llm-cli-gateway

"Without consultation, plans are frustrated, but with many counselors they succeed." — Proverbs 15:22 (LSB)

Secure local control plane for AI coding agents.

llm-cli-gateway lets supported MCP clients operate Claude Code, Codex, Gemini/Antigravity, Grok Build, Mistral Vibe, Cognition Devin, Cursor Agent, and configured HTTP API providers through one user-owned gateway while preserving native CLI sessions, local credentials, durable async jobs, validation receipts, and review workflows.

Why developers try it: use the client you are already in to delegate work to local coding agents, scope remote execution to registered workspaces, gate risky actions, survive disconnects, and collect auditable review evidence without turning those agents into a generic chat proxy.

Current signals: CI and security workflows pass on main, OpenSSF Scorecard is published, OpenSSF Best Practices is passing, releases use Sigstore signing, and the package is MIT licensed.

Quick Start

npm install -g llm-cli-gateway

Or use directly with npx from an MCP client:

{
  "mcpServers": {
    "llm-gateway": {
      "command": "npx",
      "args": ["-y", "llm-cli-gateway"]
    }
  }
}

Related MCP server: MCP-AI-Gateway

What It Provides Today

llm-cli-gateway is a single-user MCP control plane for operating AI coding agents from supported local or remote clients. It is more than a thin CLI wrapper:

• Runs registered provider CLIs and configured HTTP API providers through consistent sync and async MCP tools.

• Persists long-running jobs, supports restart-safe result collection, deduplication, cancellation, and sync-to-async deferral.

• Tracks sessions, real CLI resume paths, structured response metadata, and cache telemetry.

• Supports cache-aware promptParts, including explicit Claude cache_control when opted in.

• Can run supported provider requests inside gateway-managed git worktrees for isolated multi-agent review and implementation loops when using the local file-backed session manager. PostgreSQL-backed sessions reject this filesystem-local feature before creation. Grok, Devin, and Mistral require an explicit provider-native sessionId for a gateway worktree; fresh, createNewSession, and resumeLatest-only worktree requests fail closed because they cannot durably reselect it. Materialization suppresses repository, system, and global Git hooks, configured clean, smudge, and process checkout filters, sparse checkout, and lazy object fetching. Filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands.

• Ships personal-appliance setup surfaces: HTTP transport with bearer-token auth, doctor --json, setup UI artifacts, provider setup snippets, Docker fallback, and checked release bundles.

• Remote web connectors use MCP OAuth discovery and authorization-code setup with static client or shared-secret gates. Client secrets are generated locally, stored only as hashes, and printed only by explicit copy-once commands.

• Provider CLI requests can select registered workspaces by alias via workspace; every HTTP/tunnel request must use a registered alias, session workspace, or [workspaces].default before provider execution. Local unrestricted filesystem access is the stdio transport.

Workflow Assets

The repo ships agent-ready workflow skills under .agents/skills for async orchestration, session continuity, multi-LLM review, implement-review-fix loops, retrospective evidence walks, secure approval-gated dispatch, and Personal Agent Config Kit operations. Nine caller-facing skills are bundled in the published npm package: async-job-orchestration, multi-llm-review, session-workflow, secure-orchestration, implement-review-fix, retrospective-walk, public-demo-session, least-cost-routing, and personal-agent-config-kit. Machine-readable DAG-TOML plans live under docs/plans and setup/install-plan.dag.toml for workflows that need deterministic sequencing and verification gates.

Skill packs can be updated outside the core npm release by placing skill directories in local, operator-controlled paths. The gateway loads bundled skills first, then [skills].paths, then LLM_GATEWAY_SKILLS_PATH, then ~/.llm-cli-gateway/skills when it exists; later roots override earlier skills by name. Each skill is a directory containing SKILL.md. A root may also carry skill-pack.json to pin expected SKILL.md hashes:

[skills]
paths = ["/opt/llm-cli-gateway/skills"]

export LLM_GATEWAY_SKILLS_PATH="/opt/team-skill-pack:/opt/incident-skill-pack"

{
  "name": "team-pack",
  "version": "1.0.0",
  "skills": [
    {
      "name": "incident-retrospective",
      "sha256": "<sha256 of incident-retrospective/SKILL.md>"
    }
  ]
}

The loader is intentionally local-only: it never fetches remote Markdown at startup. To update a pack, install or replace files through your package manager or deployment system, then restart the gateway so the advertised skills://... resources refresh.

The next documentation focus is provider-specific skill and DAG-TOML pairs for each outbound CLI and API-provider family: Claude, Codex, Gemini/Antigravity, Grok, Mistral Vibe, Devin, Cursor Agent, OpenAI-compatible endpoints, Anthropic Messages, and xAI Responses. The implementation plan is tracked in docs/plans/provider-workflow-assets.dag.toml, with each provider asset expected to cover install/login checks or token-env checks, session behavior, approval modes, cache/telemetry surfaces, failure modes, and a smoke-test gate.

Trust & Supply Chain

• CI runs build, lint, format, tests, package checks, and npm audit.

• Security CI runs actionlint, zizmor, shellcheck, typos, osv-scanner, gitleaks, and lychee.

• GitHub release installer artifacts are checksummed and signed with Sigstore keyless signing.

• npm releases use a generated prod-only shrinkwrap and release security audit; GitHub Actions Trusted Publishing exchanges the job's OIDC identity for short-lived npm publish credentials.

• The npm package intentionally ships a generated, prod-only npm-shrinkwrap.json so registry installs resolve the audited release tree. Release gates regenerate it from package-lock.json, compare for parity, and run a registry-fidelity consumer install before publishing.

• Socket behavioural alerts are documented in socket.yml and under "Security Considerations" below. shellAccess and shrinkwrap are reviewed package capabilities/configuration for this CLI appliance, not hidden install behaviour.

Personal MCP Appliance

The personal-appliance contract keeps that surface intentionally narrow: one trusted user runs the gateway on a machine or volume they own, connects one MCP endpoint, and lets supported clients operate local coding agents through workspace-scoped, approval-gated, auditable requests.

The product contract is documented in docs/personal-mcp/PRODUCT_CONTRACT.md. It defines the single-user scope, security posture, target support matrix, and provider-support verification gates. Public setup guides must not claim ChatGPT, Claude web, Claude Desktop, Codex, Gemini CLI, Gemini web, or Grok inbound support until the corresponding provider/client path has been verified.

This project does not provide hosted multi-tenant credential custody. Provider credentials stay on the user's machine or user-owned deployment volume.

For a single developer who works on several workstations and repositories, the Personal Agent Config Kit guide explains the Git-synchronised personal baseline, repository overlays, immutable context stamps, and workstation-safe provider continuity model. The baseline directory is confined to a non-symlinked descendant of the local home directory, and every publish or sync revalidates its configured origin fetch and push URLs against the HTTPS/SSH-only policy. The Kit is intentionally local-caller-only; Kit provider execution and recovery of an unadmitted attempt require healthy durable SQLite or PostgreSQL async-job admission. Kit mode supports Claude, Codex, and Mistral; every other provider fails closed while it is enabled. Kit scope never inherits the gateway process cwd: Claude and Mistral require a registered workspace selection or configured default, while Codex can also use an explicit absolute workingDir. Relative Kit workingDir values are rejected before filesystem or Git inspection. It disables cross-model validation tools while enabled and is not an HTTP/OAuth configuration or remote execution feature.

For a retention-pinned non-Kit Claude MCP request configuration, follow the local-only same-host recovery procedure. It accepts no SQL or arbitrary-path override. A valid recovery invocation emits JSON and exits 0 after success or 2 for a safe refusal; invalid usage or unavailable durable storage exits 1.

Release-readiness history is tracked in docs/personal-mcp/RELEASE_READINESS.md. Dogfooding evidence (which target LLMs guided setup, what unsafe suggestions were captured, and which findings were deferred from the initial personal-appliance rollout) is in docs/personal-mcp/DOGFOODING_RESULTS.md.

Current personal-appliance artifacts include:

• Streamable HTTP startup: LLM_GATEWAY_AUTH_TOKEN=<token> npm run start:http

• Machine-readable diagnostics: npm run doctor

• Go bootstrapper: installer/ with setup, doctor --json, start, stop, status, repair, upgrade, uninstall, print-client-config, and verified bundle download commands.

• Release packaging: on the public mirror, the release workflow builds Linux binaries on GitHub-hosted ubuntu-latest and builds Windows/macOS binaries on their GitHub-hosted platform runners. The private upstream uses its internal self-hosted Linux runner. Each release publishes checksummed platform bundles with the gateway, production dependencies, and a managed Node runtime; see installer/packaging/README.md.

• Docker Compose fallback: docker/personal.compose.yml + docker/Dockerfile.personal for users who already manage containers.

• Local setup UI artifact: setup/ui/index.html

• Provider setup snippets: setup/providers/

• Cross-validation tools: review_changes, validate_with_models, second_opinion, compare_answers, red_team_review, consensus_check, ask_model, synthesize_validation, job_status, job_result, and validation_receipt (plus the validation-receipt://{validationId} resource). review_changes and durable receipts require a SQLite or PostgreSQL validation-run store and are absent in Personal Agent Config Kit mode.

Install / Upgrade / Uninstall (single binary)

Windows PowerShell:

$Version = '<version>'
$Base = "https://github.com/verivus-oss/llm-cli-gateway/releases/download/v$Version"
$InstallDir = Join-Path (Join-Path $env:LOCALAPPDATA 'Programs') 'llm-cli-gateway'
$ExeName = "llm-cli-gateway-$Version-windows-amd64.exe"
$BundleName = "llm-cli-gateway-bundle-$Version-windows-amd64.tar.gz"
$Exe = Join-Path $InstallDir 'llm-cli-gateway.exe'
$Checksums = Join-Path $InstallDir 'SHA256SUMS'
$ChecksumBundle = Join-Path $InstallDir 'SHA256SUMS.sigstore.json'
New-Item -ItemType Directory -Force $InstallDir | Out-Null
Invoke-WebRequest -UseBasicParsing "$Base/$ExeName" -OutFile $Exe
Invoke-WebRequest -UseBasicParsing "$Base/SHA256SUMS" -OutFile $Checksums
Invoke-WebRequest -UseBasicParsing "$Base/SHA256SUMS.sigstore.json" -OutFile $ChecksumBundle
cosign verify-blob $Checksums --bundle $ChecksumBundle --certificate-identity "https://github.com/verivus-oss/llm-cli-gateway/.github/workflows/release-installer.yml@refs/tags/v$Version" --certificate-oidc-issuer "https://token.actions.githubusercontent.com"
if ($LASTEXITCODE -ne 0) { throw "Sigstore verification failed for SHA256SUMS" }
function Get-ReleaseSha256($Name) {
  $line = Select-String -Path $Checksums -Pattern "^[a-fA-F0-9]{64}\s+$([regex]::Escape($Name))$" | Select-Object -First 1
  if (-not $line) { throw "No SHA256SUMS entry found for $Name" }
  return (($line.Line -split "\s+")[0]).ToLowerInvariant()
}
if ((Get-FileHash $Exe -Algorithm SHA256).Hash.ToLowerInvariant() -ne (Get-ReleaseSha256 $ExeName)) { throw "Checksum mismatch for $ExeName" }
$env:RVWR_GATEWAY_BUNDLE_URL = "$Base/$BundleName"
$env:RVWR_GATEWAY_BUNDLE_SHA256 = Get-ReleaseSha256 $BundleName
& $Exe setup
& $Exe stop
& $Exe install-bundle
& $Exe start
& $Exe status
& $Exe doctor

The Windows installer keeps a stable llm-cli-gateway.exe command in %LOCALAPPDATA%\Programs\llm-cli-gateway and adds that directory to the user PATH. Do not script against release-versioned exe names after install.

# After downloading the binary that matches your OS/arch from a release:
cosign verify-blob SHA256SUMS --bundle SHA256SUMS.sigstore.json \
  --certificate-identity "https://github.com/verivus-oss/llm-cli-gateway/.github/workflows/release-installer.yml@refs/tags/v<version>" \
  --certificate-oidc-issuer "https://token.actions.githubusercontent.com"
sha256sum --check SHA256SUMS            # verify before run (or `shasum -a 256 --check` on macOS)
chmod +x llm-cli-gateway-<ver>-<os>-<arch>
./llm-cli-gateway-<ver>-<os>-<arch> setup
./llm-cli-gateway-<ver>-<os>-<arch> install-bundle    # uses the platform bundle URL/SHA256
./llm-cli-gateway-<ver>-<os>-<arch> start
./llm-cli-gateway-<ver>-<os>-<arch> doctor

# Upgrade: replace the binary, set the new bundle env vars, run upgrade.
./llm-cli-gateway-<new>-<os>-<arch> upgrade

# Uninstall: dry-run first, then run with --yes.
./llm-cli-gateway-<ver>-<os>-<arch> uninstall
./llm-cli-gateway-<ver>-<os>-<arch> uninstall --yes

Docker fallback:

LLM_GATEWAY_AUTH_TOKEN=$(openssl rand -hex 32) \
  docker compose -f docker/personal.compose.yml up -d
docker compose -f docker/personal.compose.yml run --rm doctor

Features

Core Capabilities

• Multi-LLM Orchestration: Unified interface for Claude Code, Codex, Gemini, Grok, Mistral (Vibe), Devin, and Cursor Agent CLIs

• Session Management: Track gateway session metadata and provider-specific continuity with persistent storage

• Gateway-owned worktrees: Run supported sync or async provider requests inside a managed git worktree with the local file-backed session manager. Same-session reuse requires same-host durable ownership plus a matching live Git registration and gateway branch; manager-level named path collisions fail closed. PostgreSQL-backed sessions reject worktrees before creation because a different database-connected host cannot safely own filesystem cleanup. Grok, Devin, and Mistral require an explicit provider-native sessionId; fresh, createNewSession, and resumeLatest-only worktree requests fail closed. A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks, configured clean, smudge, and process checkout filters, sparse checkout, and lazy object fetching. Filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands. Session deletion and TTL eviction hide a durably owned worktree session while cleanup runs. If Git removal fails, the file-backed store retains a durable cleanup-pending tombstone, blocks reuse, and retries cleanup when that store is registered on the owning host. The record is finalized only after verified Git removal.

• Token Optimization: opt-in prompt/response compaction via optimizePrompt / optimizeResponse, both defaulting to false. The 44% / 37% figures are v1.0.0 measurements on a specific corpus, not a guarantee: src/optimizer.ts is a deterministic phrase-stripper, and nothing measures a ratio at runtime.

