Conduit - Universal AI Bridge for VS Code

Connect VS Code to any AI provider through a single extension. One chat interface for Grok, Claude, Gemini, ChatGPT, OpenAI Codex, OpenCode, Pi, and local models, powered by conduit-bridge.

Current version: 0.10.2

Status: Active development. Requires conduit-bridge v0.9.0+ at http://127.0.0.1:31338 (mode on chat completions: chat | plan | agent). CLI, API, and LM Studio providers are the supported transports. Web/Playwright providers were removed in the bridge.

Quick Start

Prerequisites

VS Code 1.90 or newer
conduit-bridge installed and running

Installation

Option A - From .vsix file:

code --install-extension conduit-vscode-0.10.2.vsix

Or in VS Code: Extensions > ... > Install from VSIX...

Option B - From source:

git clone https://github.com/elvatis/conduit-vscode
cd conduit-vscode
npm install --include=dev
npx @vscode/vsce package --no-dependencies
code --install-extension conduit-vscode-0.10.2.vsix

First Launch

Start the bridge: conduit-bridge start
Open VS Code, the extension activates automatically
The Conduit AI panel appears in the bottom panel area
Drag the Conduit AI tab to the secondary sidebar (right side) to place it next to Copilot / Claude Code
Click the model name in the toolbar to select your preferred model
Start chatting

Note: After first install, you must fully close and reopen VS Code (not just reload) for the Sessions panel to appear. VS Code only reads new view registrations on startup.

Features

Chat Interface

Streaming responses with full Markdown rendering (headings, code blocks, tables, lists, blockquotes, bold, italic, links)
Copy and insert-code actions on every response
Per-message model tag showing which model generated each response
Automatic context window management, trims conversation history to fit model limits
Token usage tracking via /cost

Agent Step Cards

In Agent mode, responses are structured as collapsible step cards
Each step shows an animated spinner while streaming, then a green checkmark when complete
Steps are click-to-expand/collapse for easy navigation
Models are instructed to use ### Step N: Title format for structured output

Sessions (History)

Native VS Code tree view panel (like GitHub Copilot's Sessions)
Persistent across VS Code restarts (up to 50 sessions)
Click any session to reload the full conversation
New Session button (+) and Refresh button in the panel title bar
Delete button (trash icon) on hover for each session
Sessions auto-save after each message exchange

Model Selection

Native VS Code QuickPick, opens a full-width, searchable picker at the top of the editor
Models grouped by provider (WEB-GROK, WEB-CLAUDE, WEB-GEMINI, etc.)
Tier icons: star for flagship models, half-star for mid-tier
Context window size shown next to each model (131K, 200K, 1M)
Friendly display names with version numbers (e.g. "Claude Sonnet 4.6", "Grok Expert")
Auto mode, automatically selects the best model based on task complexity
Model-mode compatibility warnings when a model doesn't support the current chat mode

Code Intelligence

Inline completions (ghost text), language-aware for 20+ languages
Inline Chat (Ctrl+I), describe a change at the cursor, review as a diff
Explain / Refactor / Generate Tests, right-click context menu on selections
Fix diagnostics, send all file errors/warnings to the AI
Terminal command suggestions, describe what you want, get a shell command
Commit message generation (Ctrl+Shift+M), generates from staged git diff

Infrastructure

Health dashboard (Conduit: Health Dashboard), real-time provider status
Bridge manager, start/stop/restart conduit-bridge from VS Code
Per-provider login, Grok, Claude, Gemini, ChatGPT login commands
Auto-start bridge, starts automatically if not running on activation
Consolidated status bar, bridge status, model count, and current model in one item

Background Agents

Conduit supports spawning background coding agents that work independently while you continue coding.

Spawn Agent

Conduit: Spawn Agent - select a model and enter a task. The agent runs in the background with its own output channel. Spawn sends mode: agent and the workspace path so the selected CLI can edit files with its own tools.

