@elvatis_com/openclaw-cli-bridge-elvatis

openclaw-cli-bridge-elvatis

OpenClaw plugin that bridges multiple AI model providers into OpenClaw via a local OpenAI-compatible proxy on port 31337. Supports CLI subprocesses, browser sessions, and direct API integrations -- all accessible through a unified slash-command interface and multi-phase pipeline.

Version: 2026.7.1 | Node.js: 22+ | OpenClaw: 2026.7.x

What's new in v2026.7.1

• Switched to date-based versioning (YYYY.M.D) aligned with OpenClaw releases.
• Plugin now compiles against OpenClaw 2026.7.1 SDK (resolved openclaw/plugin-sdk module path change).
• Claude Code CLI models re-added with the latest models: Fable 5, Sonnet 5, Opus 4.8, and legacy variants. Each runs in --print mode via stdin.
• Grok CLI added (cli-grok/grok-4.5) via --prompt-file headless mode.
• Phase 5 expanded: Perplexity API (35 models), OpenRouter (17+ models), LM Studio (dynamic discovery).
• Phase 6 added: /pipeline multi-phase AI pipeline (research, architect, implement, review).
• Project is actively maintained again. Archive notice removed.

Table of Contents

• How it works
• Phase 1 -- Codex Auth Bridge
• Phase 2 -- Local Proxy (CLI Subprocesses)
• Phase 3 -- Slash Commands
• Phase 4 -- Web Browser Providers
• Phase 5 -- API Integrations
• Phase 6 -- Pipeline
• Configuration
• Setup
• Architecture
• Requirements
• Development
• Known Issues and Fixes
• Changelog

How it works

The plugin starts a local OpenAI-compatible HTTP proxy on 127.0.0.1:31337. OpenClaw connects to it via the vllm provider. Incoming requests are dispatched to the appropriate backend based on the model path:

vllm/cli-claude/<model>   -> Claude Code CLI subprocess (stdin, --print mode)
vllm/cli-gemini/<model>   -> Gemini CLI subprocess (stdin, cwd=/tmp)
vllm/cli-grok/<model>     -> Grok CLI subprocess (--prompt-file, headless)
openai-codex/<model>      -> OpenAI API (direct, via Codex OAuth tokens)
vllm/local-bitnet/<model> -> llama-server on 127.0.0.1:8082
vllm/opencode/default     -> OpenCode CLI
vllm/pi/default           -> Pi CLI
perplexity/<model>        -> Perplexity REST API
openrouter/<model>        -> OpenRouter REST API
lmstudio/<model>          -> LM Studio REST API

All slash commands use staged switching by default: /cli-<name> stages the switch, /cli-apply applies it. Add --now to any switch command for an immediate switch.

Phase 1 -- Codex Auth Bridge

Reads existing OAuth tokens from ~/.codex/auth.json and registers an openai-codex provider in OpenClaw. No re-login is required. Tokens are refreshed automatically 30 minutes before expiry.

The openai-codex provider routes directly to the OpenAI API -- it does not go through the local proxy.

Phase 2 -- Local Proxy (CLI Subprocesses)

A local proxy on 127.0.0.1:31337 routes model calls to locally installed CLI tools. Prompts are always delivered via stdin to avoid E2BIG errors on long sessions. Message history is truncated to the last 20 messages + system message.

Claude Code CLI models

Model path	CLI invoked	Context
vllm/cli-claude/claude-fable-5	claude -p --output-format json --model claude-fable-5	1M
vllm/cli-claude/claude-sonnet-5	claude -p --output-format json --model claude-sonnet-5	1M
vllm/cli-claude/claude-opus-4-8	claude -p --output-format json --model claude-opus-4-8	1M
vllm/cli-claude/claude-opus-4-7	claude -p --output-format json --model claude-opus-4-7	1M
vllm/cli-claude/claude-opus-4-6	claude -p --output-format json --model claude-opus-4-6	1M
vllm/cli-claude/claude-sonnet-4-6	claude -p --output-format json --model claude-sonnet-4-6	1M
vllm/cli-claude/claude-haiku-4-5	claude -p --output-format json --model claude-haiku-4-5	200k

Gemini CLI models

Prompt is sent via stdin. cwd=/tmp prevents agentic mode from activating.