• Correlation ID Tracking: Full request tracing across all LLM interactions

• Cross-Tool Collaboration: LLMs can use each other via MCP (validated through dogfooding)

Observability

• Durable Flight Recorder: Provider requests and responses follow [persistence].backend: SQLite writes to ~/.llm-cli-gateway/logs.db by default, while PostgreSQL writes them to the configured database with the other durable subsystems. A PostgreSQL failure leaves the recorder unavailable or degraded and is reported generically on health surfaces; it never falls back to SQLite. Existing rows are not migrated or dual-read when the backend changes. Records include correlation IDs, duration, and token usage where the provider emits it (today: claude on stream-json/json, codex, and the API providers; grok, gemini, mistral, devin and cursor emit no usage on their CLI wire). Two exclusions are worth knowing: cross-LLM validation seats write no flight-recorder row at all, so llm_request_result cannot read one back by correlation ID, and repository-review seats are additionally excluded by design so review evidence is not retained in a non-expiring table. The retry_count and circuit_breaker_state columns are constants on the async path, which is the production path; retry and circuit breaking apply only to the direct-execute fallback. Read history through llm_request_list and llm_request_result so per-principal ownership checks remain in force. For human browsing of a SQLite deployment: datasette ~/.llm-cli-gateway/logs.db

• Structured Metadata: Tool responses include machine-readable structuredContent (model, cli, correlationId, sessionId, durationMs, token counts)

• Cache observability resources: cache-state://global, cache-state://session/{id}, and cache-state://prefix/{hash} MCP resources return aggregate cache hit/miss/savings — tokens and hashes only, no prompt text. session_get includes a cacheState block when the session has prior requests.

• Provider capability inventory: provider_tool_capabilities and provider-tools://catalog expose the gateway request fields, supported/degraded provider controls, local skill/tool discovery, and safe config-surface hints for Claude Code, Codex CLI, Gemini/Antigravity, Grok CLI/API, Mistral Vibe, Cognition Devin, and Cursor Agent. doctor --json includes a compact provider_capabilities summary for setup assistants.

Cache-aware operation

Every *_request and *_request_async tool except devin_request / devin_request_async and cursor_request / cursor_request_async accepts an optional promptParts field that structures the prompt for better cache hit rates (the Devin and Cursor headless paths take a plain prompt only). The gateway concatenates the parts in canonical order (system → tools → context → task) so that the stable prefix bytes precede the volatile task tail unchanged across calls, letting each provider's automatic prompt-caching land on the same content hash each time.

{
  "promptParts": {
    "system": "You are a helpful code reviewer.",
    "tools": "You have access to Read, Grep, Bash.",
    "context": "<long stable context block — file dumps, etc.>",
    "task": "Review the changes in src/foo.ts for security issues."
  }
}

prompt and promptParts are mutually exclusive — pass exactly one.

Per-CLI capability matrix (prefix discipline is automatic via promptParts for all providers except Devin and Cursor, which have no promptParts surface; explicit levers are provider-specific):

Claude example (explicit cacheControl)

claude_request({
  promptParts: {
    system: "You are a helpful code reviewer.",
    context: "<long stable file dump>",
    task: "Review the diff.",
    cacheControl: { system: true, context: true }, // task is never marked
  },
  outputFormat: "stream-json",
});

Gateway emits the stream-json stdin path with cache_control: {type:"ephemeral", ttl:"1h"} on marked blocks only.

Grok example (compaction)

grok_request({
  promptParts: { system: "...", context: "...", task: "..." },
  compactionMode: "segments",
  compactionDetail: "balanced",
});

Emits --compaction-mode segments --compaction-detail balanced.

See docs/personal-mcp/PROVIDER_CACHE_SURFACES.md for full surfaces, telemetry differences (e.g. Grok -p vs ACP), exact stream-json payload shapes, and cross-LLM review notes.

Opt-in flags (all default off) live under [cache_awareness] in ~/.llm-cli-gateway/config.toml.

Reliability & Performance

• Retry Logic: Exponential backoff with circuit breaker for transient failures

• Atomic File Writes: Process-specific temp files with fsync for data integrity

• Host-protection backpressure: bounded HTTP session lifecycle (max sessions + idle reaper), global and per-provider job-execution limits with a bounded FIFO queue, and a configurable per-job output cap (default 50MB). See Host-protection limits.

• NVM Path Caching: Eliminates I/O overhead on every request

• Long-Running Jobs: Non-time-bound async execution via *_request_async + polling tools

Security & Quality

• Comprehensive Testing: 1,700+ tests covering unit, integration, and regression scenarios with real CLI execution

• Input Validation: Zod schemas prevent injection attacks

• No Secret Leakage: Generic session descriptions only (file permissions 0o600)

• No ReDoS: Bounded regex patterns prevent catastrophic backtracking

• Type Safety: Strict TypeScript with comprehensive error handling

• Supply-chain hardening: a dedicated .github/workflows/security.yml runs actionlint, zizmor, shellcheck, typos, osv-scanner, gitleaks, and lychee on every push and PR (see SECURITY.md for the threat model)

Provider capability surface

Every provider is reachable through the same request, session, job, and validation machinery, but the underlying CLIs differ in what they natively expose. The table records what actually shipped per provider; discover the live surface at runtime with provider_tool_capabilities, list_models, and the provider-acp://<provider> / provider-tools://<provider> resources.

• Native ACP is reported honestly. grok, mistral, devin, and cursor expose a native ACP entrypoint, so provider-acp://<provider> carries the negotiated initialize capability set and the derived session-method availability, and the sync *_request accepts transport: "acp" (fails closed unless [acp] and the provider's runtime_enabled gate are set). ACP routing is sync-only: the *_request_async variants always run the CLI transport and do not accept transport: "acp" (nor Devin's agentType); async ACP parity is a later phase. ACP workspace selection is gateway-owned: an explicit ACP workspace must be a registered alias. A fresh remote ACP request uses that alias or [workspaces].default; a remote resume is fixed to its recorded canonical alias and cwd, and a different or unbound workspace is rejected. Local ACP may omit workspace; each unscoped process then gets a fresh private 0o700 neutral directory that is removed after the process exits, never a shared predictable temp path. claude, codex, and gemini have no native ACP entrypoint at their target CLI versions; their provider-acp:// records report native: false with no methods and no adapter-as-native masquerade, and they expose no transport: "acp" selector.

• Managed approval is Claude-only today. approvalStrategy:"mcp_managed" is executable only by the Claude CLI adapter, which launches Claude with a request-scoped generated MCP configuration and --strict-mcp-config. It permits only provisioned, gateway-owned MCP definitions and rejects dynamic npx, ambient-PATH, and Codex-config overrides. Codex, Gemini, Grok, Mistral, Devin, and Cursor reject mcp_managed before launching a provider because their current adapters cannot isolate ambient MCP configuration. For those adapters, use approvalStrategy:"legacy"; approvalPolicy has no effect.

• ACP has its own permission bridge. approvalStrategy:"mcp_managed" and any approvalPolicy are rejected when transport:"acp" is selected. Use the Claude CLI transport for managed approval. ACP host services fail closed, gated by the tool-call category the agent declares: write is denied unless allow_write_host_services is set, execute unless allow_terminal_host_services is set, and any other category is denied outright, including fetch and every unrecognised kind, so a new upstream tool kind cannot be auto-approved by default. Approval is expressed only by selecting an agent-offered single-use allow option; an agent that offers only a persistent allow_always grant is denied, so ACP never maps a raw CLI bypass input to a standing permission grant. Note the boundary is this category gate: it constrains what the agent may ask the gateway to do, not what the agent does in its own process.

• Resources are generated from the provider registry for every CLI provider: models://<provider>, sessions://<provider>, provider-acp://<provider>, provider-tools://<provider>, and provider-subcommands://<provider>.

• Model discovery is live and account-aware: the discovery listed above reaches models://<provider> and list_models, degrading to static registry facts when a live probe is unavailable (a resource read never spawns a CLI).

• Admin surfaces are discovery-driven and output-redacted. provider_admin_list and provider_admin_run are read-only for every provider. State-mutating admin operations are exposed only through provider_admin_mutate, gated behind [admin] allow_mutating_cli_admin_ops, the remote cli:admin scope, an approval gate, and an audit record. Mutating ACP session operations are likewise gated behind [acp] allow_mutating_session_ops.

• Validation commands work across every provider: review_changes, validate_with_models, second_opinion, compare_answers, red_team_review, consensus_check, ask_model, and synthesize_validation, with canonically hashed immutable receipts via validation_receipt and the validation-receipt://{validationId} resource. review_changes captures a complete, hashed Git artifact and starts repository-bound read-only reviewers.

Prerequisites

Node.js >= 24.4.0 is required (engines.node in package.json). The SQLite backend uses Node's built-in node:sqlite module, so there is no native binding to compile and no install scripts run. The 24.4 floor is where allowBareNamedParameters defaults to true, which the SQLite persistence layer relies on.

Running the source-tree release audit also requires Bash and flock from util-linux. This prerequisite was verified against commit 242e7669565ecc7c71183f5f7133791a1384ca7c. Release automation runs the audit on Ubuntu. On macOS, install a compatible flock before running npm run security:audit or the full npm run check gate.

Before using this gateway, you need to install the CLI tools you want to use:

Claude Code CLI

# Installation instructions for Claude Code
# Visit: https://docs.anthropic.com/claude-code
npm install -g @anthropic-ai/claude-code

Codex CLI

npm install -g @openai/codex
codex login

Gemini (Google Antigravity CLI)

The Gemini provider runs through Google Antigravity CLI (agy).

curl -fsSL https://antigravity.google/cli/install.sh | bash
# Docs: https://antigravity.google/docs

Grok Build CLI (xAI)

curl -fsSL https://x.ai/cli/install.sh | bash
grok login   # OAuth flow; for headless auth, set XAI_API_KEY
# Docs: https://docs.x.ai/build/overview

Mistral Vibe CLI

# Pick one — the gateway's cli_upgrade auto-detects which one you used.
curl -LsSf https://mistral.ai/vibe/install.sh | bash
pip install mistral-vibe
uv tool install mistral-vibe
brew install mistral-vibe

vibe --setup
# Complete the API-key setup locally. Do not paste the key into a chat.
# Current Vibe defaults session logging to enabled. If an older config disabled it,
# edit ~/.vibe/config.toml and set:
# [session_logging]
# enabled = true

Vibe-specific notes:

• Model selection is via the VIBE_ACTIVE_MODEL environment variable — Vibe has no --model flag. The gateway discovers ~/.vibe/config.toml / VIBE_MODELS, injects VIBE_ACTIVE_MODEL only when a model is explicitly requested or Vibe config needs recovery, and retries once after a model-not-found failure with refreshed discovery.

• permissionMode is the Vibe --agent name. Builtins are default | plan | accept-edits | auto-approve; Vibe also accepts install-gated builtins (e.g. lean) and custom agents from ~/.vibe/agents. Requests pass the selected name through for Vibe to validate. mcp_managed is not available for Vibe.

• Tool controls use Vibe's native flags. The gateway emits one --enabled-tools <tool> flag per allowedTools entry and one --disabled-tools <tool> flag per disallowedTools entry. Vibe applies disabled tools after enabled-tool filtering.

• Usage telemetry is best-effort. Vibe does not emit token or cost data in programmatic stdout. When the gateway knows Vibe's native session UUID, it reads ~/.vibe/logs/session/session_<...>/meta.json for usage and cost. A missing, malformed, or not-yet-known session log leaves those fields empty.

• No self-update: cli_upgrade --cli mistral detects whether you used pip / uv / brew and dispatches the matching upgrade command. Running vibe update is not a thing.

Installation

As an MCP server (npm)

npm install -g llm-cli-gateway

Or use directly with npx:

{
  "mcpServers": {
    "llm-gateway": {
      "command": "npx",
      "args": ["-y", "llm-cli-gateway"]
    }
  }
}

From source

git clone https://github.com/verivus-oss/llm-cli-gateway.git
cd llm-cli-gateway
npm install
npm run build

Usage

As an MCP Server

For clients that already support local stdio MCP servers, add a configuration like:

{
  "mcpServers": {
    "llm-cli-gateway": {
      "command": "node",
      "args": ["/path/to/llm-cli-gateway/dist/index.js"]
    }
  }
}

Stdio is the recommended path for unrestricted machine-local development access. HTTP MCP, including localhost HTTP and tunneled HTTPS, is treated as remote-capable for provider execution: provider tools must resolve a registered workspace alias, a session workspace, or [workspaces].default before spawning a CLI. Remote clients should pass relative workingDir, addDir, and include-directory values inside the selected workspace, and may resume only gateway-tracked sessions they own. Raw native provider session IDs are local-only. Disabling auth or using a no-auth connector path is not a filesystem bypass.

For a local CLI request with no resolved workingDir, registered workspace, or gateway-managed worktree, the child runs in a fresh private 0o700 temporary directory that is removed after the process exits. It never inherits the gateway repository cwd or its provider-native instruction context. The gateway canonicalizes the temp root and rejects or relocates it when any ancestor contains .git, AGENTS.md, AGENTS.override.md, Agents.md, AGENT.md, CLAUDE.md, Claude.md, CLAUDE.local.md, .claude/CLAUDE.md, .claude/rules/, .cursor/rules/, .cursorrules, GEMINI.md, or .vibe/config.toml, including through a symlinked or custom TMPDIR beneath that context. The list covers entries a provider discovers by walking up from its cwd. User-scope configuration such as ~/.claude/settings.json is deliberately absent: it loads on every invocation regardless of cwd, so relocating the workspace would not isolate it. Provider-native resumeLatest operations that use a cwd-scoped latest-session pointer therefore require an explicit workingDir, workspace, or configured default workspace and fail closed when none is available. Use explicit target selection whenever several repositories are active at once.

CLI request schemas accept prompts up to 100,000 characters, but operating systems also impose byte limits on individual argv elements. Codex new and resume requests stream the exact prompt over stdin. codex_fork_session remains argv-bound and rejects an oversized UTF-8 prompt before spawn as non-retryable input_too_large. Other providers whose current CLI contracts require an argv prompt use the same admission rule. Every other caller-controlled argv value is checked on its final encoded form too, including serialized agent/schema JSON, joined tool lists, instruction overrides, paths, model names, and native session IDs. The final spawn boundary checks every argv element plus the aggregate resolved command line against a conservative platform-specific byte budget and a 2,048-element cap. The aggregate byte budget excludes the environment but reserves headroom for it; on Windows, pre-resolution admission assumes the smaller npm .cmd/.bat wrapper limit until command resolution proves a native executable. Native session and resume flags on non-Kit requests are included before workspace, session, provider-artifact handoff, or durable job side effects. Claude Kit projects its eventual argv before materializing its compiled context artifact or allocating a durable Kit session. An embedded NUL byte in the command or any argv element is rejected before spawn as non-retryable invalid_input. Caller-facing results, long-lived job memory, durable job args, and async flight rows use a fixed invalid-argv marker; the optional duplicate durable payload is suppressed. None retains the rejected vector or Node's value-echoing native error. Native E2BIG, including an environment-driven failure, is normalized without retaining the native spawnargs. The gateway never truncates instructions or other values to make them fit. For stdin-backed requests, a clean provider exit is accepted only after the complete payload write callback succeeds. A closed or still-pending pipe becomes a fixed, non-sensitive incomplete-delivery failure; timeout, cancellation, and provider nonzero exits remain authoritative.