Fix Issue (Worktree Isolation)

Conduit: Fix Issue - enter a GitHub issue number. Conduit:

Creates a git worktree on a dedicated branch (fix/issue-<N>)
Spawns an agent in the isolated worktree
The agent reads the codebase, implements the fix, and commits

Parallel safety: Multiple Fix Issue commands can run simultaneously. Worktree creation is serialized via a file lock ($REPO/.git/worktree-create.lock) with 2.5s stagger to prevent .git/config.lock contention. (Credit: @m13v)

Safe Worktree Cleanup

When removing worktrees, Conduit checks branch merge status:

Merged branches: safe to remove
Unmerged + recently active (within 24h): removal blocked (prevents losing review feedback)
Unmerged + abandoned (>24h idle): allowed
Force flag: overrides all safety checks

Agent Tools

Chat Agent mode runs tools in the extension and keeps the CLI on mode: chat (read-only). Spawn and Fix Issue do not use this table; they send mode: agent and the CLI uses its own tools.

Host-side chat Agent tools:

Tool	Permission	Description
`readFile`	safe	Read file contents (with optional line range)
`writeFile`	destructive	Create or overwrite files
`listFiles`	safe	List files matching a glob pattern
`searchCode`	safe	Regex search across workspace files
`runCommand`	destructive	Execute shell commands
`readDiagnostics`	safe	Read VS Code errors and warnings
`applyDiff`	destructive	Search-and-replace within a file
`createWorktree`	destructive	Create a git worktree for parallel work
`removeWorktree`	destructive	Remove a worktree (merge-status aware)

Batch Fix Issues

Conduit: Batch Fix Issues - fix multiple GitHub issues in parallel:

Enter a label filter (e.g. bug, good-first-issue) or leave empty for all open issues
Set max issues to process
Select a model
Conduit fetches matching issues via gh, shows a confirmation dialog, then spawns agents with worktree isolation
Max 3 concurrent agents (serialized worktree creation via lock)
Progress and summary shown in a dedicated "Conduit Batch Fix" output channel

Session Panel

The Sessions tree view shows all background agents:

Running: animated spinner icon (blue)
Completed: checkmark icon (green)
Failed: error icon (red)
Click to view output, right-click to kill running agents

Model Fallback Chain

When a model fails with a transient error (rate limit, timeout, capacity, auth failure), Conduit automatically retries with the next model in the fallback chain:

Gemini 3.1 Pro High → Gemini 3.6 Flash High → Gemini 3.5 Flash High
Claude Opus 5 → Claude Sonnet 5 → Claude Haiku 4.5
Grok 4.6 → Grok 4.5
GPT-5.6 Sol → GPT-5.6 Terra

Fallback is triggered by: 429/503 errors, rate limits, capacity issues, timeouts, and auth failures. Non-transient errors (syntax, missing args) are not retried. The RouteResult metadata tracks which model was actually used and why.

Supported Models

Models are served by conduit-bridge. The extension displays whatever the bridge reports via /v1/models. Available models depend on which providers are logged in.

CLI Models (requires the matching CLI installed and authenticated)

Provider	Models	Context
Grok CLI (`cli-grok`)	Grok 4.6, Grok 4.5, Grok 4.3	256K-2M
Claude CLI (`cli-claude`)	Claude Opus 5, Sonnet 5, Fable 5, Haiku 4.5	200K-1M
Gemini CLI (`cli-gemini`)	Gemini 3.1 Pro High/Low, 3.6 Flash High/Medium/Low, 3.5 Flash	1M
Codex CLI (`cli-codex`)	GPT-5.6 Sol/Terra/Luna, GPT-5.5, GPT-5.5 Pro	400K-1M

API Models (API key saved in the bridge dashboard)

Provider	Models	Context
Claude API	Claude Fable 5, Opus 5, Sonnet 5, Haiku 4.5	200K-1M
Gemini API	Gemini 3.7/3.6/3.5 Flash, 3.1 Pro	1M
Codex API	GPT-5.6 Sol/Terra/Luna, GPT-5.5	400K-1M
OpenRouter / Perplexity	Catalog from the live provider	varies

Local Models

Provider	Models	Context
LM Studio	Whatever LM Studio is serving (`lmstudio/...`)	varies
BitNet	BitNet 1.58 2B (CPU inference, optional local endpoint)	4K

Chat Interface

The Conduit Chat panel lives in the bottom panel area by default. For the best experience, drag it to the secondary sidebar (right side) where it sits alongside GitHub Copilot and Claude Code.