Model path	Gemini model flag
vllm/cli-gemini/gemini-2.5-pro	gemini-2.5-pro
vllm/cli-gemini/gemini-2.5-flash	gemini-2.5-flash
vllm/cli-gemini/gemini-3.1-pro-preview	gemini-3.1-pro-preview
vllm/cli-gemini/gemini-3-pro-preview	gemini-3-pro-preview (legacy alias)
vllm/cli-gemini/gemini-3-flash-preview	gemini-3-flash-preview

Grok CLI

Model path	CLI invoked
vllm/cli-grok/grok-4.5	grok --prompt-file <tmpfile> (headless)

Codex CLI (OAuth)

These models use the openai-codex provider and call the OpenAI API directly via your Codex OAuth session. They do not go through the local proxy.

Model path	Notes
openai-codex/gpt-5.3-codex	Tested
openai-codex/gpt-5.3-codex-spark	Tested
openai-codex/gpt-5.2-codex	Tested
openai-codex/gpt-5.4	May require upgraded OAuth scope
openai-codex/gpt-5.5	May require upgraded OAuth scope
openai-codex/gpt-5.1-codex-mini	Tested

Other CLI providers

Model path	CLI invoked	Notes
vllm/opencode/default	opencode run "prompt"	Requires opencode CLI
vllm/pi/default	pi -p "prompt"	Requires pi CLI
vllm/local-bitnet/bitnet-2b	llama-server on port 8082	CPU inference, no API key

Phase 3 -- Slash Commands

All slash commands use staged switching by default. A staged switch is queued but not applied until you run /cli-apply, preventing mid-session model disruption. Use --now to apply immediately.

Claude Code CLI

Command	Model
/cli-fable5	vllm/cli-claude/claude-fable-5
/cli-sonnet5	vllm/cli-claude/claude-sonnet-5
/cli-opus	vllm/cli-claude/claude-opus-4-8
/cli-opus47	vllm/cli-claude/claude-opus-4-7
/cli-opus46	vllm/cli-claude/claude-opus-4-6
/cli-sonnet	vllm/cli-claude/claude-sonnet-4-6
/cli-haiku	vllm/cli-claude/claude-haiku-4-5

Gemini CLI

Command	Model
/cli-gemini	vllm/cli-gemini/gemini-2.5-pro
/cli-gemini-flash	vllm/cli-gemini/gemini-2.5-flash
/cli-gemini3	vllm/cli-gemini/gemini-3.1-pro-preview
/cli-gemini3-pro-preview	vllm/cli-gemini/gemini-3-pro-preview
/cli-gemini3-flash	vllm/cli-gemini/gemini-3-flash-preview

Codex CLI

Command	Model
/cli-codex	openai-codex/gpt-5.3-codex
/cli-codex-spark	openai-codex/gpt-5.3-codex-spark
/cli-codex52	openai-codex/gpt-5.2-codex
/cli-codex54	openai-codex/gpt-5.4
/cli-codex55	openai-codex/gpt-5.5
/cli-codex-mini	openai-codex/gpt-5.1-codex-mini

Grok CLI

Command	Model
/cli-grok	vllm/cli-grok/grok-4.5

Other CLI providers

Command	Model
/cli-opencode	vllm/opencode/default
/cli-pi	vllm/pi/default
/cli-bitnet	vllm/local-bitnet/bitnet-2b

Perplexity API

Command	Model
/plex-opus	Claude Opus via Perplexity
/plex-sonnet	Claude Sonnet via Perplexity
/plex-haiku	Claude Haiku via Perplexity
/plex-gpt5	GPT-5 via Perplexity
/plex-gpt54	GPT-5.4 via Perplexity
/plex-gpt55	GPT-5.5 via Perplexity
/plex-grok4	Grok 4 via Perplexity
/plex-gemini	Gemini via Perplexity
/plex-sonar	Perplexity Sonar (web-search)