This generic stdio example is not provider-support verification for the Personal MCP Appliance. Client-specific setup guides for ChatGPT, Claude web, Claude Desktop, Codex, Gemini CLI, Gemini web, and Grok remain gated by the provider-support matrix in docs/personal-mcp/PRODUCT_CONTRACT.md.

Available Tools

Cross-LLM Validation Tools

The personal-appliance surface exposes simplified validation tools for non-developer clients. These tools start provider CLI jobs through the durable async job manager and return normalized provider status plus raw job references.

• validate_with_models: ask two or more providers to independently validate a question.

• review_changes: capture one complete Git review artifact, fence repository content as untrusted data, and start read-only independent reviewers. See Repository change review.

• second_opinion: ask one provider to review an answer.

• red_team_review: challenge a plan, answer, or document for risks and failure modes.

• consensus_check: check whether providers agree with a claim.

• ask_model: ask one provider through the simplified surface.

• synthesize_validation: run an explicit judge model after provider results have been collected. General validation requires the caller's question and terminal normalized results. A review_changes run instead reloads its exact owned durable results from validationId; caller-supplied question/results are ignored.

• list_available_models: list the models each provider CLI exposes through the simplified surface.

• job_status and job_result: poll and collect validation job outputs.

• validation_receipt: retrieve the canonically hashed immutable receipt of a terminal cross-LLM validation run by validationId (returns minted | pending | verification_failed | expired_unminted | not_found, own-or-not-found). verification_failed means a stored receipt exists but disagrees with its durable run, which is a defect to investigate; expired_unminted only ever means absence. format: "markdown" renders a human-readable report; includeRawResponses inlines complete provider answer text when the linked job still exposes identity-verified output. Registered only when the attached job store provides the durable validation-run store capability (sqlite and postgres).

The same receipt is also exposed as the validation-receipt://{validationId} MCP resource (same durable gate and own-or-not-found owner scoping).

The validation report preserves per-provider disagreement. Optional judge synthesis is explicit about which provider produced the judge job.

Repository change review

review_changes accepts an absolute local workingDir or a registered workspace, then resolves scope: "auto" | "uncommitted" | "branch" | "commit". It can take an explicit Git base, literal repository-relative paths, stance: "standard" | "adversarial", reviewer models, an optional judgeModel, trustCursorWorkspace, and fail-closed artifact/prompt byte ceilings.

Cursor refuses to review a directory it does not trust, so a cursor seat is granted --trust only when the reviewed directory is a registered [[workspaces.repos]] path whose providers include cursor (a gateway worktree beneath one counts, and the nearest enclosing registration decides). On any other directory the cursor seat is skipped with an actionable reason, and the rest of the roster still runs. trustCursorWorkspace: true accepts the grant for one durable review run. The consent applies to Cursor seats in that run's reviewer roster and to its planned Cursor judge when synthesize_validation launches it later. It is bound to the run owner, repository, and planned judge, so another principal or review run cannot replay it. That is a real decision rather than a formality: a trusted folder is also where cursor loads project rules and AGENTS.md, so the repository under review gains some influence over its own reviewer, which the fenced review prompt otherwise forbids. The artifact keeps committed, staged, unstaged, and regular non-ignored untracked file evidence separate. It forces tracked diffs to remain readable even when in-tree attributes mark them as non-diffable. The review-evidence.v2 artifact exposes committedPatch, stagedPatch, and unstagedPatch independently; each segment carries its sorted path inventory, encoding, exact byte length, SHA-256 identity, and content. This prevents an index change and its worktree-only reversal from canceling out. The artifact is collision-fenced, byte-counted, SHA-256 identified, race-checked, and never truncated. In auto mode, a diverged branch is reviewed from its merge base with working-tree evidence included. Otherwise, a dirty tree selects uncommitted changes, while a clean tree falls back to the last commit (HEAD^..HEAD) without working-tree evidence. Unsafe untracked file types or a repository mutation during capture cause a refusal.

The tool starts asynchronous provider jobs and returns a validationId, exact artifact and prompt identities, file inventory, and one rawJobReference per reviewer. Poll those references with validation job_status and collect them with validation job_result, not the similarly named llm_job_* tools. If a judge was requested, wait for every reviewer to become terminal, then call synthesize_validation with the validationId and the same workingDir or workspace selector. Continue collecting results for progress and human visibility, but do not pass them as review evidence: for a review_changes run, the gateway ignores caller-supplied question and providerResults, reloads the exact owned durable linked terminal jobs, and reconstructs requested but unavailable seats as skipped. General validation synthesis still requires a caller-supplied question and terminal normalized results.

The review surface is registered only with durable SQLite or PostgreSQL job and validation-run storage. Each CLI review job retains the exact fenced prompt in its expiry-bound payload_json; its persisted argv contains only a hash marker. The non-expiring flight recorder does not receive repository-review prompts. Configured HTTP/API reviewer seats require explicit allowApiUpload:true because the complete artifact leaves the local CLI boundary. Remote HTTP/OAuth workspace reviews reject API reviewer uploads even with that flag. Treat the durable job store as sensitive. Its retention is unbounded by default, and an operator can set [persistence.retention].jobs when a bounded record is required. When judgeModel is an HTTP/API provider, review_changes binds that explicit consent, the judge provider, the resolved repository, and the caller identity to the durable validationId. The later synthesize_validation call must provide that id and the same repository selector. The stored judge, repository, owner, and upload consent are authoritative. The gateway atomically claims the planned judge once, so concurrent or repeated synthesis cannot start a second judge. A follow-up argument cannot grant or override upload consent.

LLM Request Tools

claude_request

Execute a Claude Code request with optional session management.

Parameters:

• prompt (string, optional*): The prompt to send (1-100,000 chars). *Exactly one of prompt or promptParts is required (mutually exclusive)

• model (string, optional): Model name or alias (use list_models for available values; supports latest)

• outputFormat (string, optional): Output format (text|json|stream-json), default: stream-json — the gateway parses NDJSON usage events for token/cost observability; override to text only when you want unparsed stdout

• sessionId (string, optional): Specific session ID to use. Under mcp_managed, native continuation is a high-risk input because it can inherit an unverified provider posture; it requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1, but does not select a full-permission profile.

• continueSession (boolean, optional): Continue the active session. It has the same managed-approval requirement as sessionId. Because Claude --continue selects by cwd, it requires workingDir or a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default. That workspace may optionally supply a gateway worktree. The request fails closed when no selection supplies a stable cwd.

• createNewSession (boolean, optional): Always create a new session

• forkSession (boolean, optional): Fork the resumed session instead of appending to it. Under mcp_managed, it is a high-risk native-fork input that requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1, but stays bounded.

• allowedTools (string[], optional): Restrict Claude tools to this allow-list. A non-empty allow-list is a high-risk managed input because it can change the tool posture.

• disallowedTools (string[], optional): Explicitly deny listed Claude tools

• permissionMode (string, optional): Claude permission mode (default|acceptEdits|plan|auto|dontAsk|bypassPermissions); preferred over dangerouslySkipPermissions. bypassPermissions is a direct full-permission request under mcp_managed.

• dangerouslySkipPermissions (boolean, optional): Deprecated, maps to permissionMode: "bypassPermissions"; permissionMode wins when both are set. It is a direct full-permission request under mcp_managed.

• agent (string, optional): Named sub-agent to run as. A non-empty value is a high-risk managed input because it can change tool and permission posture.

• agents (string, optional): Inline agent definitions JSON. A non-empty value is a high-risk managed input for the same reason.

• systemPrompt / appendSystemPrompt (string, optional): Replace or extend the system prompt. A non-empty value is a high-risk managed input.

• systemPromptFile / appendSystemPromptFile (string, optional): Replace or extend the system prompt from a file. A non-empty file path is a high-risk managed input.

• safeMode (boolean, optional): Start Claude with local customizations disabled, including CLAUDE.md, skills, plugins, hooks, MCP, commands, and agents. true is a high-risk managed input.

• bare (boolean, optional): Start Claude in minimal mode, skipping local customization discovery. true is a high-risk managed input.

• debugFile (string, optional): Write Claude debug output to a file. A non-empty path is a high-risk managed input.

• maxBudgetUsd (number, optional): Budget cap in USD for the request

• maxTurns (integer, optional): Agent-loop turn cap

• effort (string, optional): Reasoning effort (low|medium|high|xhigh|max)

• fallbackModel (string, optional): Auto-fallback model when the default is overloaded

• jsonSchema (string, optional): JSON Schema literal constraining structured output

• addDir (string[], optional): Additional workspace directories. A non-empty value is a high-risk managed input.

• noSessionPersistence (boolean, optional): Ephemeral session (not persisted to disk)

• settingSources / settings / tools (optional): Setting sources to load, settings JSON path/literal, built-in tool restriction. Non-empty setting sources, settings, or tool selections are high-risk managed inputs.

• pluginDir / pluginUrl (string[], optional): Load Claude plugins from local directories or URLs. Non-empty values are high-risk managed inputs.

• excludeDynamicSystemPromptSections (boolean, optional): Trim dynamic system prompt sections

• approvalStrategy (string, optional): "legacy" (default) or "mcp_managed". Managed mode uses acceptEdits by default and forces strictMcpConfig:true, so Claude uses only the gateway-generated MCP configuration. A direct full-permission request requires all of an explicit caller request, an approval-manager approval, and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1. Other high-risk inputs require the approval and operator setting too, but remain bounded and do not themselves select full permission.

• approvalPolicy (string, optional): "strict", "balanced", or "permissive"

• mcpServers (string[], optional): Names of MCP servers to expose to Claude (default: none). Legacy requests resolve names from the local registry or Codex MCP config; unknown names are reported as unavailable. Under mcp_managed, Claude uses only the generated configuration and only registry entries explicitly provisioned as gateway-owned local commands are eligible. Dynamic npx launchers, ambient-PATH commands, and Codex-config overrides are rejected. Configure and deploy the managed entries in the gateway environment.

• strictMcpConfig (boolean, optional): In legacy mode this defaults to false; set true to require only the generated MCP config and fail if requested servers are unavailable. Under mcp_managed, the gateway forces it to true and a caller-supplied false cannot weaken that boundary.

• optimizePrompt (boolean, optional): Optimize prompt for token efficiency (44% reduction), default: false

• optimizeResponse (boolean, optional): Optimize response for token efficiency (37% reduction), default: false

• correlationId (string, optional): Request trace ID (auto-generated if omitted)

• idleTimeoutMs (integer, optional): Kill a stuck process after output inactivity; 30,000 to 3,600,000 ms. Idle enforcement applies only when outputFormat is stream-json; it is ignored for text/json, which produce no output until the run completes

• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ). A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, so filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands. Requesting a worktree is a high-risk managed input that requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1, but remains bounded.

• promptParts (object, optional): Cache-aware structured prompt { system?, tools?, context?, task }; mutually exclusive with prompt

• forceRefresh (boolean, optional): Bypass dedup and force a fresh CLI run, default: false

Workspace boundary: stdio callers may use machine-local paths directly. HTTP/tunnel callers must pass workspace or rely on a configured default/session workspace; path fields are then validated relative to that workspace. [workspaces].allow_unregistered_working_dir is an inert legacy key: it is still accepted so old configs keep loading, but nothing reads it at either value, and setting it now logs a warning at startup. It never allowed arbitrary HTTP working directories or additional directories.

Response extras:

• approval: Approval decision record when approvalStrategy="mcp_managed"

• mcpServers: Requested/enabled/missing MCP servers for this call

Example:

{
  "prompt": "Write a Python function to calculate fibonacci numbers",
  "model": "sonnet",
  "continueSession": true,
  "optimizePrompt": true,
  "optimizeResponse": true
}

codex_request

Execute a Codex request with optional session tracking.

Parameters:

• prompt (string, optional*): The prompt to send (1-100,000 chars). *Exactly one of prompt or promptParts is required (mutually exclusive)

• model (string, optional): Model name or alias (use list_models for available values; supports latest, recommended: gpt-5.5)

• fullAuto (boolean, optional): Deprecated — expands to --sandbox workspace-write only (current Codex no longer accepts approval-policy flags); prefer sandboxMode

• sandboxMode (string, optional): Codex sandbox (read-only|workspace-write|danger-full-access).

• dangerouslyBypassApprovalsAndSandbox (boolean, optional): Request Codex's full approvals-and-sandbox bypass.

• dangerouslyBypassHookTrust (boolean, optional): Request Codex hook-trust bypass.

• approvalStrategy (string, optional): "legacy" is the only executable strategy. "mcp_managed" is rejected before Codex launches because the adapter cannot isolate ambient MCP configuration.

• approvalPolicy (string, optional): Has no effect for Codex because mcp_managed is unavailable.

• mcpServers (string[], optional): Metadata only. It does not configure or isolate Codex MCP servers.

• sessionId (string, optional): Session identifier for tracking.

• resumeLatest (boolean, optional): Resume a previous Codex session (codex exec resume --last). Do not rely on which session --last selects or on the resumed working directory (#258): --last is cwd-filtered upstream and the child is still spawned with the gateway-resolved cwd. Verify the target, or start a fresh session when it must be certain. Ignored if sessionId is set.

• createNewSession (boolean, optional): Always create a new session

• forceRefresh (boolean, optional): Bypass dedup and force a fresh CLI run, default: false

• outputFormat (string, optional): text (default) or json (--json JSONL events for token usage extraction)

• outputSchema (string|object, optional): Codex --output-schema, path or inline JSON Schema.

• workingDir (string, optional): Working root for this session (-C/--cd; new sessions only). Personal Agent Config Kit mode requires an absolute path.

• addDir (string[], optional): Additional writable workspace directories (one --add-dir per entry; new sessions only).

• ephemeral (boolean, optional): Codex --ephemeral (no session persistence)

• images (string[], optional): Image attachments (one -i <path> per entry).

• profile (string, optional): Codex --profile <name> (new sessions only; ignored with a logged warning on resume).

• configOverrides (object, optional): Codex -c key=value overrides. Local callers only; remote HTTP/OAuth requests are rejected.

• enable / disable (string[], optional): Codex --enable / --disable feature overrides. They are -c features.* equivalents and are also local-only.

• ignoreRules / ignoreUserConfig (boolean, optional): Codex --ignore-rules / --ignore-user-config.

• outputLastMessage (string, optional): Codex --output-last-message <path>.

• oss (boolean, optional): Codex --oss, selecting the open-source provider.

• localProvider (string, optional): Codex --local-provider <name>.

• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ). A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, so filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands.