Toolbar

The toolbar at the bottom of the chat has:

+ - Attach context (current selection or file from disk)
Mode button (Ask/Edit/Agent/Plan) - click to switch chat mode via QuickPick
Model button (e.g. "Claude Sonnet 4.6") - click to switch model via QuickPick
Settings - open Conduit extension settings
Send - send your message (or press Enter)

Chat Modes

Click the mode button in the toolbar to switch between modes:

Mode	Purpose	System Behavior
Ask	Answer questions about code	Conversational, explains concepts, provides examples
Edit	Modify and refactor code	Focuses on producing code changes, minimal explanation
Agent	Plan and build features	Host-side tool loop in the extension. CLI providers stay read-only.
Plan	Create implementation plans	Calls the CLI's native plan function through conduit-bridge (`mode: plan`)

Mode Compatibility

Models are classified into tiers that determine which modes they support:

Tier	Modes	Examples
Tier 1 (flagship)	Ask, Edit, Agent, Plan	Claude Opus 4.6, GPT-5.4 Pro, Gemini 3.1 Pro
Tier 2 (mid-tier)	Ask, Edit, Plan	Claude Haiku 4.5, Grok Fast, GPT-5.3 Instant
Tier 3 (fast)	Ask only	GPT-5 Thinking Mini, BitNet 2B

If you select a mode that your current model doesn't support, Conduit shows a warning with a suggested alternative model.

Agent Mode

In Agent mode, models produce structured output with step-by-step reasoning. Each step is rendered as a collapsible card in the chat:

While the model streams a step, it shows an animated spinner
When a step finishes (the next step begins), it shows a green checkmark
When the full response is done, all steps show checkmarks
Click any step header to expand/collapse its contents

This gives you a clear overview of the model's reasoning process without being overwhelmed by long responses.

Model Selection

Click the model name in the toolbar (or use /model) to open the model picker.

The QuickPick shows all available models grouped by provider:

  Auto  best for task
  --- WEB-GROK ---
  * Grok Expert                  131K context - Ask, Edit, Agent, Plan
    Grok Fast                    131K context - Ask, Edit, Plan
  --- WEB-CLAUDE ---
  * Claude Sonnet 4.6            200K context - Ask, Edit, Agent, Plan
    Claude Opus 4.6              200K context - Ask, Edit, Agent, Plan
  ...

Type to search/filter models
The checkmark shows the currently selected model
Star icons indicate model tier (full star = tier 1, half star = tier 2)
Supported modes are shown next to each model
Auto mode picks the best model per message based on complexity

Auto Model Selection

When set to Auto, Conduit analyzes your message to determine complexity:

Simple (short questions, "explain this", "fix typo") → fast models like Grok Fast, Gemini 3 Fast
Moderate (code changes, debugging) → mid-tier models like Gemini 3 Thinking, Claude Sonnet
Complex (architecture, multi-file, "build a system") → flagship models like Claude Opus, GPT-5.4 Pro

Session Management

The Sessions panel appears below the Chat panel (or as a separate collapsible section). It works like GitHub Copilot's session history.

Actions

New Session (+) - start a fresh conversation (saves the current one)
Refresh - reload the session list
Click a session - load that conversation into the chat
Rename (edit icon on hover) - give the session a custom name
Delete (trash icon on hover) - remove a session permanently

Session Rename

You can rename any session to keep track of what you were working on:

Use /rename My Feature Work in the chat input
Or click the edit icon next to a session in the Sessions panel
Custom names persist and are shown in the session list

Model Switch Handoff

When models change mid-conversation (either via Auto mode or manual switch), Conduit automatically injects a compressed summary of the previous context. This means the new model understands what was discussed before and can continue seamlessly.

Working Context Persistence

Each session stores a "working summary", a compressed snapshot of what you were working on. When you switch between sessions, this summary is restored so no context is lost.

Sessions are stored in VS Code's global state and persist across restarts. Up to 50 sessions are kept.

Move to Secondary Sidebar

To place Conduit next to GitHub Copilot and Claude Code in the secondary sidebar (right side), run the command "Conduit: Move to Secondary Sidebar" from the command palette (Ctrl+Shift+P).

Slash Commands

Type / in the chat input to see autocomplete suggestions.