OpenRouter API

Command	Model
/or-opus	Claude Opus
/or-sonnet	Claude Sonnet
/or-haiku	Claude Haiku
/or-gpt4o	GPT-4o
/or-gpt41	GPT-4.1
/or-o3	o3
/or-gemini	Gemini 2.5 Pro
/or-gemini-flash	Gemini 2.5 Flash
/or-grok3	Grok 3
/or-deepseek	DeepSeek R1
/or-deepseek-v3	DeepSeek V3
/or-llama4	Llama 4 Maverick

LM Studio

Command	What it does
/lms	Switch to the currently active model in LM Studio
/lms-models	List all models loaded in LM Studio
/lms-status	Show LM Studio connection status
/lms-use <model-id>	Switch to a specific LM Studio model by ID

Utility commands

Command	What it does
/cli-back	Restore the model active before the last /cli-* switch
/cli-apply	Apply a staged model switch
/cli-pending	Show the currently staged switch (if any)
/cli-test [model]	Health check the proxy without switching your active model
/cli-list	List all registered models grouped by provider
/cli-help	Full reference card with sections, expiry info, examples, and dashboard link

/cli-back details

Before every /cli-* switch, the current model is saved to ~/.openclaw/cli-bridge-state.json. /cli-back reads that file, restores the previous model, then clears the file. State survives gateway restarts.

/cli-test details

Accepts a short form (cli-sonnet) or full path (vllm/cli-claude/claude-sonnet-4-6). Default when no argument is given: cli-claude/claude-sonnet-4-6. Reports response content, latency, and confirms your active model is unchanged.

Phase 4 -- Web Browser Providers

Routes requests through real browser sessions on the provider's web UI. Uses persistent Chromium profiles -- sessions survive gateway restarts. No API key required.

The first time you use a provider, run /xxx-login to authenticate. If headless login fails, a headed browser opens for manual login (5-minute timeout). After login, cookies are saved to a persistent profile.

On gateway restart, sessions are automatically restored from saved profiles (~25 seconds after start).

Grok (grok.com -- SuperGrok subscription)

Model	Notes
web-grok/grok-4	Full model
web-grok/grok-3	Previous generation
web-grok/grok-3-fast	Faster variant
web-grok/grok-3-mini	Lightweight
web-grok/grok-3-mini-fast	Fastest

Command	What it does
/grok-login	Authenticate via X.com OAuth, save session to ~/.openclaw/grok-profile/
/grok-status	Show session validity and cookie expiry
/grok-logout	Clear session

Gemini (gemini.google.com)

Model	Notes
web-gemini/gemini-3-1-pro	Gemini 3.1 Pro
web-gemini/gemini-2-5-pro	Gemini 2.5 Pro
web-gemini/gemini-2-5-flash	Gemini 2.5 Flash
web-gemini/gemini-3-pro	Gemini 3 Pro
web-gemini/gemini-3-flash	Gemini 3 Flash

Command	What it does
/gemini-login	Authenticate, save cookies to ~/.openclaw/gemini-profile/
/gemini-status	Show session validity and cookie expiry
/gemini-logout	Clear session

Claude.ai

Command	What it does
/claude-login	Authenticate, save session to ~/.openclaw/claude-profile/
/claude-status	Show session validity
/claude-logout	Clear session

ChatGPT

Command	What it does
/chatgpt-login	Authenticate, save session to ~/.openclaw/chatgpt-profile/
/chatgpt-status	Show session validity
/chatgpt-logout	Clear session

All providers at a glance

/bridge-status

Shows all four providers with login state and cookie expiry in one view.

Status dashboard

Live HTML dashboard at: http://127.0.0.1:31337/status

Auto-refreshes every 10 seconds. Includes provider health, active requests, request log, fallback history, and live log viewer.

Phase 5 -- API Integrations

Pure REST integrations -- no subprocess overhead, no browser.

