Add Atomic Chat as a first-class local provider.

[yanalialiuk]

Summary

Ship the Atomic Chat plugin with auto-discovery on localhost:1337, register the provider in models.dev fixtures, and document setup for VS Code and CLI.

Context

Atomic Chat runs local models and exposes an OpenAI-compatible API at http://127.0.0.1:1337/v1 (alternate port 1338). Kilo Code already supports local backends like LM Studio via @ai-sdk/openai-compatible, but users had to configure baseURL and model IDs manually, and Atomic Chat did not appear as a first-class provider with discovery or loaded-model checks.

This PR adds Atomic Chat as a built-in local provider: default plugin, catalog entry, documentation, icon, and VS Code migration mapping—so users can pick Atomic Chat in the UI and get models populated automatically when the app is running.

Implementation

@kilocode/plugin-atomic-chat (new package)

• Auto-discovery: If provider.atomic-chat.options.baseURL is unset or still the default origin (http://127.0.0.1:1337), probe ports 1337 and 1338 with a health check, then call GET /v1/models.
• Config hooks: Merge discovered models into provider.atomic-chat, set options.baseURL, and show success/error toasts.
• chat.params hook: Validate the selected model is loaded (with cache/retry) before sending a request; surface a clear error if not.
• 12 Vitest tests covering discovery, config enhancement, and validation paths.

Default plugin registration

• atomic-chat-feature.ts — resolve and dedupe @kilocode/plugin-atomic-chat.
• KilocodeDefaultPlugins — append the plugin alongside the indexing plugin (same builtin-origin pattern).

Catalog, UI, migration

• Add atomic-chat to packages/opencode/test/tool/fixtures/models-api.json (@ai-sdk/openai-compatible, env ATOMIC_CHAT_API_KEY, default API http://127.0.0.1:1337/v1).
• Provider icon + sprite; legacy migration atomicChat → atomic-chat.
• Docs: pages/ai-providers/atomic-chat.md, nav, and local-models page (Atomic Chat listed first under self-hosted).

Notes for reviewers

• models-snapshot.js remains gitignored; the fixture is the source of truth for tests (regenerate locally via packages/opencode/script/generate.ts if needed).
• Fixture models are placeholders; runtime discovery from the live API is authoritative when the plugin runs.
• Optional follow-up: upstream atomic-chat on models.dev for live catalog fetches outside the bundled snapshot path.

How to Test

Prerequisites

1. Install Atomic Chat and enable the local API (default port 1337).
2. Load a model in the app.
3. Confirm the API responds:

curl http://127.0.0.1:1337/v1/models

Automated

cd packages/plugin-atomic-chat && bun test
cd packages/opencode && bun test test/kilocode/config/atomic-chat-default-plugin.test.ts

Manual — VS Code

-Build/run this branch of the Kilo Code extension.
-Open Settings → Providers → Atomic Chat.
-Confirm models appear (from discovery) and chat works with a loaded model id from /v1/models.
-Stop or unload the model in Atomic Chat and retry chat — expect a validation error or toast.

Manual — CLI

In ~/.config/kilo/kilo.jsonc (or ./kilo.jsonc):

{
  "provider": {
    "atomic-chat": {
      "options": {
        "baseURL": "http://127.0.0.1:1337/v1"
      }
    }
  },
  "model": "atomic-chat/<id-from-v1-models>"
}

Run the Kilo CLI; confirm the default plugin discovers the server without manually adding @kilocode/plugin-atomic-chat.

Edge cases

Custom baseURL (non-default host/port): auto-probe should not override user config.
Atomic Chat on port 1338: should be detected when 1337 is down.

[kilo-code-bot]

Code Review Summary

Status: 1 Issue Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	0
SUGGESTION	1

Issue Details (click to expand)

SUGGESTION

File	Line	Issue
packages/plugin-atomic-chat/src/utils/validation/validate-config.ts	17	validateConfig silently mutates the config object (sets atomic.npm, atomic.name, atomic.options) — side-effectful validation is surprising; move defaults to enhanceConfig or document/rename

Resolved since last review

• ✅ CRITICAL: Broken not_found branch in categorizeError — two nested ifs with no return and a bare labeled object expression — replaced with a single, correct multi-line condition and a proper return statement
• ✅ WARNING: retryWithBackoff retrying against stale model status cache — fixed via validationAttempt counter + forceRefresh on retry
• ✅ WARNING: Two separate ModelStatusCache singletons in chat-params-hook and enhance-config — fixed by introducing shared-model-status-cache.ts
• ✅ WARNING: enhanceConfig / getModels making separate health + discovery calls — fixed by deduplicating into a single fetchModelsEndpoint call
• ✅ WARNING: fetchModelsDirect silently returning [] on non-OK responses — now throws on non-success status
• ✅ WARNING: Promise.race timeout in config-hook leaving dangling promises — replaced with AbortController + AbortSignal threaded through enhanceConfig
• ✅ SUGGESTION: Hard-coded port 1337 probe — now probing both 1337 and 1338 from ATOMIC_CHAT_PROBE_PORTS
• ✅ WARNING: kilocode_change annotations were too broad in dialog-provider.tsx — narrowed to inline markers

Files Reviewed (incremental — 1 file)

• packages/plugin-atomic-chat/src/utils/index.ts — ✅ CRITICAL resolved (not_found branch now correctly structured with return)

Fix these issues in Kilo Cloud

---

_{Reviewed by claude-4.6-sonnet-20260217 · 292,328 tokens}

_{Review guidance: REVIEW.md from base branch main}

[yanalialiuk]

.

[yanalialiuk]

.

[yanalialiuk]

.

[kilo-code-bot]

WARNING: Empty catch block silently swallows errors — violates style guide

The style guide requires that catch blocks never be empty. At minimum, bind the error and log it so failures are visible:

} catch (err) {
  // network/timeout error — treat as unhealthy
  return false
}

Same applies to the analogous blocks in discoverAtomicChatModels (line 61) and autoDetectAtomicChat (line 82).