Command	Description
`/help`	Show all available commands, context mentions, and keyboard shortcuts
`/fix`	Fix errors and warnings in the current file
`/explain`	Explain the selected code
`/tests`	Generate tests for the selected code
`/refactor [instruction]`	Refactor selected code (optional instruction)
`/plan [task]`	Switch to Plan mode and ask the CLI native planner (`mode: plan`)
`/commit`	Generate a commit message from staged git changes
`/clear`	Clear the current chat (without saving)
`/new`	Save current chat and start a new session
`/cost`	Show estimated token usage for this session
`/model [name]`	Switch model by name, or list all available
`/rename [name]`	Rename the current session
`/mode [ask\|edit\|agent\|plan]`	Switch chat mode

Context Mentions

Add context to your messages by typing # followed by a mention:

Mention	What it attaches
`#file:src/main.ts`	Full content of the file
`#file:src/main.ts:10-20`	Lines 10-20 of the file
`#selection`	Current editor selection
`#problems`	All errors/warnings in the current file
`#workspace`	Lightweight workspace folder structure overview
`#codebase`	Deep search: file tree + contents of up to 30 source files (~80K chars)
`#terminal`	Terminal output (select text in terminal first)

Examples

Explain this function #file:src/utils/parser.ts:45-80
Fix the errors in this file #problems
How does authentication work? #codebase
Where are the API routes? #workspace
Refactor #selection based on patterns in #file:src/helpers.ts

#workspace vs #codebase

#workspace is lightweight, just the folder structure. Use it when you need a quick overview of where things are.
#codebase is deep, includes the folder structure PLUS the actual contents of up to 30 prioritized source files. Use it when the model needs to understand how your code works. Files are prioritized: config files first, then entry points, then by directory depth.

Code Intelligence

Right-click Context Menu

Select code in the editor and right-click to access:

Conduit: Explain Selected Code - get a detailed explanation
Conduit: Refactor Selected Code - suggest improvements
Conduit: Generate Tests for Selection - create unit tests
Conduit: Inline Edit - edit code with a prompt

Inline Completions (Ghost Text)

Conduit provides inline code suggestions as you type, similar to GitHub Copilot:

Works for 20+ programming languages
Respects language-specific conventions
Configurable delay (conduit.inlineTriggerDelay)
Toggle on/off with Conduit: Toggle Inline Suggestions

Inline Chat

Press Ctrl+I (or Cmd+I on Mac) anywhere in the editor to open the inline chat:

Type a description of what you want to change (e.g. "add error handling", "convert to async")
Conduit generates the code change
A diff view opens showing the proposed changes
Click Accept to apply or Reject to discard

Custom Instructions

Customize Conduit's behavior for your project by creating instruction files:

Project-level (checked into repo)

.conduit/instructions.md - Conduit-specific instructions
CLAUDE.md - also recognized (Claude Code compatible)
.github/copilot-instructions.md - also recognized (Copilot compatible)

User-level (global, all projects)

~/.conduit/instructions.md

Example `.conduit/instructions.md`

## Project Context
This is a React 19 + TypeScript project using Tailwind CSS.
The backend is a Node.js Express API with PostgreSQL.

## Coding Conventions
- Use functional components with hooks
- Prefer named exports
- Use descriptive variable names
- Always add JSDoc comments to exported functions
- Tests use vitest with React Testing Library

## Important
- Never use `any` type
- Always handle loading and error states
- API calls go through the `/lib/api` module

Keyboard Shortcuts

Shortcut	Action
`Ctrl+Shift+G`	Open/focus the Conduit Chat panel
`Ctrl+I`	Inline Chat - edit code at cursor
`Ctrl+Shift+I`	Inline Edit - edit selected code with prompt
`Ctrl+Shift+E`	Explain selected code
`Ctrl+Shift+M`	Generate commit message from staged changes
`Enter`	Send message in chat
`Shift+Enter`	Insert new line in chat

Configuration

Open settings with Ctrl+, and search for "conduit", or use the gear icon in the chat toolbar.

Setting	Default	Description
`conduit.proxyUrl`	`http://127.0.0.1:31338`	Base URL of the conduit-bridge proxy
`conduit.apiKey`	`cli-bridge`	API key for the proxy
`conduit.defaultModel`	`cli-gemini/gemini-2.5-pro`	Default model (full ID from the model picker)
`conduit.inlineSuggestions`	`true`	Enable inline ghost-text completions
`conduit.inlineTriggerDelay`	`600`	Delay in ms before requesting inline suggestion
`conduit.contextLines`	`80`	Lines of surrounding code to include as context
`conduit.includeOpenFiles`	`true`	Include other open editor tabs as additional context
`conduit.maxOpenFilesContext`	`3`	Maximum number of open files to include
`conduit.terminalIntegration`	`true`	Enable terminal command suggestions
`conduit.autoStatusBar`	`true`	Show connection status in the status bar

Provider Setup

Each AI provider is authenticated in conduit-bridge, not inside VS Code. The extension reads /v1/status (connected, loginType) and /v1/models.