Perplexity API

Set PERPLEXITY_API_KEY in ~/.openclaw/.env. Provides access to 35 models across OpenAI, Anthropic, Google, xAI, NVIDIA, and Perplexity-native providers.

Notable model: plex-sonar uses Perplexity Sonar for web-search-augmented responses. Recommended as the default research phase model in /pipeline.

OpenRouter API

Set OPENROUTER_API_KEY in ~/.openclaw/.env. Provides 17+ models:

• Claude Opus, Sonnet, Haiku
• GPT-4o, GPT-4.1, o3
• Gemini 2.5 Pro, Gemini 2.5 Flash
• Grok 3
• DeepSeek R1, DeepSeek V3
• Llama 4 Maverick

LM Studio

Set LM_STUDIO_URL in ~/.openclaw/.env (default: http://127.0.0.1:1234). Connects to any LM Studio instance on the network. Models are discovered dynamically -- no hardcoded model list required.

/lms              Switch to the currently active LM Studio model
/lms-models       List available models
/lms-status       Show connection status
/lms-use <id>     Switch to a specific model

Phase 6 -- Pipeline

/pipeline runs a multi-phase AI workflow where each phase uses a different model and each phase's output feeds into the next as context.

Phases: research -> architect -> implement -> review

Usage:

/pipeline "topic"
/pipeline "topic" research=plex-sonar architect=cli-fable5 implement=or-sonnet review=or-deepseek
/pipeline "topic" --skip=research,architect implement=or-o3 review=plex-opus

Default model assignments:

Phase	Default model
research	plex-sonar
architect	cli-opus
implement	cli-sonnet
review	plex-gpt55

Model names are the slash-command names from /cli-list without the leading /. Any model reachable via the bridge can be used in any phase.

Configuration

API keys

All API keys go in ~/.openclaw/.env:

PERPLEXITY_API_KEY=pplx-...
OPENROUTER_API_KEY=sk-or-v1-...
LM_STUDIO_URL=http://192.168.177.4:1234   # optional, default: http://127.0.0.1:1234

Plugin config

In ~/.openclaw/openclaw.json under plugins.entries.openclaw-cli-bridge-elvatis.config:

{
  "enableCodex": true,         // register openai-codex from Codex CLI auth (default: true)
  "enableProxy": true,         // start local CLI proxy server (default: true)
  "proxyPort": 31337,          // proxy port (default: 31337)
  "proxyApiKey": "cli-bridge", // key between OpenClaw vllm provider and proxy (default: "cli-bridge")
  "proxyTimeoutMs": 300000,    // base CLI subprocess timeout in ms (default: 300s)
  "modelTimeouts": {           // per-model timeout overrides in ms (optional)
    "cli-claude/claude-fable-5":      360000,
    "cli-claude/claude-sonnet-5":     300000,
    "cli-claude/claude-opus-4-8":     360000,
    "cli-claude/claude-opus-4-7":     360000,
    "cli-claude/claude-opus-4-6":     300000,
    "cli-claude/claude-sonnet-4-6":   180000,
    "cli-claude/claude-haiku-4-5":     90000,
    "cli-gemini/gemini-2.5-pro":      180000,
    "cli-gemini/gemini-2.5-flash":     90000,
    "openai-codex/gpt-5.5":          300000,
    "openai-codex/gpt-5.4":          300000,
    "openai-codex/gpt-5.3-codex":    180000,
    "openai-codex/gpt-5.1-codex-mini": 90000
  }
}

Required: OpenClaw LLM idle timeout

OpenClaw's default llm.idleTimeoutSeconds is 60 seconds, which is too short for CLI subprocesses. Without this setting you will see exit 143 / status:408 / FailoverError: LLM request timed out.

Add to ~/.openclaw/openclaw.json:

{
  "agents": {
    "defaults": {
      "llm": {
        "idleTimeoutSeconds": 300  // must be >= your longest per-model timeout
      }
    }
  }
}

Cron-triggered agents automatically have idleTimeoutSeconds: 0 (disabled) in OpenClaw, so they are not affected.

