different model without re-applying edits that already landed.
The timeline is generated from the ledger, so the displayed history and the resumed state cannot disagree.
Cost and token accounting are shown when the configured model catalog declares prices.

Lesson

I built the chat panel first and it was the wrong object. The useful unit is not the conversation, it is the ledger of what actually changed on disk — once that existed, provider failover became a data problem instead of a prompt-engineering problem.

Source

CodeRelay — repository ↗

Connected

Epistemic Guard — Both are AI and SYSTEMS work, and both are built with TypeScript, VS Code Extension API, Node.js.
ContextGuard — Connected through AI and SYSTEMS.
CausalGuard — Connected through SYSTEMS and AI.
Bnlang for VS Code — Both are SYSTEMS work, and both are built with TypeScript, VS Code Extension API.