Start the bridge (autostart is fine) on http://127.0.0.1:31338
Open Ctrl+Shift+P → Conduit: Open Bridge Dashboard
Save API keys on the API page, or sign in to each CLI (grok login, claude, codex login, Gemini CLI)
In VS Code, connected CLI/API/local providers show as Connected in the Bridge Manager and Health Dashboard. Their models appear in the picker.

CLI connected means the CLI is installed and authenticated, not merely on PATH.

Checking Provider Status

Use Ctrl+Shift+P → Conduit: Health Dashboard or Conduit: Bridge Manager to see which providers are connected and which models are available.

Testing

Conduit has a comprehensive test suite with 283 tests across 17 test files.

Running Tests

npm test                    # run all tests
npm run test:coverage       # run with coverage report

Test Categories

Test File	Tests	Coverage
`agent-parser.test.ts`	25	Agent output parsing, step card extraction
`agent-tools.test.ts`	18	Tool execution (readFile, writeFile, applyDiff, etc.)
`worktree-tools.test.ts`	17	Worktree lock serialization, merge-status safety
`cli-runner-failover.test.ts`	9	Bridge failover chain
`aahp-context.test.ts`	10	AAHP v3 context detection, loading, block building
`llm-tool-validation.test.ts`	14	Tool catalog schema (optional live check via `LLM_MODEL`)
`model-registry.test.ts`	39	Model capabilities, tiers, auto-selection
`sessions-tree-provider.test.ts`	19	Session tree, background agent status
`proxy-client.test.ts`	19	HTTP streaming, error handling
`chat-view-provider.test.ts`	17	Chat webview, slash commands
`context-builder.test.ts`	16	Editor context collection
`mention-parser.test.ts`	11	#file, #selection, #codebase parsing
`custom-instructions.test.ts`	8	Instruction file loading
`utils.test.ts`	17	Shared utilities
`config.test.ts`	4	Settings reader
`inline-provider.test.ts`	7	Ghost-text completions

Troubleshooting

"No view is registered with id: conduit.sessionsView"

Fully close and reopen VS Code. VS Code only reads new view registrations on startup, not on extension reload.

Models not showing up

The model list comes from the bridge (/v1/models). Check:

Is the bridge running? Run Conduit: Check Proxy Status
Is the provider logged in? Run the login command for that provider
Check the Health Dashboard for provider status

Empty responses / "No response received"

The provider may not be authenticated, run the login command
The model may not support your request type
Check the bridge logs: Conduit: Show Bridge Logs

Chat panel not visible

Open the command palette (Ctrl+Shift+P)
Run Conduit: Open Chat
If the panel appears at the bottom, right-click the tab and select "Move to Secondary Side Bar"

Inline completions not working

Check that conduit.inlineSuggestions is enabled in settings
Ensure the bridge is running and at least one model is available
Try increasing conduit.inlineTriggerDelay if suggestions are too slow

Bridge not starting

The extension tries to auto-start the bridge on activation. If it fails:

Check if conduit-bridge is installed: conduit-bridge --version
Check if port 31338 is already in use
Start manually: conduit-bridge start
Check bridge logs: Conduit: Show Bridge Logs

Development

Setup

git clone https://github.com/elvatis/conduit-vscode
cd conduit-vscode
npm install --include=dev

Build and Run

npm run dev     # watch mode with source maps
npm run build   # production build (minified)
npm run lint    # eslint
npm test        # run tests (vitest, 283 tests)

Press F5 in VS Code to launch the Extension Development Host for debugging.

CI

CI (.github/workflows/ci.yml) runs on every push/PR to main. It builds, runs the vitest suite (including tool-catalog schema tests), and uploads coverage artifacts.