• promptParts (object, optional): Cache-aware structured prompt { system?, tools?, context?, task }; mutually exclusive with prompt

• optimizePrompt (boolean, optional): Optimize prompt for token efficiency, default: false

• optimizeResponse (boolean, optional): Optimize response for token efficiency, default: false

• correlationId (string, optional): Request trace ID (auto-generated if omitted)

• idleTimeoutMs (integer, optional): Kill a stuck Codex process after output inactivity; 30,000 to 3,600,000 ms

Response extras:

• mcpServers: Requested MCP-server metadata for this call

Example:

{
  "prompt": "Create a REST API endpoint",
  "model": "gpt-5.5",
  "sandboxMode": "workspace-write",
  "optimizePrompt": true
}

codex_fork_session

Fork an existing Codex session into a new branch (codex fork <SESSION_ID|--last> <prompt>), preserving the original session's history while the fork diverges. Unlike Codex new and resume requests, this command remains argv-bound and rejects oversized UTF-8 prompts as non-retryable input_too_large.

Parameters:

• prompt (string, required): Prompt text for the forked session (1-100,000 chars)

• sessionId (string, optional): Codex session UUID to fork from (mutually exclusive with forkLast).

• forkLast (boolean, optional): Fork the most recent Codex session instead of naming one.

• model (string, optional): Model name or alias (e.g. gpt-5.5, latest)

• sandboxMode (string, optional): Codex sandbox (read-only|workspace-write|danger-full-access).

• approvalStrategy (string, optional): "legacy" is the only executable strategy. "mcp_managed" is rejected before Codex launches because the adapter cannot isolate ambient MCP configuration.

• approvalPolicy (string, optional): Has no effect for Codex because mcp_managed is unavailable.

• correlationId (string, optional): Request trace ID (auto-generated if omitted)

• idleTimeoutMs (number, optional): Idle timeout in ms (30s-1h, omit for CLI default)

gemini_request

Execute a Google Antigravity CLI (agy) request with session support.

Parameters:

• prompt (string, optional*): The prompt to send (1-100,000 chars). *Exactly one of prompt or promptParts is required (mutually exclusive)

• model (string, optional): Model name or alias (use list_models for available values; supports latest, pro, flash)

• sessionId (string, optional): Session ID to resume.

• resumeLatest (boolean, optional): Resume the latest session automatically.

• createNewSession (boolean, optional): Always create a new session

• approvalMode (string, optional): Antigravity approval mode in legacy mode: default leaves agy prompted, auto_edit emits --mode accept-edits, plan emits --mode plan, and yolo emits --dangerously-skip-permissions.

• approvalStrategy (string, optional): "legacy" is the only executable strategy. "mcp_managed" is rejected before Antigravity launches because the adapter cannot isolate ambient MCP configuration.

• approvalPolicy (string, optional): Has no effect for Antigravity because mcp_managed is unavailable.

• includeDirs (string[], optional): Additional workspace directories (passed as --add-dir).

• project (string, optional): Select the Antigravity project for this session (--project <ID>); mutually exclusive with newProject.

• newProject (boolean, optional): Create a new Antigravity project for this session (--new-project); mutually exclusive with project.

• sandbox (boolean, optional): Run Antigravity in sandbox mode (--sandbox)

• workingDir (string, optional): Local Antigravity process working directory. Stdio/local callers may pass local paths directly; remote HTTP/OAuth callers must use relative paths inside a selected registered workspace. includeDirs adds read paths but does not select cwd.

• workspace (string, optional): Registered gateway workspace alias that selects the Antigravity process cwd for remote HTTP/OAuth callers.

• outputFormat (string, optional): text (default), json, or stream-json. The async job recorder launches the transcript-capable stream-json wire independently of the requested presentation.

• mcpServers (string[], optional): Metadata only. Antigravity manages its own MCP configuration; this field does not create an allowlist.

• allowedTools, policyFiles, adminPolicyFiles, attachments (string[], optional) and skipTrust (boolean, optional): Unsupported by Antigravity CLI. Non-empty values, or skipTrust: true, are rejected with an explanatory error.

• yolo (boolean, optional): Auto-approve all; equivalent to approvalMode: "yolo". Emits --dangerously-skip-permissions in legacy mode.

• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ). A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, so filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands.

• promptParts (object, optional): Cache-aware structured prompt { system?, tools?, context?, task }; mutually exclusive with prompt

• optimizePrompt (boolean, optional): Optimize prompt for token efficiency, default: false

• optimizeResponse (boolean, optional): Optimize response for token efficiency, default: false

• correlationId (string, optional): Request trace ID (auto-generated if omitted)

• idleTimeoutMs (integer, optional): Total-runtime bound, not an idle timer: this provider emits no output until it exits, so the process is killed after this duration even while healthy. 30,000 to 3,600,000 ms, default 3,600,000 ms

• forceRefresh (boolean, optional): Bypass dedup and force a fresh CLI run, default: false

Response extras:

• mcpServers: Requested MCP-server metadata for this call

Example:

{
  "prompt": "Explain quantum computing",
  "model": "latest",
  "resumeLatest": true,
  "optimizePrompt": true
}

grok_request

Execute a Grok CLI (xAI) request with session support.

Parameters:

• prompt (string, optional*): The prompt to send (1-100,000 chars). *Exactly one of prompt or promptParts is required (mutually exclusive)

• model (string, optional): Model name or alias (e.g. grok-4.6, grok-4.5). Prefer live discovery via list_models or models://grok: the legacy grok-build id was removed upstream and now hard-fails with Invalid params: "unknown model id".

• transport (string, optional): "cli" (default) runs the Grok CLI; "acp" routes through Grok's native grok agent stdio transport when [acp].enabled and the provider's runtime_enabled are set (fails closed otherwise). Both transports reject approvalStrategy:"mcp_managed"; approvalPolicy has no effect. Sync-only: grok_request_async always runs the CLI transport and does not accept transport

• outputFormat (string, optional): "plain" (default), "json", or "streaming-json"

• sessionId (string, optional): Session ID to resume (--resume <id>).

• resumeLatest (boolean, optional): Resume the most recent session in the current cwd (--continue).

• createNewSession (boolean, optional): Always create a new session

• alwaysApprove (boolean, optional): Auto-approve all tool executions (--always-approve) in legacy mode.

• permissionMode (string, optional): default|acceptEdits|auto|dontAsk|bypassPermissions|plan.

• effort (string, optional): low|medium|high|xhigh|max

• reasoningEffort (string, optional): Reasoning effort for reasoning models

• approvalStrategy (string, optional): "legacy" is the only executable strategy. "mcp_managed" is rejected before Grok launches because the adapter cannot isolate ambient MCP configuration.

• approvalPolicy (string, optional): Has no effect for Grok because mcp_managed is unavailable.

• mcpServers (string[], optional): Metadata only. Grok manages its own MCP configuration via grok mcp; this field does not create an allowlist.

• allowedTools (string[], optional): Allowed built-in tools (passed as --tools comma list).

• disallowedTools (string[], optional): Disallowed built-in tools (passed as --disallowed-tools comma list)

• maxTurns (integer, optional): Agent-loop iteration cap (--max-turns)

• workingDir (string, optional): Working directory for this invocation (--cwd)

• sandbox (string, optional): Sandbox profile for filesystem/network access (--sandbox, freeform; also via GROK_SANDBOX).

• rules (string, optional): Extra rules appended to the system prompt (--rules; supports @file prefix).

• systemPromptOverride (string, optional): Replace the agent's system prompt entirely.

• allow / deny (string[], optional): Permission allow/deny rules (one --allow/--deny per entry).

• compactionMode (string, optional): summary (default) |transcript|segments

• compactionDetail (string, optional): none|minimal|balanced|verbose (segments mode only)

• agent (string, optional): Agent name or definition file path.

• agents (string|object, optional): Inline subagent definitions JSON.

• disableWebSearch (boolean, optional): Disable web search and remote retrieval tools

• todoGate (boolean, optional): Enable runtime turn-end TodoGate (session-scoped)

• verbatim (boolean, optional): Send the prompt exactly as given (also skips gateway prompt optimisation)

• promptFile / promptJson / single (optional): Single-turn prompt from a file / JSON blocks / literal.

• experimentalMemory / noMemory (boolean, optional): Enable/disable cross-session memory.

• noAltScreen / noPlan / noSubagents (boolean, optional): Disable alt screen / plan mode / subagent spawning

• oauth (boolean, optional): Use OAuth during authentication.

• restoreCode (boolean, optional): Check out the original session commit when resuming.

• leaderSocket (string, optional): Custom leader socket path (--leader-socket, Grok 0.2.32+; default ~/.grok/leader.sock) targeting an isolated leader process, for example a local or branch Grok build.

• nativeWorktree (boolean|string, optional): Grok's own --worktree flag (true means bare, string means named); distinct from the gateway worktree option.

• worktreeRef (string, optional): Branch/tag/commit to base the native worktree on (--worktree-ref); requires nativeWorktree.

• forkSession (boolean, optional): Fork the resumed session into a new branch instead of appending to it.

• jsonSchema (string|object, optional): JSON Schema (string or object) constraining structured output (--json-schema)

• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ). Grok requires an explicit provider-native sessionId; fresh, createNewSession, and resumeLatest-only worktree requests are rejected because they cannot durably reselect the worktree. A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, so filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands.

• promptParts (object, optional): Cache-aware structured prompt { system?, tools?, context?, task }; mutually exclusive with prompt

• optimizePrompt (boolean, optional): Optimize prompt for token efficiency, default: false

• optimizeResponse (boolean, optional): Optimize response for token efficiency, default: false

• correlationId (string, optional): Request trace ID (auto-generated if omitted)

• idleTimeoutMs (integer, optional): Kill a stuck process after output inactivity; 30,000 to 3,600,000 ms

• forceRefresh (boolean, optional): Bypass dedup and force a fresh CLI run, default: false

Example:

{
  "prompt": "Summarize the latest commit message in 1 sentence",
  "model": "grok-4.6",
  "effort": "low"
}

Durable job results & automatic dedup

Every async job is persisted to a job store as it transitions through running → completed/failed/canceled. This makes the gateway a durable collection layer:

• Re-issuing a request is safe. Identical *_request / *_request_async calls within the dedup window (default 1 hour) short-circuit onto the existing running or completed job — the caller gets back the same job ID instead of starting a duplicate run. This directly fixes the "agent times out polling, re-issues, and the whole job starts over" failure mode.

• llm_job_status and llm_job_result work across gateway restarts. Job rows are retained until an operator configures a bound; callers can collect results long after the in-memory cache has evicted them.

• A job is marked orphaned only when its owning gateway instance is provably gone, never because another instance restarted. Each instance holds a periodic heartbeat lease and stamps every job it owns; the recovery sweep orphans a queued/running job only when that job's own lease has expired. On a shared store (backend = "postgres") this means a fresh instance never orphans another live instance's in-flight jobs. The captured partial output of a genuinely orphaned job remains readable, and a stale-then-reviving owner that later finishes self-heals to the correct terminal state (issue #139).

• Pass forceRefresh: true on any request tool to bypass dedup and force a fresh CLI run.

Persistence configuration

The durable backend is configured by ~/.llm-cli-gateway/config.toml (override with LLM_GATEWAY_CONFIG=/path/to/config.toml). Example:

[persistence]
backend = "sqlite"                          # "sqlite" | "memory" | "postgres" | "none"
path = "~/.llm-cli-gateway/logs.db"         # for sqlite
# dsn = "postgresql://user:pw@host/db"      # for postgres
# retentionDays = 30                       # optional legacy JOB-store bound; omitted means unbounded
dedupWindowMs = 3600000
acknowledgeEphemeral = false                # required to enable async tools with memory backend

# One retention policy, over every subsystem. Every destructive bound is OFF
# unless you write a number, because deleting prompt, response, or transcript
# history is destructive and no upgrade should do it for you. Unknown keys are
# refused rather than silently applying no bound.
# `llm-cli-gateway doctor --json` -> .storage.retention reports what each bound
# would delete BEFORE you set it.
[persistence.retention]
# jobs = 30                                 # overrides retentionDays above
# requests = 90                             # flight-recorder transcripts, incl. bodies
# wedgedValidationRuns = 30                 # validation runs nothing can ever finalize
# sweepIntervalMs = 3600000                 # how often the sweeper ticks

# Issue #139 durable orphan-recovery lease (defaults shown). Each instance
# advances a per-job lease on every heartbeat; the sweep orphans a job only
# after its own lease expires, so a fresh instance never orphans another live
# instance's jobs on a shared store. Validated: leaseTtl >= 2*heartbeat and
# httpJobGrace >= leaseTtl.
instanceHeartbeatMs = 15000                 # heartbeat cadence
instanceLeaseTtlMs = 90000                  # per-job lease TTL (6x heartbeat)
httpJobGraceMs = 300000                     # extra grace for no-pid http jobs (5 min)
orphanSweepIntervalMs = 30000               # reaper cadence
instanceGcMs = 3600000                      # gateway_instances GC horizon
# ownsOrphanRecovery = false                # DEPRECATED (#139): superseded by the lease; parsed + warned, no longer used

# Optional, postgres only. One credential per class of work, for a deployment
# that has provisioned the RBAC in docs/plans/postgres-security-hardening.md.
# `app` is not a key here: the runtime credential is [persistence].dsn above.
# `migrate` is not a key either, and must not be held by a running gateway.
# A role left out degrades onto `app`, which llm_process_health reports as
# `persistence.roles.degraded` rather than implying separation is in force.
# [persistence.roles]
# reader = "postgresql://llmgw_reader@host/db"       # transcript read-back
# analytics = "postgresql://llmgw_analytics@host/db" # aggregates, no body text
# retention = "postgresql://llmgw_retention@host/db" # job expiry

Backends:

• sqlite (default) — durable, file-backed. Safe for single-instance deployments.

• postgres: PostgreSQL is authoritative for every backend-governed durable subsystem, including async jobs, dedup, orphan recovery, HTTP jobs, validation state, session metadata, and the flight recorder. File-backed operator records such as approvals, admin audit, and the workspace registry are not SQLite/PostgreSQL engine choices. Use this for multi-instance or service deployments. Requires the optional peer dependency pg to be installed alongside the gateway.

• memory — in-process Map. Lost on gateway exit. Requires acknowledgeEphemeral = true to be loaded. Suitable for tests and ephemeral CI gateways.

• none — no store. *_request_async, llm_job_status, llm_job_result, and llm_job_cancel are NOT registered on the gateway. This is a structural invariant: agents that try to call async tools against a gateway with backend = "none" get a clean "tool not found" at connect time instead of silent in-memory loss after the 1-hour TTL. Use llm_process_health to inspect the resolved persistence state programmatically.

backend = "postgres" is the whole engine decision. The configured dsn must be a PostgreSQL URL. The gateway does not infer deployment topology, resolve the host, restrict PostgreSQL to a local target, or render a DSN-derived target on diagnostics. If migration, connection, or an operation fails, the affected PostgreSQL subsystem fails closed and the gateway does not open SQLite as a fallback. Health surfaces identify the selected engine and report PostgreSQL failures generically.

Changing engines does not move data. Existing SQLite and PostgreSQL rows stay where they are: there is no automatic migration, dual-write, or cross-engine read. Plan and execute any data move separately before cutover if old history must remain available through the gateway.

backend = "none" and LLM_GATEWAY_LOGS_DB=none are different switches. The first disables async job persistence; the second disables request history. Setting the backend to "none" leaves the flight recorder writing, and disabling the recorder leaves the job store alone. Both are stated at startup in the Storage: block on stderr. llm_process_health reports the selected engine, recorder state, role configuration, and deprecated-input decisions without rendering the configured PostgreSQL target.