Setup

1. Install the plugin

The plugin loads from ~/.openclaw/extensions/openclaw-cli-bridge-elvatis/. Enable it in ~/.openclaw/openclaw.json:

{
  "plugins": {
    "entries": {
      "openclaw-cli-bridge-elvatis": { "enabled": true }
    }
  }
}

Or install from ClawHub:

clawhub install openclaw-cli-bridge-elvatis

2. Add API keys

Add keys for the API providers you want to use to ~/.openclaw/.env (see Configuration).

3. Install CLI tools

Install the CLI tools for any CLI providers you want to use:

npm i -g @anthropic-ai/claude-code    # for cli-claude/*
npm i -g @google/gemini-cli           # for cli-gemini/*
# grok, opencode, pi: install per their own docs

4. Restart the gateway

openclaw gateway restart

Check the logs for:

[cli-bridge] proxy ready on :31337
[cli-bridge] registered N commands (use /cli-list to see all)
[cli-bridge] openai-codex provider registered

5. Explore available models

/cli-list

6. Test without switching your model

/cli-test
/cli-test cli-gemini

7. Switch models and restore

/cli-fable5               # stage a switch to Claude Fable 5
/cli-apply                # apply the staged switch
...
/cli-back                 # restore the previous model

Architecture

OpenClaw agent
  |
  +-- openai-codex/*  ---------------------------------> OpenAI API (direct)
  |    auth: ~/.codex/auth.json OAuth tokens
  |
  +-- vllm/cli-claude/*  ---+
  |   vllm/cli-gemini/*  ---+
  |   vllm/cli-grok/*    ---+--> 127.0.0.1:31337 (openclaw-cli-bridge proxy)
  |   vllm/opencode/*    ---+       |
  |   vllm/pi/*          ---+       +-- cli-claude/*  -> claude -p --output-format json --model <m>
  |   vllm/local-bitnet/* -+        |                   stdin=prompt, cwd=homedir()
  |                                 |
  |                                 +-- cli-gemini/*  -> gemini -m <model> -p ""
  |                                 |                   stdin=prompt, cwd=/tmp
  |                                 |
  |                                 +-- cli-grok/*    -> grok --prompt-file <tmpfile>
  |                                 |
  |                                 +-- opencode/*    -> opencode run "prompt"
  |                                 |
  |                                 +-- pi/*          -> pi -p "prompt"
  |                                 |
  |                                 +-- local-bitnet/* -> llama-server on :8082
  |
  +-- perplexity/*  ---------------------------------> Perplexity REST API
  |    key: PERPLEXITY_API_KEY in ~/.openclaw/.env
  |
  +-- openrouter/*  ---------------------------------> OpenRouter REST API
  |    key: OPENROUTER_API_KEY in ~/.openclaw/.env
  |
  +-- lmstudio/*    ---------------------------------> LM Studio REST API
       url: LM_STUDIO_URL in ~/.openclaw/.env

Proxy endpoints:
  /health          -> {"status":"ok"}
  /healthz         -> detailed JSON (version, uptime, provider status, model count)
  /status          -> HTML dashboard (auto-refreshes every 10s)
  /v1/models       -> OpenAI-compatible model list
  /v1/provider-sessions       -> active CLI sessions + stats
  /v1/provider-sessions/:id   -> DELETE to remove a session
  /api/dashboard-data         -> pre-rendered HTML sections (AJAX polling target)
  /api/logs/stream            -> SSE stream of debug.log (real-time)

Staged model switching:
  /cli-<name>       -> stages switch, saves current model to ~/.openclaw/cli-bridge-state.json
  /cli-apply        -> applies staged switch (calls openclaw models set <model>)
  /cli-pending      -> shows staged switch
  /cli-back         -> restores previous model, clears state file