Package for Distribution

npx @vscode/vsce package --no-dependencies
# produces conduit-vscode-0.10.2.vsix

Project Structure

conduit-vscode/
  src/
    extension.ts              - activation, command registration
    commands.ts               - command palette registrations (spawn, fix issue, etc.)
    chat-view-provider.ts     - main chat webview (sidebar), slash commands, markdown
    chat-panel.ts             - chat panel webview host
    sessions-tree-provider.ts - native sessions tree view + background agent sessions
    models-tree-provider.ts   - model picker tree view
    model-registry.ts         - model capabilities, display names, tiers, auto-selection
    proxy-client.ts           - HTTP/streaming client for the bridge
    bridge-status.ts          - normalize conduit-bridge /v1/status (connected + loginType)
    cli-runner.ts             - background agents via conduit-bridge /v1/chat/completions
    cost-tracker.ts           - token usage parsing and cost estimation per agent session
    agent-loop.ts             - multi-turn agent loop with tool execution
    agent-parser.ts           - agent output parsing (step cards, tool calls)
    agent-tools.ts            - workspace tools (read/write/search/worktree/diff)
    agent-types.ts            - shared type definitions for the agent system
    aahp-context.ts           - AAHP v3 context auto-detection and injection
    mention-parser.ts         - #file, #selection, #workspace, #codebase parsing
    context-builder.ts        - editor context collection
    bridge-manager.ts         - bridge lifecycle management
    bridge-panel.ts           - bridge manager webview
    inline-provider.ts        - ghost-text inline completions
    inline-chat.ts            - Ctrl+I inline chat with diff
    custom-instructions.ts    - .conduit/instructions.md loader
    commit-message.ts         - git commit message generation
    config.ts                 - settings reader
    health-panel.ts           - health dashboard webview
    status-bar.ts             - consolidated status bar item
    utils.ts                  - shared utilities
  src/__tests__/              - 17 test files, 283 tests (vitest)
  dist/
    extension.js              - bundled output (esbuild)
  media/
    icon.png                  - extension icon
    icon.svg                  - extension icon (vector)
    sidebar-icon.svg          - panel icon

Related Projects

Project	Description	Overlap
aahp-orchestrator	VS Code extension for AAHP v3 context injection	Context building for coding agents
aahp-runner	Autonomous CLI agent runner	CLI agent spawning, backend routing
aahp-cron	Pipeline orchestrator for multi-repo agent runs	Batch agent execution pattern

Roadmap

Open issues tracking planned features:

#	Feature	Status
#6	Model failover chain for agent sessions	✅ Done
#7	AAHP context integration for agent sessions	✅ Done
#8	Multi-issue batch mode (aahp-cron pattern)	✅ Done
#9	Agent output streaming to session panel	✅ Done
#10	Shared agent backend abstraction with aahp-runner	✅ Done
#11	LLM tool-call validation CI (multi-model smoke test)	✅ Done
#12	Agent session persistence and resume	✅ Done
#13	Cost tracking per agent session	✅ Done

Changelog

v0.10.2 (2026-09-03)

Type checking, which this extension never had: esbuild bundles without checking types, so npm run build passed over 159 errors while TRUST.md recorded the opposite as verified. npm run typecheck now runs in CI
157 of those came from tsconfig.json declaring no "types", so node globals resolved to nothing. The two real ones are fixed
The sessions tree can clear its highlight again: refresh took string | undefined while the event it listens to emits string | null, and null is what "no active session" looks like
No change to chat, providers or the bridge protocol

v0.10.1 (2026-09-03)

.github/workflows/release.yml builds the .vsix from the tagged commit and publishes it as the release asset in one run, so a release cannot exist without its download. Three of the seventeen releases cut so far are still empty (v0.1.0, v0.2.0, v0.6.0) and v0.10.0 had no asset for five minutes after publication; the release object and the artifact were two separate manual acts
The workflow refuses a tag that disagrees with package.json, is not an ancestor of main, or has no matching CHANGELOG.md section, and re-runs aahp verify, aahp check, aahp doctor, the build and the tests before publishing. It asserts afterwards that the release is not a draft and carries a fully uploaded, non-empty asset whose bytes are the bytes the gates passed
Note when tagging: GitHub loads the workflow from the tagged commit, so a tag cut before this file exists runs nothing at all, silently
Release hygiene only; no runtime change. v0.10.0 shipped with no .vsix attached to its GitHub release, which is this project only distribution channel, and with a README that still read 0.9.0 and install snippets that still read 0.7.6
versionSites is now configured, so aahp check fails when the version in package.json disagrees with README, CHANGELOG.md, STATUS.md or DASHBOARD.md. The gate existed all along and reported SKIP - no versionSites configured
CI runs aahp check alongside verify and doctor; it never ran before, so no governance gate could fail the build
Removed the em dashes the repo own forbidden-patterns rule bans, including the one in the displayName shown in the VS Code extensions list

v0.10.0 (2026-09-03)