DATABASE_URL is deprecated and never overrides [persistence]. When it is the only selector, it acts as a deprecated alias for backend = "postgres" plus dsn, so sessions, jobs, validation state, and request history stay on one engine. Against an explicit backend, or against a dsn it disagrees with, it is ignored with a reason: the gateway does not abort, because the outcome is fully determined and is the one the config file describes.

For PostgreSQL, apply the schema with a schema-owner or dedicated migration role before starting a DML-only gateway role:

DATABASE_URL='postgresql://<user>:<password>@<host>/<database>' npm run migrate

The runner serializes migration work with an advisory lock and records a SHA-256 of each migration in schema_migrations.checksum_sha256. On later runs, a malformed or mismatched recorded checksum stops the runner before it calculates or applies pending migrations. A NULL checksum is an explicit legacy row from before checksum recording. It is allowed with a warning but is never backfilled, because current source files cannot prove what SQL ran historically. Do not edit released migration files or populate ledger checksums manually. The runner preserves the historical 002/003 SQL and applies compatibility only while one of those legacy versions remains pending; forward migration 018 repairs an already-recorded legacy session/view layout. Release checks reject a source edit to a published migration file.

Legacy environment variables (deprecated; emit a warning at startup):

• LLM_GATEWAY_LOGS_DB / LLM_GATEWAY_JOBS_DB: when no backend is explicitly configured, none selects backend = "none" and any other value selects backend = "sqlite" with that path. An explicit [persistence].backend wins. LLM_GATEWAY_LOGS_DB still independently disables the recorder on every backend and paths it when the selected recorder engine is SQLite.

• LLM_GATEWAY_JOB_RETENTION_DAYS — overrides retentionDays.

• LLM_GATEWAY_DEDUP_WINDOW_MS — overrides dedupWindowMs.

• LLM_GATEWAY_ACKNOWLEDGE_EPHEMERAL — 1/true/yes sets acknowledgeEphemeral = true.

Host-protection limits ([http] and [limits])

The gateway bounds HTTP session growth and async/sync job execution so a burst of clients or requests cannot drive unbounded memory, process, CPU, or provider-request growth. All keys live in the same ~/.llm-cli-gateway/config.toml; defaults are conservative but chosen not to surprise local stdio development.

[http]                              # HTTP MCP transport session lifecycle
max_sessions = 100                  # max concurrent live sessions; excess initialize returns HTTP 429
session_idle_ttl_ms = 1800000       # 30 min: reap a session idle longer than this (no client DELETE needed)
session_reaper_interval_ms = 60000  # 1 min: how often the idle reaper sweeps

[limits]                            # async + sync job-execution backpressure (per gateway process)
max_running_jobs = 32               # global concurrent running jobs (process CLI + HTTP API)
max_running_jobs_per_provider = 16  # per-provider concurrent running jobs
max_queued_jobs = 128               # bounded wait queue; a full queue rejects new work
queue_timeout_ms = 120000           # 2 min: max time a job waits in the queue before failing
completed_job_memory_ttl_ms = 3600000  # 1 h: in-memory retention for finished jobs (durable rows kept separately)
max_job_output_bytes = 52428800     # 50 MB: per-job stdout+stderr cap

Failure modes (all deterministic and safe to retry):

• HTTP session cap reached: the initialize request returns 429 with Retry-After: 5 and a structured { error, code: "session_capacity", retryable: true } body. No new session is created.

• Idle HTTP session: the reaper closes it (transport + gateway server) once idle past session_idle_ttl_ms, independent of the client sending DELETE. A session with an in-flight request is never reaped mid-request.

• Job limiter saturated: when the running limit is reached and the queue is full, *_request / *_request_async and the direct-sync fallback return a retryable saturated error (structuredContent.errorCategory = "saturated", retryable: true). Nothing is spawned. When the queue has room the job waits (FIFO, per-provider fair) up to queue_timeout_ms, then fails with the same category.

• Sync direct execution: the SYNC_DEADLINE_MS=0 and storeless/backend="none" paths acquire the same process permit before spawning, so no execution bypasses the limiter.

• Output overflow: a job whose combined stdout+stderr exceeds max_job_output_bytes is failed (exit code 126), its process terminated, its completion persisted, and its run slot released.

• In-memory vs durable retention: completed_job_memory_ttl_ms only ages finished jobs out of the in-memory map. The durable job store is unbounded by default and can be bounded with [persistence.retention].jobs or the legacy [persistence].retentionDays, so results stay readable via llm_job_result / llm_request_result after in-memory eviction.

Live counters are exposed on GET /healthz (unauthenticated, HTTP transport) and via the llm_process_health tool backpressure block: session current/max/oldest-age/idle-TTL/saturation, running and queued job counts globally and per provider, limiter saturation counters, configured TTL/output caps, and parent-process RSS/heap. These surfaces report counts, ages, and bytes only, never prompt text, response content, tokens, session IDs, bearer/OAuth tokens, API keys, or machine secrets.

For production user services, pair the in-process limits above with systemd's outer guardrails so an unexpected bug, provider CLI leak, or evaluation burst cannot consume the host:

systemctl --user edit llm-cli-gateway.service

[Service]
MemoryMax=2G
TasksMax=512

Choose values for your workload: MemoryMax should cover the gateway process, the configured max_running_jobs provider children, and normal output buffering; TasksMax should exceed the process/thread count implied by max_running_jobs plus the HTTP server and SQLite work, but still be far below host exhaustion. If systemd terminates the service at those limits, durable jobs can be inspected after restart and llm_process_health.backpressure should be used to tune [http], [limits], MemoryMax, and TasksMax together.

Per-project isolation

By default, gateway state is global per user, not per project. With no overrides, every Claude Code window across every repo spawns its own gateway subprocess but they all read and write the same state:

• ~/.llm-cli-gateway/logs.db when [persistence].backend = "sqlite" (async jobs + flight recorder). With backend = "postgres", those rows live in the configured database instead. All destructive retention bounds default to OFF. [persistence.retention].jobs prunes complete job records, [persistence.retention].requests prunes the flight recorder's requests and gateway_metadata tables, and [persistence.retention].wedgedValidationRuns prunes validation runs that can never be finalized with their validation_run_jobs links. An operator must opt in to each bound. validation_receipts is immutable by design and is never pruned; a finalized run is never pruned either, so no receipt is ever orphaned.

  Deleting rows frees SQLite pages but never bytes, so the file does not shrink. llm-cli-gateway storage compact --yes returns the space, with the gateway stopped, because a VACUUM holds an exclusive lock for the length of a full rewrite. doctor --json -> .storage.retention reports the resolved bounds, what a sweep would delete, and how many bytes a compaction would return. See docs/plans/durable-state-lifecycle.dag.toml.

• ~/.llm-cli-gateway/sessions.json (gateway session metadata when using the default file session backend)

• ~/.llm-cli-gateway/config.toml (resolved config)

When [persistence].backend = "postgres" selects the PostgreSQL session manager, the session metadata lives in PostgreSQL instead of sessions.json. This is usually what you want: session_list from repo A can show sessions from repo B, an async job started in window A can be polled from window B, and the 1-hour dedup window catches re-issues across windows. Gateway-managed worktrees are the exception: they are filesystem-local and therefore fail closed with PostgreSQL session storage. SQLite WAL mode protects the default job/flight-recorder database, while the file session manager uses locked atomic writes.

Per-project durable-job isolation

If unrelated repositories should not share async jobs, flight-recorder rows, or deduplication, point each project at its own persistence config. In .claude/settings.local.json for the project:

{
  "mcpServers": {
    "llm-gateway": {
      "env": {
        "LLM_GATEWAY_CONFIG": "${workspaceFolder}/.gateway/config.toml"
      }
    }
  }
}

…and put a per-project config.toml in the repo:

[persistence]
backend = "sqlite"
path = "/srv/repos/.../my-repo/.gateway/logs.db"

Now every gateway subprocess spawned for this repo's Claude Code window reads its own config and writes its durable jobs, flight-recorder rows, and deduplication state to its own SQLite file. Other repos keep using the global default. This [persistence] override does not move the default file-backed sessions.json, so it does not isolate session lists. llm_process_health.persistence.sources.configFile lets an agent confirm which persistence config it is actually running under.

Agent-executable spec (DAG-TOML)

If you want an LLM agent to perform this setup deterministically — rather than reading the prose above and guessing — copy the following DAG-TOML into the repo (e.g. docs/planning/per-project-gateway-isolation.toml) and point your agent at it. The schema is agent-assurance template_kind = "implementation-dag". The agent MUST execute units in layer order, must not skip the verification unit, and must treat any failed gate as blocking.

[meta]
schema_version    = "1.0.0"
template_kind     = "implementation-dag"
docs              = "https://github.com/verivus-oss/agent-assurance/blob/main/SPEC.md"
confidentiality   = "public"
title             = "Per-project llm-cli-gateway durable-job isolation"
spec              = "https://github.com/verivus-oss/llm-cli-gateway#per-project-durable-job-isolation"
created           = "YYYY-MM-DD"
total_units       = 5
tier1_units       = ["U01","U02","U03","U04","U05"]
tier2_units       = []
tier3_units       = []

# ============================================================================
# [policy.agent] — persona for the agent performing the configuration.
# ============================================================================

[policy.agent]
name                 = "Gateway Persistence Isolator"
role                 = "Configuration Engineer"
purpose              = "Configure the llm-cli-gateway MCP server so its SQLite durable state is scoped to THIS repository instead of the per-user default at ~/.llm-cli-gateway/. The default file-backed session metadata remains shared."
validation_type      = "Structural + Runtime Verification"
workflow_initiator   = false
description          = "Writes a repo-local config.toml, registers an LLM_GATEWAY_CONFIG override in .claude/settings.local.json, restarts the MCP server, and confirms via llm_process_health that the gateway is now reading the repo-local config and writing to the repo-local SQLite path."

[policy.agent.orchestration]
consumes_events      = ["PerProjectIsolationRequested"]
produces_events      = ["PerProjectDurableJobIsolationComplete"]

[policy.agent.responsibilities]
items = [
  "Create the repo-local gateway data directory and add it to .gitignore.",
  "Write a config.toml that pins backend=sqlite to a repo-local path.",
  "Register the LLM_GATEWAY_CONFIG env override in .claude/settings.local.json (NOT .mcp.json — that file is committed and shared).",
  "Trigger an MCP server reconnect.",
  "Verify via llm_process_health that the resolved configFile and dbPath are the repo-local values.",
]

# ============================================================================
# [policy.instance] — concrete paths the agent fills in for THIS repo.
# Agent MUST replace <REPO_ABS_PATH> with the absolute path to the repo
# before emitting any artefact. Relative paths in config.toml MUST be
# expanded to absolute — the gateway does not re-resolve them per cwd.
# ============================================================================

[policy.instance]
repo_abs_path                  = "<REPO_ABS_PATH>"           # e.g. /srv/repos/me/my-project
gateway_data_dir_relative      = ".gateway"                  # repo-relative directory
config_toml_relative           = ".gateway/config.toml"
sqlite_db_relative             = ".gateway/logs.db"
claude_local_settings_relative = ".claude/settings.local.json"
gitignore_relative             = ".gitignore"
mcp_server_name                = "llm-gateway"               # must match the entry in .mcp.json

# ============================================================================
# [policy.gates] — blocking checks. Any failure stops the workflow.
# ============================================================================

[policy.gates]
gate_repo_abs_path_resolved    = "policy.instance.repo_abs_path must NOT be the literal string '<REPO_ABS_PATH>' when U01 starts."
gate_config_is_committed       = "policy.instance.config_toml_relative MAY be committed. policy.instance.claude_local_settings_relative MUST NOT be committed (it is per-developer). Agent MUST verify .gitignore covers .claude/settings.local.json if absent."
gate_no_legacy_env_leak        = "Agent MUST grep the shell init files for LLM_GATEWAY_LOGS_DB / LLM_GATEWAY_JOBS_DB. An explicit [persistence].backend wins for job persistence, but LLM_GATEWAY_LOGS_DB still disables the recorder on every backend and paths it when SQLite is selected. Both variables remain deprecated. The agent reports either as a finding and asks the operator to unset it before proceeding."
gate_health_confirms_isolation = "U05 MUST observe llm_process_health.persistence.sources.configFile == policy.instance.repo_abs_path + '/' + policy.instance.config_toml_relative AND llm_process_health.persistence.path == policy.instance.repo_abs_path + '/' + policy.instance.sqlite_db_relative. Anything else means the override did not take effect."