Browser provider profiles:
  ~/.openclaw/grok-profile/      Grok persistent Chromium session
  ~/.openclaw/gemini-profile/    Gemini persistent Chromium session
  ~/.openclaw/claude-profile/    Claude.ai persistent Chromium session
  ~/.openclaw/chatgpt-profile/   ChatGPT persistent Chromium session
  ~/.openclaw/cookie-expiry.json Unified cookie expiry tracking

Metrics and state:
  ~/.openclaw/cli-bridge/debug.log       Request lifecycle logs (1MB rotation, 2 files)
  ~/.openclaw/cli-bridge/metrics.json    Persistent per-model metrics (debounced 5s writes)
  ~/.openclaw/cli-bridge/cli-sessions.json  CLI session registry
  ~/.openclaw/cli-bridge-state.json      Previous model for /cli-back

Requirements

Requirement	Notes
OpenClaw gateway	Tested with 2026.7.x
Node.js 22+	Required
claude CLI	For cli-claude/* models
gemini CLI	For cli-gemini/* models
grok CLI	For cli-grok/* models
opencode CLI	For vllm/opencode/default
pi CLI	For vllm/pi/default
codex CLI	For openai-codex/* (Phase 1 auth bridge)
llama-server on port 8082	For local-bitnet/bitnet-2b
PERPLEXITY_API_KEY	For Perplexity provider
OPENROUTER_API_KEY	For OpenRouter provider
LM Studio running	For LM Studio provider
System Chrome / Chromium	For web browser providers (Phase 4)

CLI tools and API keys are optional -- install only what you plan to use.

Development

npm run lint        # eslint (TypeScript-aware)
npm run typecheck   # tsc --noEmit
npm test            # vitest run
npm run ci          # lint + typecheck + test

Debug log: tail -f ~/.openclaw/cli-bridge/debug.log

Known Issues and Fixes

spawn E2BIG (fixed in v0.2.1)

Symptom: CLI error for cli-claude/...: spawn E2BIG after ~500+ messages.

Cause: Gateway injects large values into process.env at runtime. Spreading it into spawn() exceeds Linux's ARG_MAX (~2MB).

Fix: buildMinimalEnv() passes only HOME, PATH, USER, XDG_RUNTIME_DIR, DBUS_SESSION_BUS_ADDRESS, and required auth keys.

Claude Code 401 / timeout on OAuth login (fixed in v0.2.21)

Symptom: /cli-test cli-claude/* times out after 30s; logs show 401 Invalid authentication credentials.

Cause: buildMinimalEnv() did not forward XDG_RUNTIME_DIR and DBUS_SESSION_BUS_ADDRESS. Claude Code authenticated via claude.ai OAuth (Claude Max) stores tokens in the system keyring (Gnome Keyring / libsecret) and needs these env vars to read them.

Affects: Only systems using claude auth OAuth login. API-key users (ANTHROPIC_API_KEY) are not affected.

Fix: Added XDG_RUNTIME_DIR and DBUS_SESSION_BUS_ADDRESS to buildMinimalEnv().

Gemini agentic mode / hangs (fixed in v0.2.4)

Symptom: Gemini hangs, returns wrong answers, or says "directory does not exist".

Cause: @file syntax triggers Gemini's agentic mode -- it scans the working directory and treats prompts as task instructions.

Fix: Stdin delivery (gemini -p "" with prompt via stdin) and cwd=/tmp.

Exit 143 / status 408 / FailoverError (see Configuration)

Symptom: Requests to CLI models time out with exit 143 or FailoverError: LLM request timed out.

Cause: OpenClaw's default llm.idleTimeoutSeconds is 60s. CLI subprocesses often need more time before producing the first token.

Fix: Set agents.defaults.llm.idleTimeoutSeconds to 300 or higher in ~/.openclaw/openclaw.json. See Configuration.