Deleted the local model tables: conduit-bridge v0.9.0 reports context_window, max_output_tokens and display_name, so MODEL_LIMITS, PROVIDER_FALLBACK_LIMITS, MODEL_DISPLAY_NAMES and MODEL_TIERS are gone (129 lines). They went stale the moment the bridge started discovering catalogs instead of shipping a pinned list
Every model reports every chat mode. Tiering silently withheld Agent from any id this build had not heard of, which with a discovered catalog was 21 of 36 CLI models
Trimming honours max_prompt_chars, so a prompt cannot exceed what the transport carries
local-* models keep a conservative 8192-token assumption: the bridge never sees them
Requires conduit-bridge v0.9.0

v0.9.1 (2026-09-02)

Request timeout is a budget on the whole turn, not an idle timer, because no CLI provider streams and the bridge writes nothing until the child exits. Configurable via conduit.requestTimeout, default 330s, above the bridge 300s CLI ceiling
A timeout no longer triggers the fallback chain; it used to spend four budgets and four killed CLI runs before anything appeared
Typed RequestTimeoutError: the old substring test classified every case backwards
Stop aborts the request, which is what makes the bridge kill the CLI
Removed the retired cli-gemini/gemini-3.5-flash-high from the pickers and fallbacks
Background spawn fails early without a workspace instead of an opaque HTTP 400

v0.9.0 (2026-09-02)

Plan chat sends mode: plan plus workspace cwd so conduit-bridge calls each CLI native planner
Background spawn/fix-issue sends mode: agent with cwd so CLI providers can write the workspace
VS Code chat Agent mode stays host-side and sends mode: chat
Requires conduit-bridge v0.7.0+ (mode on chat completions)
Removed the weekly llm-validation.yml GitHub Action (schema tests stay in CI)

v0.7.6 (2026-07-17)

Dev deps: @typescript-eslint/eslint-plugin ^8.64.0 (aligned with @typescript-eslint/parser 8.64.0), eslint ^10.7.0, @types/node ^26.1.1
Held back typescript 6 -> 7: @typescript-eslint/eslint-plugin@8.64 peers on typescript@">=4.8.4 <6.1.0", so the lint toolchain cannot resolve against TS 7 yet; added a Dependabot ignore for the typescript major
No runtime changes (dev-only bumps; bundled dist/extension.js unchanged from v0.7.5)

v0.7.5 (2026-07-17)

Security: command-injection hardening (CWE-78) in agent tool executors - branch names and command tokens from LLM tool args are validated, all git/shell sinks moved to execFile/execFileSync with no shell (toolCreateWorktree, toolRunCommand, toolRemoveWorktree, getBranchStatus, batchFixIssues label validation); 32 regression tests added
Security: transitive brace-expansion bump (GHSA-jxxr-4gwj-5jf2, ReDoS)
Dev deps: eslint ^10.6.0, vitest + @vitest/coverage-v8 4.1.10, @typescript-eslint/eslint-plugin ^8.63.0, @typescript-eslint/parser ^8.64.0, esbuild ^0.28.1, @types/node ^26.1.0; transitive vite resolves 8.1.3
Re-pinned @types/vscode to ~1.90.0 to match engines.vscode (vsce packaging requirement) and added a Dependabot ignore rule so it stops re-bumping it
Packaging: .vscodeignore extended so the .vsix no longer bundles AAHP/CI repo tooling (scripts/, schema/, .scg-history/, .github/, CLAUDE.md)
CI: supply-chain-guard now tracks the moving @v5 release branch; actions/checkout 7, setup-node 7, setup-python 6, codeql-action 4, upload-artifact 7; Dependabot exempted from the aahp-verify handoff gate; AAHP gate tooling synced to v3.5.0
Docs: handoff docs reconciled with reality (multi-turn agent loop T-016 confirmed shipped, closed #52; Marketplace listing T-006 dropped by decision, closed #53)
314 tests passing

v0.7.4 (2026-05-17)

Dev deps: typescript 5.x to ^6.0.3, @types/node to ^25.7.0, @types/vscode to ^1.118.0
Dev deps: @vitest/coverage-v8 and vitest to ^4.1.6
No runtime changes; closes Dependabot PRs #40, #41, #42, #43

v0.7.3 (2026-05-05)