# ============================================================================
# [policy.evidence] — what each unit must emit so the work is auditable.
# ============================================================================

[policy.evidence]
per_unit_required_fields = [
  "unit_id",                  # U01..U05
  "status",                   # "completed" | "failed"
  "artefact_paths",           # files written / modified
  "stdout_tail",              # last 20 lines of any command output
  "verification_quote",       # for U05, the verbatim llm_process_health.persistence block
]
findings_required_fields = [
  "gate_id",                  # which gate failed
  "observed",
  "expected",
  "remediation",
]

# ============================================================================
# Units. Execute in layer order. U01..U03 modify the working tree; U04
# triggers a reconnect; U05 is the verification gate that decides success.
# ============================================================================

[units.U01]
name           = "create-repo-local-data-dir"
summary        = "mkdir -p <repo>/.gateway and append /.gateway/ to .gitignore (creating .gitignore if missing). The gateway will write logs.db, logs.db-wal, logs.db-shm here — none should be committed."
layer          = 0
tier           = 1
status         = "pending"
depends_on     = []
blocks         = ["U02"]
estimated_loc  = 5
files_modify   = [".gitignore"]
produces       = ["ART:gateway-data-dir"]
consumes       = []

[units.U02]
name           = "write-config-toml"
summary        = "Write <repo>/.gateway/config.toml with [persistence] backend='sqlite' and path=<absolute-path-to-repo>/.gateway/logs.db. Path MUST be absolute. Do NOT use ~ — the gateway expands ~ but [persistence].path is read literally if not prefixed with ~/, and Claude Code may launch the gateway with a HOME that surprises you."
layer          = 1
tier           = 1
status         = "pending"
depends_on     = ["U01"]
blocks         = ["U03"]
estimated_loc  = 10
files_modify   = [".gateway/config.toml"]
produces       = ["ART:gateway-config"]
consumes       = ["ART:gateway-data-dir"]

[units.U03]
name           = "register-llm-gateway-config-env-in-claude-local-settings"
summary        = "Add (or merge) an mcpServers.<mcp_server_name>.env entry in .claude/settings.local.json that sets LLM_GATEWAY_CONFIG to the absolute path of .gateway/config.toml. Do NOT modify .mcp.json — that file is committed and the path would be wrong for every other developer. If .claude/settings.local.json already has an mcpServers.<mcp_server_name> entry, the agent MUST merge into the existing env map (preserving other keys), not overwrite the whole entry."
layer          = 2
tier           = 1
status         = "pending"
depends_on     = ["U02"]
blocks         = ["U04"]
estimated_loc  = 20
files_modify   = [".claude/settings.local.json"]
produces       = ["ART:claude-local-settings"]
consumes       = ["ART:gateway-config"]

[units.U04]
name           = "trigger-mcp-reconnect"
summary        = "Ask the operator to run /mcp in Claude Code (or restart Claude Code) so the gateway subprocess is re-spawned under the new env. The agent cannot do this itself — MCP server lifecycle is owned by the host."
layer          = 3
tier           = 1
status         = "pending"
depends_on     = ["U03"]
blocks         = ["U05"]
estimated_loc  = 0
files_modify   = []
produces       = ["OUT:mcp-reconnected"]
consumes       = ["ART:claude-local-settings"]

[units.U05]
name           = "verify-via-llm-process-health"
summary        = "Call llm_process_health and assert the returned persistence block satisfies policy.gates.gate_health_confirms_isolation. Quote the verbatim persistence block in evidence. If the assertion fails, the agent MUST NOT mark the workflow complete — it must emit a finding under policy.evidence.findings_required_fields, naming the observed vs. expected configFile/path, and stop."
layer          = 4
tier           = 1
status         = "pending"
depends_on     = ["U04"]
blocks         = []
estimated_loc  = 5
files_modify   = []
produces       = ["ART:durable-isolation-verification","OUT:per-project-durable-isolation-complete"]
consumes       = ["OUT:mcp-reconnected"]

Why this matters for agents: the gateway has multiple configuration surfaces (TOML file, env-var overrides, two different MCP settings files) and one easy mistake, editing the committed .mcp.json instead of the local-only .claude/settings.local.json, will silently break the per-project persistence scope for every other developer on the repo. The DAG above encodes the correct sequence, the verification gate, and the failure modes explicitly so an agent can execute it without inference. It deliberately does not claim to isolate the default file-backed session store.

mistral_request

Run a Mistral Vibe agentic coding request. Like grok_request in shape, but with Vibe's specific surface:

• model (string, optional): Vibe model alias (for example mistral-medium-3.5 or latest). The resolved value is injected via the VIBE_ACTIVE_MODEL environment variable; omit it to let the gateway discover Vibe config and avoid stale hardcoded defaults.

• transport (string, optional): "cli" (default) runs the Vibe CLI; "acp" routes through Vibe's native vibe-acp transport when [acp].enabled and the provider's runtime_enabled are set (fails closed otherwise). Both transports reject approvalStrategy:"mcp_managed"; approvalPolicy has no effect. Sync-only: mistral_request_async always runs the CLI transport and does not accept transport

• permissionMode: the Vibe --agent name: builtins default | plan | accept-edits | auto-approve, or any install-gated/custom agent. Requests emit the supplied name in legacy mode.

• allowedTools (string[], optional): One --enabled-tools <tool> flag per entry.

• disallowedTools (string[], optional): One --disabled-tools <tool> flag per entry, applied after the enabled-tool filter.

• outputFormat (string, optional): Vibe 2.x values are "text", "json", or "streaming"; legacy aliases "plain" and "stream-json" are accepted and normalized before spawn.

• sessionId / resumeLatest / createNewSession: standard session controls. Current Vibe defaults session logging to enabled; if an older config has [session_logging] enabled = false, doctor --json surfaces an actionable next-action.

• trust (boolean, optional): Emit --trust so Vibe trusts the cwd for this invocation only (not persisted; skips the interactive trust prompt).

• maxTurns (integer, optional): Agent-loop iteration cap (--max-turns, programmatic mode only)

• maxPrice (number, optional): Interrupt when cumulative cost crosses this USD cap (--max-price, programmatic mode only)

• maxTokens (integer, optional): Cap cumulative prompt + completion tokens (--max-tokens, programmatic mode only)

• workingDir (string, optional): Change to this directory before running (--workdir)

• addDir (string[], optional): Additional writable workspace directories (one --add-dir per entry).

• approvalStrategy (string, optional): "legacy" is the only executable strategy. "mcp_managed" is rejected before Vibe launches because the adapter cannot isolate ambient MCP configuration.

• approvalPolicy (string, optional): Has no effect for Vibe because mcp_managed is unavailable.

• mcpServers (string[], optional): Metadata only. Vibe reads its MCP configuration from VIBE_HOME config; this field does not create an allowlist.

• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ). Mistral requires an explicit provider-native sessionId; fresh, createNewSession, and resumeLatest-only worktree requests are rejected because they cannot durably reselect the worktree. A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, so filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands.

• promptParts (object, optional): Cache-aware structured prompt { system?, tools?, context?, task }; mutually exclusive with prompt

• optimizePrompt / optimizeResponse (boolean, optional): Token-efficiency optimisation, default: false

• correlationId (string, optional): Request trace ID (auto-generated if omitted)

• idleTimeoutMs (integer, optional): Total-runtime bound, not an idle timer: this provider emits no output until it exits, so the process is killed after this duration even while healthy. 30,000 to 3,600,000 ms, default 3,600,000 ms

• forceRefresh (boolean, optional): Bypass dedup and force a fresh CLI run, default: false

devin_request

Run a Cognition Devin CLI request synchronously (headless print mode, devin -p). Auto-defers to a pollable job past the sync deadline when async jobs are enabled.

Parameters:

• prompt (string, optional*): Prompt text for Devin CLI (1-100,000 chars). Required in practice; promptFile is additive

• model (string, optional): Model name or alias (e.g. opus, latest)

• transport (string, optional): "cli" (default) runs the Devin CLI; "acp" routes through Devin's native devin acp transport when [acp].enabled and the provider's runtime_enabled are set (fails closed otherwise). Both transports reject approvalStrategy:"mcp_managed"; approvalPolicy has no effect. Sync-only: devin_request_async always runs the CLI transport and accepts neither transport nor agentType

• agentType (string, optional): ACP agent variant for transport: "acp" (devin acp --agent-type): "summarizer" (no tools, text summary) or "review" (read-only plus shell code-review); ignored for the CLI transport

• permissionMode (string, optional): Devin CLI permission mode (--permission-mode): auto (auto-approves read-only tools), accept-edits (also auto-approves workspace edits), smart (also auto-runs actions a fast model judges safe), dangerous (auto-approves all). Omit to use Devin's headless default

• approvalStrategy (string, optional): "legacy" is the only executable strategy. "mcp_managed" is rejected before Devin launches because the adapter cannot isolate ambient MCP configuration.

• approvalPolicy (string, optional): Has no effect for Devin because mcp_managed is unavailable.

• promptFile (string, optional): Load the initial prompt from a file (--prompt-file)

• sessionId (string, optional): Devin session ID to resume (--resume <id>). The gw-* id minted for a brand-new session is not resumable via sessionId; continue with resumeLatest: true

• resumeLatest (boolean, optional): Resume the most recent Devin session in the selected cwd (--continue). Requires workingDir, workspace, or a configured default workspace.

• createNewSession (boolean, optional): Force a new session

• workingDir (string, optional): Local Devin process cwd. CLI transport only.

• workspace (string, optional): Registered gateway workspace alias that selects the Devin process cwd.

• worktree (boolean|object, optional): Run the Devin CLI request inside a gateway-owned Git worktree. Devin requires an explicit provider-native sessionId; fresh, createNewSession, and resumeLatest-only worktree requests are rejected because they cannot durably reselect the worktree. A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, so filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands.

• optimizePrompt / optimizeResponse (boolean, optional): Token-efficiency optimisation, default: false

• correlationId (string, optional): Request trace ID (auto-generated if omitted)

• idleTimeoutMs (integer, optional): Total-runtime bound, not an idle timer: this provider emits no output until it exits, so the process is killed after this duration even while healthy. 30,000 to 3,600,000 ms, default 3,600,000 ms

• forceRefresh (boolean, optional): Bypass dedup and force a fresh CLI run, default: false

cursor_request

Run a Cursor Agent request synchronously. The default CLI transport uses headless print mode (cursor-agent --print) and auto-defers to a pollable job past the sync deadline when async jobs are enabled. Set transport: "acp" to use Cursor's native cursor-agent acp transport when [acp].enabled, [acp.providers.cursor].enabled, and [acp.providers.cursor].runtime_enabled are enabled. ACP accepts prompt, model, a gateway-owned ACP sessionId, and a registered gateway workspace alias. Both Cursor transports reject approvalStrategy:"mcp_managed"; approvalPolicy has no effect on CLI and is rejected on ACP.

Parameters:

• prompt (string, required): Prompt text for Cursor Agent CLI (1-100,000 chars)

• model (string, optional): Model name or alias (for example gpt-5, sonnet-4-thinking, or latest)

• mode (string, optional): Cursor mode, "plan" or "ask" (--mode)

• outputFormat (string, optional): "text" (default), "json", or "stream-json"

• transport (string, optional): "cli" (default) or "acp"; ACP fails closed unless its global and Cursor provider gates are enabled. ACP accepts only prompt, model, sessionId, and a registered workspace; non-default Cursor CLI controls (mode, non-text outputFormat, workingDir, non-empty addDir, force, autoReview, sandbox, trust, resumeLatest, createNewSession, optimization or compression, idleTimeoutMs, and forceRefresh) are rejected. Sync-only: cursor_request_async always runs the CLI transport and does not accept transport

• force (boolean, optional): Emit --force for non-interactive operation.

• autoReview (boolean, optional): Emit --auto-review.

• sandbox (string, optional): "enabled" or "disabled" (--sandbox).

• trust (boolean, optional): Emit --trust for this invocation.

• workspace (string, optional): On transport: "cli", a Cursor workspace path or name (--workspace); remote HTTP/OAuth callers must pass a registered workspace alias, while local stdio callers may pass paths. An unregistered relative local value is preserved verbatim as a provider-native saved-workspace name and is never resolved against the gateway process cwd; pass an absolute path to select a local directory cwd. On transport: "acp", it must be a registered gateway workspace alias. A fresh remote ACP request uses the supplied alias or [workspaces].default; a remote ACP resume stays bound to its recorded canonical alias and cwd

• workingDir (string, optional): Local Cursor Agent process working directory, CLI transport only. Distinct from workspace, which is Cursor's own selector; passing an absolute workspace path that disagrees with workingDir is rejected rather than silently ranked. Rejected on transport: "acp", which resolves its own scope, rather than being accepted and discarded.

• addDir (string[], optional): Additional workspace roots (one --add-dir per entry); remote HTTP/OAuth callers must use registered workspace roots.

• sessionId (string, optional): On transport: "cli", a Cursor chat/session ID to resume (--resume <id>). The gw-* id minted for a brand-new gateway session is not resumable through the CLI transport; continue with resumeLatest: true. On transport: "acp", pass the gateway-owned ACP session ID returned by an earlier ACP call; Cursor-native and CLI session IDs are rejected

• resumeLatest (boolean, optional): CLI only: resume the most recent Cursor chat (--continue). true is rejected on ACP.

• createNewSession (boolean, optional): CLI only: force a new session. true is rejected on ACP.

• approvalStrategy (string, optional): "legacy" is the only executable CLI strategy. "mcp_managed" is rejected before Cursor launches; ACP has its own permission bridge and also rejects it.

• approvalPolicy (string, optional): Has no effect on CLI because mcp_managed is unavailable. ACP rejects it because ACP uses its own permission bridge.

• optimizePrompt / optimizeResponse (boolean, optional): Token-efficiency optimisation, default: false

• correlationId (string, optional): Request trace ID (auto-generated if omitted)

• idleTimeoutMs (integer, optional): Total-runtime bound, not an idle timer: this provider emits no output until it exits, so the process is killed after this duration even while healthy. 30,000 to 3,600,000 ms, default 3,600,000 ms. This holds for the default text invocation; with outputFormat: "stream-json" Cursor streams incrementally, so the same timer behaves as a genuine idle window

• forceRefresh (boolean, optional): Bypass dedup and force a fresh CLI run, default: false

claude_request_async / codex_request_async / gemini_request_async / grok_request_async / mistral_request_async / devin_request_async / cursor_request_async

Start a long-running Claude, Codex, Gemini, Grok, Mistral, Devin, or Cursor request without waiting for completion in the same MCP call.

Use this flow when analysis/runtime can exceed client tool-call limits:

1. Start job with *_request_async

2. Poll with llm_job_status

3. Read output with llm_job_result

4. Optionally stop with llm_job_cancel

Async request tools accept the same strategy fields as their sync variants:

• claude_request_async supports approvalStrategy:"legacy" (default) or "mcp_managed", and approvalPolicy:"strict"|"balanced"|"permissive". Managed mode forces strictMcpConfig:true and launches Claude with only the generated MCP config.

• Every other *_request_async tool supports only executable approvalStrategy:"legacy". It rejects mcp_managed before provider launch because ambient MCP configuration cannot be isolated; approvalPolicy has no effect.

• mcpServers configures the Claude generated config only. For the other CLI adapters it is metadata only and does not create an MCP allowlist.

llm_job_status

Return lifecycle status (queued, running, completed, failed, canceled, orphaned), metadata, and bounded normalized progress for an async job. afterProgressSeq returns events with a greater sequence number and progressLimit selects up to 64 events in forward sequence order. Continue with nextAfterSeq; highWaterSeq (and its compatibility alias lastSeq) reports the highest sequence observed for the job, while hasMore reports whether another retained page is immediately available. The progress snapshot reports capability (structured, activity_only, or lifecycle_only), lastActivityAt, cursor/high-water metadata, droppedCount, and events with a phase, kind, timestamp, safe message, and source. Claude stream-JSON, Codex JSONL, Grok streaming-JSON, Gemini stream-JSON, Mistral streaming, and Cursor stream-JSON expose structured activity. Codex validation and repository review calls do not request JSONL and therefore report activity_only. HTTP/API jobs report lifecycle_only; other process output modes expose only privacy-safe activity/lifecycle signals. Raw reasoning, provider-supplied tool names, tool arguments, paths, provider IDs, and output text are not copied into progress messages. Tool-start activity uses the fixed message Using a provider tool.

llm_job_watch

Wait up to 30 seconds for new normalized progress on an owned async job. Pass the response's nextAfterSeq as afterProgressSeq; the response uses the same snapshot shape as llm_job_status. When the MCP request carries a progress token, the gateway emits notifications/progress only while that watch call is active. The tool rechecks owner access throughout the wait and returns the same own-or-not-found result as other job tools.

llm_job_result

Return captured stdout/stderr for an async job. By default it returns a display-oriented result, with maxChars limiting each stdout and stderr stream to 1,000 through 2,000,000 characters per call (200,000 by default).