Changelog

v2026.7.1

• Switched to date-based versioning (YYYY.M.D) aligned with OpenClaw.
• Plugin now compiles against OpenClaw 2026.7.1 SDK (resolved openclaw/plugin-sdk module path).
• Claude Code CLI models re-added: Fable 5, Sonnet 5, Opus 4.8, plus legacy variants.
• Grok CLI added (cli-grok/grok-4.5).
• Perplexity API: 35 models, new slash commands (/plex-*).
• OpenRouter: 17+ models, new slash commands (/or-*).
• LM Studio: dynamic model discovery, /lms* commands.
• Phase 6 pipeline (/pipeline) added.
• Archive notice removed -- project actively maintained.

v3.11.5

• /cli-opus now targets Claude Opus 4.8; /cli-opus47 and /cli-opus46 kept for explicit older routing.
• /cli-gemini3 now targets Gemini 3.1 Pro Preview; /cli-gemini3-pro-preview kept as legacy alias.
• Adds GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro Preview timeout and fallback defaults.
• Test coverage updated for refreshed model timeout and fallback defaults.

v3.11.2

• New /cli-codex55 slash command routes to openai-codex/gpt-5.5.
• Per-model timeouts now cover Opus 4.8, Opus 4.7, and GPT-5.5 (previously missing, fell back to base proxyTimeoutMs).

v3.11.1

• SDK import path fix for OpenClaw 2026.7.1 (openclaw/plugin-sdk/provider-auth-result submodule).

v3.11.0

• Real token usage from Claude CLI (--output-format json): input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens now flow through to OpenClaw's session table.
• Claude Opus 4.7 added (/cli-opus47).
• OpenClaw 2026.5.x compatibility verified.

v3.10.0 -- v3.10.8

• v3.10.8: Include dist/ in npm tarball; openclaw.extensions field for OpenClaw 2026.5.16 compat. Dependency bumps.
• v3.10.6: Pass GEMINI_CLI_TRUST_WORKSPACE=1 to Gemini CLI subprocess.
• v3.10.3: Gemini session resume disabled. Filter Gemini "YOLO mode" warnings.
• v3.10.2: Codex session resume disabled; cwd fixed to homedir().
• v3.10.0: Intelligent prompt routing (src/prompt-router.ts); keyword-based model selection.

v3.9.0 -- v3.9.1

• v3.9.1: Sonnet 60s stale timeout (was 30s). Task payload rescue from subagent {"content":"..."} wrappers.
• v3.9.0: Changelog and documentation update.

v3.8.0

• Root cause fix: Claude CLI must run from homedir(), not from project directories. Running from a project dir triggers agentic mode and prompt injection rejection. This caused the 90% Sonnet failure rate.
• Orchestration test script added (test/orchestration-test.ts).

v3.0.0

• Dashboard v2: sidebar navigation with 9 sections.
• Live log viewer: SSE-powered real-time log streaming.
• AJAX polling replaces full-page meta-refresh (preserves scroll position).

v2.8.0

• Gemini API provider (gemini-api/gemini-2.5-flash, gemini-api/gemini-2.5-pro): direct SDK integration with native image generation support.

v2.6.1

• Startup warning when idleTimeoutSeconds is not set or is under 120s.

v2.5.0

• Graceful timeout handling: SIGTERM -> SIGKILL with 5s grace period.
• Per-model timeout profiles: Opus 5min, Sonnet 3min, Haiku 90s.

v2.3.0

• OpenAI tool calling protocol support for all CLI models.
• Multimodal content support (images/audio from webchat).
• Autonomous execution mode for all CLI runners.

v2.0.0

• Phase 2: local OpenAI-compatible proxy, stdin delivery, prompt truncation.

v1.9.0

• Model fallback chains: gemini-2.5-pro -> Flash, claude-opus -> Sonnet -> Haiku.
• /healthz JSON endpoint.
• Staged model switching (/cli-apply, /cli-pending).

v1.8.0

• BitNet local inference (local-bitnet/bitnet-2b, llama-server on port 8082).
• /bridge-status shows BitNet server health.

v1.6.0

• Persistent Chromium profiles for all four web providers.
• Sessions survive gateway restarts automatically.

v0.1.x

• Phase 1: Codex CLI OAuth auth bridge.

License

MIT