Dev deps: @types/node major bump 22.x to ^25.5.0
Dev deps: eslint major bump 8.x to ^10.1.0
No runtime changes; closes Dependabot PRs #30 and #34

v0.7.2 (2026-05-05)

Security: vite 8.0.5 (path traversal in .map handling, GHSA-4w7w-66w2-5vf9)
Security: esbuild ^0.27.7 (dev server bypass, GHSA-67mh-4wv8-2f99)
Security: brace-expansion transitive bump (GHSA-f886-m6hf-6m8v)
Aligned @typescript-eslint/parser to ^8.59.2 to match the plugin and unblock CI peer-dep resolution
Bumped vitest and @vitest/coverage-v8 to ^4.1.5

v0.7.1 (2026-03-24)

Fix: resolve .cmd wrapper to actual JS entry on Windows when spawning the bridge

v0.7.0 (2026-03-18)

CI workflow (build + test on push/PR) and weekly LLM validation workflow (#11)
Shared agent backend abstraction, refactored cli-runner.ts (#10)
Session persistence and resume: sessions survive VS Code restarts (#12)
Cost tracking per agent session with token parsing and budget limits (#13)
Resume, Remove, Clear commands for session management (#12)
Cost summary command with per-model breakdown (#13)
Branch protection enabled on main
295+ tests across 19 test files

v0.6.0 (2026-03-18)

Agent backends: Claude CLI, Gemini CLI, OpenAI Codex, OpenCode, Pi
Background agent sessions with spawn/monitor/kill
Git worktree isolation for parallel agent work
Worktree lock serialization (prevents .git/config.lock contention)
Merge-status aware worktree cleanup
Fix Issue command (auto-worktree + agent spawn)
Model fallback chain definitions
Live agent output streaming to session panel (#9)
277 tests across 17 test files

v0.5.0

Reliable agent loop with tool execution
Multi-model auto-selection
Local model support (BitNet)
Smart fallback on model errors

v0.4.0

Per-provider sessions with model-aware completions
Streaming metadata display (timestamps, duration, tokens)
Inline chat with diff preview

v0.3.0

Initial release: chat interface, model picker, inline completions
Session management with persistence
Custom instructions support

Name		Name	Last commit message	Last commit date
.ai/handoff		.ai/handoff
.github		.github
media		media
schema		schema
scripts		scripts
src		src
.gitignore		.gitignore
.vscodeignore		.vscodeignore
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
aahp.config.json		aahp.config.json
package-lock.json		package-lock.json
package.json		package.json
tsconfig.json		tsconfig.json
vitest.config.ts		vitest.config.ts

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

Conduit - Universal AI Bridge for VS Code

Table of Contents

Quick Start

Prerequisites

Installation

First Launch

Features

Chat Interface

Agent Step Cards

Sessions (History)

Model Selection

Code Intelligence

Infrastructure

Background Agents

Spawn Agent

Fix Issue (Worktree Isolation)

Safe Worktree Cleanup

Agent Tools

Batch Fix Issues

Session Panel

Model Fallback Chain

Supported Models

CLI Models (requires the matching CLI installed and authenticated)

API Models (API key saved in the bridge dashboard)

Local Models

Chat Interface

Toolbar

Chat Modes

Mode Compatibility

Agent Mode

Model Selection

Auto Model Selection

Session Management

Actions

Session Rename

Model Switch Handoff

Working Context Persistence

Move to Secondary Sidebar

Slash Commands

Context Mentions

Examples

#workspace vs #codebase

Code Intelligence

Right-click Context Menu

Inline Completions (Ghost Text)

Inline Chat

Custom Instructions

Project-level (checked into repo)

User-level (global, all projects)

Example .conduit/instructions.md

Keyboard Shortcuts

Configuration

Provider Setup

Checking Provider Status

Testing

Running Tests

Test Categories

Troubleshooting

"No view is registered with id: conduit.sessionsView"

Models not showing up

Empty responses / "No response received"

Chat panel not visible

Inline completions not working

Bridge not starting

Development

Setup

Build and Run

CI

Package for Distribution

Project Structure

Related Projects

Roadmap

Changelog

v0.10.2 (2026-09-03)

v0.10.1 (2026-09-03)

Example `.conduit/instructions.md`