For complete, resumable retrieval of a large provider stream, set rawOutput:true. The response's result includes independent stdoutOffsetChars, stdoutTotalChars, and stdoutNextOffsetChars fields, plus matching stderr* fields. Start both offsets at zero, then pass each non-null next offset back as stdoutOffsetChars or stderrOffsetChars until that stream's next offset is null. Raw pages are not display-parsed or compressed. On the local stdio surface, pages concatenate in stream order to the captured stdout or stderr stream. Remote callers use the same offset protocol, but provider-session-ID ranges are redacted before pages are returned, including an ID that crosses a page boundary. Treat remote rawOutput:true as resumable, sanitized output, not byte-for-byte captured provider output.

Non-zero offsets are rejected in default display mode. Display mode can parse a streaming provider format, reconstruct a Codex reply, or compress text, so a slice of the captured stream cannot safely resume or concatenate the displayed result. Use rawOutput:true whenever an application must resume output collection.

Every non-Kit CLI job records the exact launched argv, the resolved cwd scope, the repository HEAD when available, bounded SHA-256 instruction-file evidence, the capture grammar, and whether the transcript was captured whole, captured to the stated 50 MiB aggregate ceiling, or not captured. Claude, Codex, Gemini, Grok, Mistral, and Cursor are launched with their richest transcript-capable wires. Devin is launched with a gateway-owned ATIF export path; the regular, non-symlink export is copied into the durable job row after process exit and the source is then removed. Local stdio callers can page the native transcript in raw mode. Remote HTTP/OAuth callers cannot retrieve raw job results, native transcripts, instruction paths, or replay context.

llm_job_cancel

Cancel a running async job.

approval_list

List recent Claude MCP-managed approval decisions recorded by the gateway.

Parameters:

• limit (number, optional): Max records (1-500), default: 50

• cli (string, optional): Filter by "claude", "codex", "gemini", "grok", "mistral", "devin", or "cursor"

Approval records are persisted to ~/.llm-cli-gateway/approvals.jsonl.

llm_request_result

Read back any persisted request — sync or async — by its correlation ID. Every response echoes its ID in structuredContent.correlationId; pass it here to recover the persisted prompt/response after the inline result is gone. Reads the flight recorder, so it works independently of async-job persistence (returns "not found" when flight recording is disabled).

Parameters:

• correlationId (string, required): Correlation ID from a prior request

• maxChars (number, optional): Max chars of the persisted response to return (1,000-2,000,000)

• includePrompt (boolean, optional): Include the full persisted prompt text, default: false

• includeJobRecord (boolean, optional): Include the linked raw provider capture and replay context for local stdio callers, default: false. Remote callers receive neither transcripts nor host paths.

llm_request_list

List recent persisted requests newest-first without a correlation ID, which is how you find one. Every other flight-recorder read is keyed by an id handed out inline to the caller that made the request, so an agent that did not make the call (or whose context was compacted since) starts here. Returns metadata only: pass a returned correlationId to llm_request_result for the bodies, or a returned asyncJobId to llm_job_status. A caller only ever sees its own requests. An empty list is not proof nothing ran, and it does not mean the work left no record: cross-LLM validation seats write no row here, but they write validation_runs and validation_run_jobs, and each of those links a job row holding the launched argv and the provider output (read it with validation_receipt, then llm_job_result). Flight recording can also be disabled. Jobs and requests are both unbounded by default; an operator can configure their bounds independently.

Parameters:

• limit (number, optional): Max rows, 1-200, default: 25

• since (string, optional): ISO-8601 lower bound, e.g. 2026-08-21T00:00:00Z

• cli (string, optional): Restrict to one provider as recorded

• sessionId (string, optional): Restrict to one gateway session id

llm_process_health

Report gateway process health: async-job manager state, the resolved durable persistence configuration, and flight-recorder health on that configured engine. With backend = "postgres", the recorder reports the opaque postgresql target and generic PostgreSQL failures, and never falls back to SQLite. Its warning also states that rows left in the previous engine were not migrated. The explicit LLM_GATEWAY_LOGS_DB=none recorder-off switch remains separate, as do the intentionally ephemeral memory and disabled none job-store modes.

upstream_contracts

Return the gateway's declared provider CLI contracts, optionally probing the installed binaries for drift.

Parameters:

• cli (string, optional): Filter (claude|codex|gemini|grok|mistral|devin|cursor)

• probeInstalled (boolean, optional, default false): Run local --help probes and compare advertised flags against the declared contract — strongly recommended after any provider CLI upgrade. The probe reports missingFlags, extraFlags, acknowledgedExtraFlags (known upstream-only flags filtered from extraFlags), discoveredFlags, and stale-marker warnings.

Session Management Tools

session_create

Create a new session for a specific CLI.

Parameters:

• cli (string, required): CLI to create session for ("claude", "codex", "gemini", "grok", "mistral", "devin", "cursor")

• description (string, optional): Description for the session

• setAsActive (boolean, optional): Set as active session, default: true

Example:

{
  "cli": "claude",
  "description": "Code review session",
  "setAsActive": true
}

session_list

List all sessions, optionally filtered by CLI.

Parameters:

• cli (string, optional): Filter by CLI ("claude", "codex", "gemini", "grok", "mistral", "devin", "cursor")

Response includes:

• Total session count

• Session details (ID, CLI, description, timestamps, active status)

• Active session IDs for each CLI

session_set_active

Set the active session for a specific CLI.

Parameters:

• cli (string, required): CLI to set active session for

• sessionId (string, required): Session ID to activate (or null to clear)

session_get

Retrieve details for a specific session.

Parameters:

• sessionId (string, required): Session ID to retrieve

session_delete

Delete a specific session.

Parameters:

• sessionId (string, required): Session ID to delete

session_clear_all

Clear all sessions, optionally for a specific CLI.

Parameters:

• cli (string, optional): Clear sessions for specific CLI only

Utility Tools

list_models

List available models for each CLI.

Parameters:

• cli (string, optional): Specific provider to list models for ("claude", "codex", "gemini", "grok", "mistral", "devin", "cursor", or an enabled API provider name). When one or more [providers.<name>] API providers are enabled, the unfiltered response also carries an apiProviders array (each entry tagged providerKind: "api"); see API providers (HTTP).

Response includes:

• Model names and descriptions

• Best use cases for each model

• CLI-specific information

• defaultModel and defaultModelSource when a default is explicitly configured

• modelMetadata with source/confidence (fallback, config, env, observed)

• aliases and warnings when configured or when discovery degrades gracefully

The registry treats explicit configuration as authoritative. Bundled fallback models are low-confidence hints, and Gemini models observed in local session history are merged as low-confidence entries only; they do not become the default model.

Model registry environment overrides:

# Explicit defaults
CLAUDE_DEFAULT_MODEL=haiku
CODEX_DEFAULT_MODEL=<codex-model-id>
GEMINI_DEFAULT_MODEL=gemini-2.5-flash

# Additional models: comma/newline list, JSON array, or JSON object of model->description
GEMINI_MODELS='{"gemini-team-default":"Team-approved Gemini model"}'

# Aliases
GEMINI_MODEL_ALIASES='team=gemini-team-default'
LLM_GATEWAY_MODEL_ALIASES='codex.fast=gpt-5.3-codex-spark,gemini.fast=gemini-team-default'

# Deterministic config/discovery paths
CODEX_CONFIG_PATH=/path/to/config.toml
CLAUDE_SETTINGS_PATH=/path/to/settings.json
CLAUDE_SETTINGS_LOCAL_PATH=/path/to/settings.local.json
GEMINI_SETTINGS_PATH=/path/to/settings.json
GEMINI_HISTORY_ROOT=/path/to/.gemini/tmp

# Disable local model-history discovery
LLM_GATEWAY_DISABLE_MODEL_DISCOVERY=1

provider_tool_capabilities

Report the provider tool and feature capability catalog. Use this before orchestrating provider-specific requests so callers can distinguish supported controls, provider-owned configuration, ignored parity fields, and unsupported inputs.

Parameters:

• cli (string, optional): Provider filter ("claude", "codex", "gemini", "grok", "mistral", "devin", "cursor", "grok_api", or an enabled API provider name)

• includeSkills (boolean, default true): Include bounded local skill discovery

• includeProviderTools (boolean, default true): Include provider-native tools extracted from discovered skills

• includeUnsupported (boolean, default true): Include explicit unsupported/degraded input records

• includePaths (boolean, default false): Include raw local filesystem paths in discovery output

• refresh (boolean, default false): Bypass the short-lived capability cache

The response schema is provider-tool-capabilities.v2. Capability discovery is read-only and bounded; raw local paths are redacted unless includePaths is explicitly true, and secret-bearing auth files are not read.

Equivalent MCP resources:

• provider-tools://catalog: full provider catalog

• provider-tools://claude

• provider-tools://codex

• provider-tools://gemini

• provider-tools://grok

• provider-tools://grok_api

• provider-tools://mistral

• provider-tools://devin

• provider-tools://cursor

• provider-tools://<api-provider> for each enabled [providers.<name>] API provider

doctor --json also emits a compact provider_capabilities block with the same schema version, per-provider request tool names, supported feature names, unsupported input names, config-surface counts, discovery counts, and resource URIs. This block is intended for setup assistants that need a concise capability summary without local skill bodies or raw paths. When API providers are enabled, doctor --json additionally emits an api_providers health block; see API providers (HTTP).

cli_versions

Report installed CLI versions.

Parameters:

• cli (string, optional): Specific CLI to inspect ("claude", "codex", "gemini", "grok", "mistral", "devin", "cursor")

cli_upgrade

Plan or run an upgrade for one CLI.

Parameters:

• cli (string, required): CLI to upgrade ("claude", "codex", "gemini", "grok", "mistral", "devin", "cursor")

• target (string, optional): Package tag/version/target, default: latest

• dryRun (boolean, optional): Return the upgrade plan without running it, default: true

• timeoutMs (number, optional): Upgrade timeout when dryRun=false

Upgrade strategies:

• Claude latest: claude update

• Claude explicit target: claude install <target>

• Codex latest: codex update

• Codex explicit target: npm install -g @openai/codex@<target>

• Gemini latest: agy update (Antigravity self-update; explicit version targets are unsupported)

• Grok latest: grok update

• Grok explicit target: grok update --version <target>

• Mistral (Vibe): dispatches to the detected installer (pip/uv/brew); errors with guidance when none is detected (Vibe ships no self-update command)

• Devin latest: devin update (self-update; explicit version targets are unsupported)

• Cursor latest: cursor-agent update (self-update; explicit version targets are unsupported)

Example dry run:

{
  "cli": "gemini",
  "target": "latest",
  "dryRun": true
}

API providers (HTTP)

In addition to the spawnable CLI tools, the gateway can route requests to first-class HTTP/API providers (OpenRouter and other OpenAI-compatible endpoints, the Anthropic Messages API, and the xAI Responses API). These use Node's built-in HTTP client (node:https, or node:http for loopback endpoints) rather than spawning a CLI, and they run through the same request/job/validation/flight-recorder machinery as the CLI tools, so they reach parity for sessions, async jobs, dedup, retries, usage/cost capture, and cross-LLM validation.

Configuring a provider

API providers are declared as [providers.<name>] blocks in ~/.llm-cli-gateway/config.toml (override with LLM_GATEWAY_CONFIG). The <name> becomes the provider's identity across every tool and resource and must not collide with a spawnable CLI name (claude, codex, gemini, grok, mistral, devin, cursor); a collision is rejected with a warning and the provider is disabled.

# OpenRouter (OpenAI-compatible). The key is read from the named env var at
# request time and is never written to config, logs, the flight recorder, or
# the dedup key.
[providers.openrouter]
kind = "openai-compatible"                 # "openai-compatible" | "anthropic" | "xai-responses"
base_url = "https://openrouter.ai/api/v1"
api_key_env = "OPENROUTER_API_KEY"         # env var NAME, not the key itself
default_model = "x-ai/grok-2"
models = ["x-ai/grok-2", "anthropic/claude-sonnet-4.6"]   # optional allowlist: an explicit model must be listed (omit model to use default_model)
usage_include = true                       # OpenRouter token/cost reporting (usage:{include:true})

# Keyless-local: an openai-compatible provider on a loopback base_url (Ollama,
# llama.cpp) is enabled with no api_key_env at all.
[providers.ollama]
kind = "openai-compatible"
base_url = "http://127.0.0.1:11434/v1"
default_model = "qwen2.5"

A provider is enabled when its api_key_env resolves to a non-empty value, OR it is a keyless-local openai-compatible provider on a loopback base_url. base_url must use https unless it targets localhost/loopback. A schema-invalid single [providers.<name>] block disables only itself (with a warning) and leaves the other providers untouched; a TOML syntax error anywhere in the file is different, it makes the whole config fall back to defaults.

The pre-existing [providers.xai] block keeps its dedicated grok_api_request tool and xAI identity. It is also exposed through the generic surface like any other enabled provider, so a configured xAI key registers both grok_api_request and the generic api_xai_request (and the xai entry appears in apiProviders, models://xai, etc.).

Request tools

For each enabled provider, the gateway registers api_<name>_request (and api_<name>_request_async when async jobs are enabled). They accept the same shape as the CLI request tools:

• prompt (or the cache-aware promptParts { system?, tools?, context?, task }, mutually exclusive), optional system

• model (omit to use default_model; an explicit value is rejected if outside a configured models allowlist), maxOutputTokens, temperature, topP

• reasoningEffort (none|low|medium|high): forwarded only by the xai-responses adapter; accepted but ignored by the other kinds

• timeoutMs, optimizePrompt, optimizeResponse, forceRefresh, correlationId

• sessionId / createNewSession for continuity on the synchronous api_<name>_request (see below); on the async variant they are currently inert

Responses (and the 50 MB output cap, dedup window, cancellation, retention) behave exactly as for the CLI tools, and HTTP requests are logged to the flight recorder with status/usage/cost like everything else.

Continuity

Continuity is capability-typed per provider kind and never stores conversation content in the session record (the gateway's no-transcript-in-sessions invariant holds):

• xai-responses uses real server-side continuation: the gateway persists the provider's previous_response_id in session metadata and threads it back on resume, self-healing on a stale-handle 404.

• openai-compatible and anthropic are stateless-resend: the session tracks active/owner state for principal isolation, but the caller resends prior context (no server-side conversation handle exists).

Discovery surfaces

Enabled API providers appear, alongside the CLI providers, across the discovery surfaces. The model, capability, doctor, and resource surfaces below omit their API field entirely when no API providers are enabled, so their output is byte-identical to before; llm_process_health is the exception (it always carries an apiProviders array that is simply empty when none are enabled):

• list_models / list_available_models: an apiProviders array tagged providerKind: "api" with defaultModel and the optional allowlist, omitted entirely when no API providers are enabled.

• llm_process_health: an always-present outboundProviders.apiProviders array (empty when none are enabled) carrying the same projection plus each provider's circuit-breaker state.

• provider_tool_capabilities: per-api-kind capability metadata. Model + sampling + (for xai-responses) reasoning + continuity are supported; allow/deny tool lists, MCP servers, local skills, and workspace/worktree controls are not (those are CLI-only).

• doctor --json: an api_providers health block (kind, base_url, default_model, models, the key env var name, whether the key is present, and login guidance), omitted entirely when none are enabled. The optional flag doctor --json --probe-api-providers adds a per-provider endpoint-reachability result (a bare GET, treating any HTTP response as reachable). The probe is opt-in and off by default: a normal doctor run opens no socket and spends no tokens (reachable stays null).

• MCP resources: models://<provider> for every enabled provider and sessions://<provider> for continuity-tracked kinds, plus provider-tools://<provider>. The provider-subcommands:// resources stay CLI-only (API providers have no subcommands).

The API key value is never emitted on any of these surfaces (only the env var name and a presence boolean). Because base_url is config-supplied and may legally carry URL userinfo, the diagnostic surfaces (doctor, login guidance) redact any embedded credentials before displaying it; the actual request path and the reachability probe still use the original configured URL.

Security note

The resolved API key is excluded from payloadJson, the dedup key, logs, and the flight recorder. For ordinary non-Kit async jobs, the request prompt is persisted in plaintext in the async job store (SQLite at [persistence].path, default ~/.llm-cli-gateway/logs.db, or Postgres rows under backend = "postgres") and is not covered by secret redaction. This mirrors the CLI tools, whose prompt is persisted in argsJson whenever it is passed as a command argument rather than streamed over stdin. review_changes deliberately retains the complete fenced CLI prompt in expiry-bound payloadJson, stores only a hash marker in argsJson, and does not copy that prompt into the flight recorder. Personal Agent Config Kit durable rows use a separate privacy boundary: they do not persist compiled instructions or request arguments, and they withhold provider output and errors while retaining only the recovery and integrity state the Kit needs. Treat the job store as sensitive at rest. See Security Considerations and the Personal Agent Config Kit guide.

Session Management

How It Works

1. Gateway metadata, not transcripts: Session records track ownership, timestamps, active pointers, and provider metadata. They do not store a conversation transcript.

2. Storage backend: The session manager follows [persistence].backend in config.toml: sqlite/none keep ~/.llm-cli-gateway/sessions.json, postgres selects the PostgreSQL session manager. DATABASE_URL is a DEPRECATED override retained for compatibility; it warns once and is refused outright when it disagrees with an explicit [persistence].dsn. Note the PostgreSQL session store has no TTL eviction, where the file store reaps at 30 days; see the 3.1.0 notes.

3. Provider-native continuity: A gateway session ID is tracking metadata, not automatically a provider-native resume ID. Native behavior remains provider-specific: claude_request with continueSession:true uses Claude's latest conversation in a stable selected working directory and fails closed without workingDir or a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default. That workspace may optionally supply a gateway worktree. codex_request needs a real Codex UUID for sessionId, or resumeLatest:true, which resumes a previous session. Do not rely on which session --last selects or on the resumed working directory (under review).

4. Caller isolation: HTTP/OAuth callers can retrieve or reuse only sessions they own. Their session projection hides local paths and native provider identifiers.

5. Personal Kit: With Personal Agent Config Kit enabled, Claude, Codex, and Mistral use a separate, context-bound active-session pointer and retain a native continuation handle only in the current gateway process. See the Personal Agent Config Kit guide.

Session Workflow

// 1. Create a gateway tracking record when you need one.
// This ID is not automatically a native provider resume ID.
await callTool("session_create", {
  cli: "claude",
  description: "Debugging session",
  setAsActive: true,
});

// 2. Use the provider's documented native continuity control.
await callTool("claude_request", {
  prompt: "What's the bug in this code?",
  continueSession: true, // Claude's latest conversation in this working directory
});

await callTool("codex_request", {
  prompt: "Review the proposed fix.",
  resumeLatest: true, // or pass a real Codex UUID as sessionId
});

// 3. Inspect or remove gateway tracking records.
await callTool("session_list", { cli: "claude" });
await callTool("session_delete", {
  sessionId: "session-id-to-delete",
});

session_delete removes gateway tracking from caller-visible session surfaces. For file-backed sessions, it also runs cleanup for associated gateway-owned lifecycle resources, such as a managed worktree. A durably owned worktree session remains as a hidden cleanup-pending tombstone until Git removal is verified; failed removal is retried when that store is registered on the owning host. It does not delete a provider's own stored conversation.

Configuration

Environment Variables

• DEBUG: Enable debug logging (set to any value)

  DEBUG=1 node dist/index.js

• LLM_GATEWAY_APPROVAL_POLICY: Default approval policy for a Claude mcp_managed request when it does not pass approvalPolicy (strict, balanced, permissive). It has no effect on non-Claude adapters because they reject mcp_managed before launch.

  LLM_GATEWAY_APPROVAL_POLICY=strict node dist/index.js

• LLM_GATEWAY_APPROVAL_ALLOW_BYPASS: Applies only to Claude approvalStrategy:"mcp_managed". A direct bypassPermissions request, or an unverified execution posture such as settings, instruction overrides, plugins, additional directories, worktrees, or native continuation, is denied by default regardless of approval score. Ordinary managed Claude requests use --permission-mode acceptEdits.

  Set this to 1/true only to allow an explicit caller request through Claude's normal approval decision. The environment setting alone never escalates an ordinary managed Claude request. This setting does not enable mcp_managed on Codex, Gemini, Grok, Mistral, Devin, or Cursor: those adapters reject it before provider launch.

  LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1 node dist/index.js

• LLM_GATEWAY_TRUSTED_PRINCIPAL_HEADER: Name of an HTTP header carrying the authenticated user identity asserted by a trusted front door (any identity-aware reverse proxy / IdP). When set, the gateway adopts that header value as the request's ownership principal — but only for requests authenticated with the gateway's own static bearer token (i.e. the trusted upstream proxy), never from an arbitrary remote client. Off by default; IdP-agnostic. Lets a proxy-fronted multi-user deployment carry per-user identity into the gateway.

  LLM_GATEWAY_TRUSTED_PRINCIPAL_HEADER=x-gateway-principal node dist/index.js

• LLM_GATEWAY_OAUTH_REQUIRE_CONSENT / LLM_GATEWAY_OAUTH_CONSENT_SECRET: Opt-in human-consent gate for the built-in OAuth server. When enabled (REQUIRE_CONSENT=1, or implied by setting CONSENT_SECRET), /oauth/authorize renders an operator approval page (CSRF-protected) and issues an authorization code only after the dedicated consent password is entered — instead of auto-issuing. CONSENT_SECRET is the plaintext password (hashed in memory; or persist a consent_secret_hash in [http.oauth]). Off by default; remote OAuth refuses to enable consent without a secret to verify.

  LLM_GATEWAY_OAUTH_REQUIRE_CONSENT=1 LLM_GATEWAY_OAUTH_CONSENT_SECRET='choose-a-strong-code' node dist/index.js

• LLM_GATEWAY_CONFIG: Path to the gateway TOML config (default: ~/.llm-cli-gateway/config.toml). See Persistence configuration above for the [persistence] schema.

• LLM_GATEWAY_SKILLS_PATH: Extra local skill-pack roots to load at startup, separated by the host path delimiter (: on Linux/macOS, ; on Windows). These paths are appended after [skills].paths; ~/.llm-cli-gateway/skills still loads last when present.

• LLM_GATEWAY_LOGS_DB: Deprecated. When no backend is explicitly configured, selects backend = "sqlite" with this path (or backend = "none" when set to none). An explicit backend wins for job persistence. The variable still independently disables the recorder on every backend and paths it when SQLite is selected. Emits a deprecation warning at startup; migrate to config.toml.

  # Custom path
  LLM_GATEWAY_LOGS_DB=/var/log/gateway/logs.db node dist/index.js
  # Disable the recorder; with no explicit backend, also disable async persistence
  LLM_GATEWAY_LOGS_DB=none node dist/index.js

• LLM_GATEWAY_REDACT_LOGGED_SECRETS: Redact recognisable secrets (provider/cloud/VCS keys, bearer tokens, JWTs, PEM private keys, key=value secret assignments) from the prompt/system/response copies written to the flight-recorder log. Enabled by default; set to 0/false/off/no to store content verbatim. Only the audit log is affected — live sync responses and async llm_job_result output are never altered.

  # Opt out of flight-recorder secret redaction
  LLM_GATEWAY_REDACT_LOGGED_SECRETS=0 node dist/index.js

CLI-Specific Settings

Each CLI can be configured through its own configuration files:

• Claude Code: ~/.claude/settings.json

• Codex: ~/.codex/config.toml

• Gemini / Antigravity (agy): ~/.gemini/settings.json

Development

Project Structure

llm-cli-gateway/
├── src/
│   ├── index.ts              # Main MCP server and tool definitions
│   ├── executor.ts           # CLI execution with timeout support
│   ├── session-manager.ts    # Session management logic
│   └── __tests__/
│       ├── executor.test.ts  # Unit tests for executor
│       └── integration.test.ts # Integration tests
├── dist/                     # Compiled JavaScript
├── package.json
├── tsconfig.json
└── vitest.config.ts

Running Tests

# Run all tests
npm test

# Run unit tests only
npm run test:unit

# Run integration tests only
npm run test:integration

# Watch mode
npm run test:watch

Building

npm run build

Starting the Server

npm start

Error Handling

The gateway provides detailed error messages for common issues:

CLI Not Found

Error executing claude CLI:
spawn claude ENOENT

The 'claude' command was not found. Please ensure claude CLI is installed and in your PATH.

External Timeout / Legacy Timeout Option

Error executing codex CLI: Command timed out
Process timed out after 120000ms

Invalid Parameters

Prompt cannot be empty
Prompt too long (max 100k chars)

Logging

Logs are written to stderr (stdout is reserved for MCP protocol):

[INFO] 2026-01-24T05:00:00.000Z - Starting llm-cli-gateway MCP server
[INFO] 2026-01-24T05:00:01.000Z - claude_request invoked with model=sonnet, prompt length=150
[INFO] 2026-01-24T05:00:05.000Z - claude_request completed successfully in 4523ms, response length=2048
[ERROR] 2026-01-24T05:00:10.000Z - codex CLI execution failed: spawn codex ENOENT

Enable debug logging:

DEBUG=1 node dist/index.js

Troubleshooting

CLIs Not Found

Make sure the CLIs are installed and in your PATH:

which claude
which codex
which agy

The gateway extends PATH to include common locations:

• ~/.local/bin

• /usr/local/bin

• /usr/bin

• All ~/.nvm/versions/node/*/bin directories

Permission Errors

If you encounter permission errors, ensure the CLI tools have proper permissions:

chmod +x $(which claude)
chmod +x $(which codex)
chmod +x $(which agy)

Session Storage Issues

These file checks apply only to the default file session manager. When [persistence].backend = "postgres" selects PostgreSQL session storage, sessions.json is not authoritative. Do not delete or edit session storage while gateway processes, provider children, or Personal Agent Config Kit attempts are active.

1. Check file permissions after stopping local gateway processes:

ls -la ~/.llm-cli-gateway/

2. Use session_delete or session_clear_all for intentional gateway-record cleanup. Do not manually edit the file as a normal recovery method.

3. If a file-backed store is unreadable, inspect a copy only after the gateway is stopped:

cat ~/.llm-cli-gateway/sessions.json

For a stuck Personal Agent Config Kit attempt, do not reset sessions.json; follow the local-only recovery procedure in the Personal Agent Config Kit guide.

Performance

Timeouts

The gateway does not enforce a default execution timeout for LLM CLI requests.

If your MCP client/runtime enforces per-tool-call deadlines, use async tools (*_request_async + llm_job_status/llm_job_result) so long-running jobs can complete outside a single call window.

Concurrent Requests

The gateway supports concurrent requests across different CLIs. Each request spawns a separate process.

Security Considerations

• Input Validation: All prompts are validated (min 1 char, max 100k chars)

• API-provider keys: For [providers.<name>] HTTP providers, the gateway reads the key from the named environment variable at request time only. The resolved key is excluded from the persisted payloadJson, the dedup key, logs, and the flight recorder, and is never surfaced on the discovery/diagnostic surfaces (which report only the env var name and a presence boolean). base_url userinfo is redacted on the diagnostic surfaces. See API providers (HTTP).

• Prompt persistence at rest: Ordinary non-Kit async job rows store the request prompt in plaintext (HTTP payloadJson, and CLI argsJson whenever the prompt is passed as a command argument rather than streamed over stdin); this is not covered by secret redaction. review_changes is explicit about its different CLI layout: the complete fenced prompt is retained in expiry-bound payloadJson, persisted argv contains only its hash marker, and the prompt is not copied into the flight recorder. Personal Agent Config Kit durable rows do not persist compiled instructions or request arguments and withhold provider output and errors, retaining only privacy-safe recovery and integrity state. The SQLite job-store file (default ~/.llm-cli-gateway/logs.db, configurable via [persistence].path) is chmoded to 0o600 on non-Windows hosts; the Postgres backend stores the corresponding fields in database rows. Treat either backend as sensitive and scope/rotate it like any prompt log. Set [persistence].backend = "none" to disable the async job store entirely (the *_request_async / llm_job_* tools are then not registered). See the Personal Agent Config Kit guide for its narrower durable-record boundary.

• Command Execution: Uses spawn with separate arguments (not shell execution)

• No Eval: No dynamic code evaluation in our source (see "Socket alerts" below for the transitive ajv codegen case)

• Sandboxing: Consider running in containers for production use

• npm publish control: npm releases are gated by the generated prod-only shrinkwrap, release security audit, packed-consumer checks, and GitHub Actions Trusted Publishing with short-lived OIDC-derived npm publish credentials

• Release signing: GitHub release installer artifacts are signed with Sigstore keyless signing; verify SHA256SUMS.sigstore.json before trusting the checksum file

Socket alerts — context for reviewers

If you're vetting llm-cli-gateway through Socket or a similar supply-chain scanner, you'll see behavioural alerts and some dependency-ownership alerts. They are accurate descriptions of what the package does and what it depends on. The reviewed shellAccess and shrinkwrap entries are configured in socket.yml for repository/PR policy surfaces, but Socket's public package page may still display them for the published npm artifact; the rationale remains documented here and in the package.

The currently flagged surfaces are not new in 2.6.x: the 2.3.0, 2.4.0, 2.5.0, and 2.6.3 npm tarballs all include npm-shrinkwrap.json, and all include the same dist/executor.js child-process spawn surface used to run provider CLIs. The socket.yml policy for 2.4.0, 2.5.0, 2.6.0, and 2.6.3 is materially the same for shellAccess; this README now adds the missing shrinkwrap disclosure as well.

See socket.yml for the same context in machine-readable form.

Contributing

1. Fork the repository

2. Create a feature branch

3. Make your changes

4. Run tests: npm test

5. Build: npm run build

6. Submit a pull request

License

MIT. See LICENSE for details.

Support

For issues and questions:

• Open an issue on GitHub

• Check existing issues and documentation

• Review CLI-specific documentation for CLI-related problems

Changelog

See CHANGELOG.md for detailed release history.