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.

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 requests inside gateway-managed git worktrees for isolated multi-agent review and implementation loops.
• 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, and secure approval-gated dispatch. Seven 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, and public-demo-session. 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; the publish token is fetched at runtime from Azure Key Vault through GitHub OIDC to Entra.
• 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.

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: the release workflow builds Linux binaries on the local self-hosted runner, builds Windows/macOS binaries on GitHub-hosted runners, then 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: 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).

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 and resume conversations across all CLIs with persistent storage
• Gateway-owned worktrees: Run any sync or async provider request inside a managed git worktree, with per-session reuse and cleanup
• Token Optimization: Automatic 44% reduction on prompts, 37% on responses (opt-in)
• Correlation ID Tracking: Full request tracing across all LLM interactions
• Cross-Tool Collaboration: LLMs can use each other via MCP (validated through dogfooding)

Observability

• SQLite Flight Recorder: Every request/response logged to ~/.llm-cli-gateway/logs.db with correlation IDs, token usage, duration, retry counts, and circuit breaker state. Browse with Datasette: 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. 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.
• 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: validate_with_models, second_opinion, compare_answers, red_team_review, consensus_check, ask_model, and synthesize_validation, with durable signed receipts via validation_receipt and the validation-receipt://{validationId} resource.

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"]
    }
  }
}

Prerequisites

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

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/cli-overview

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 auth login
# 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, so any name is passed through and Vibe validates availability. The gateway's programmatic-mode default is auto-approve; pick a stricter mode explicitly if you need approval gates.
• allowedTools is allow-list only — the gateway emits one --enabled-tools <tool> flag per entry. disallowedTools is accepted in the schema for caller-side parity but is silently ignored at the CLI boundary (a logger.info warning records the no-op).
• 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; disabling auth or using a no-auth connector path is not a filesystem bypass.

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.
• 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.
• 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 immutable receipt of a terminal cross-LLM validation run by validationId (returns minted | pending | expired_unminted | not_found, own-or-not-found). format: "markdown" renders a human-readable report; includeRawResponses inlines provider answer text. 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.

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
• continueSession (boolean, optional): Continue the active session
• createNewSession (boolean, optional): Always create a new session
• forkSession (boolean, optional): Fork the resumed session instead of appending to it
• allowedTools (string[], optional): Restrict Claude tools to this allow-list
• disallowedTools (string[], optional): Explicitly deny listed Claude tools
• permissionMode (string, optional): Claude permission mode (default|acceptEdits|plan|auto|dontAsk|bypassPermissions); preferred over dangerouslySkipPermissions
• dangerouslySkipPermissions (boolean, optional): Deprecated — maps to permissionMode: "bypassPermissions"; permissionMode wins when both are set
• agent (string, optional): Named sub-agent to run as
• agents (string, optional): Inline agent definitions JSON
• systemPrompt / appendSystemPrompt (string, optional): Replace or extend the system prompt
• 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
• 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
• excludeDynamicSystemPromptSections (boolean, optional): Trim dynamic system prompt sections
• approvalStrategy (string, optional): "legacy" (default) or "mcp_managed"
• approvalPolicy (string, optional): "strict", "balanced", or "permissive"
• mcpServers (string[], optional): Names of MCP servers to expose to Claude (default: none). The gateway resolves each name to a launch command from its local registry / Codex MCP config; unknown names are reported as unavailable. Configure the servers your deployment uses in the gateway environment.
• strictMcpConfig (boolean, optional): Require Claude to use only supplied MCP config, default: true (request fails if any requested server is unavailable)
• 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
• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ)
• 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 a stdio/local legacy setting and does not allow 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 bypass flags
• approvalStrategy (string, optional): "legacy" (default) or "mcp_managed"
• approvalPolicy (string, optional): "strict", "balanced", or "permissive"
• mcpServers (string[], optional): MCP servers expected for Codex execution context
• sessionId (string, optional): Session identifier for tracking
• resumeLatest (boolean, optional): Resume the most recent Codex session in the current cwd (codex exec resume --last); 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)
• 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
• ignoreRules / ignoreUserConfig (boolean, optional): Codex --ignore-rules / --ignore-user-config
• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ)
• 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:

• approval: Approval decision record when approvalStrategy="mcp_managed"
• mcpServers: Requested MCP servers 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.

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)
• 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. Only default (prompted execution) and yolo (emits --dangerously-skip-permissions) are accepted; auto_edit and plan are rejected with an error.
• approvalStrategy (string, optional): "legacy" (default) or "mcp_managed"
• approvalPolicy (string, optional): "strict", "balanced", or "permissive"
• 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)
• outputFormat (string, optional): text only. Antigravity print mode emits text; json and stream-json are rejected.
• mcpServers, 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. Retained in the schema for caller parity.
• yolo (boolean, optional): Auto-approve all; equivalent to approvalMode: "yolo". Emits --dangerously-skip-permissions
• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ)
• 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

Response extras:

• approval: Approval decision record when approvalStrategy="mcp_managed"
• mcpServers: Requested MCP servers 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-build, latest)
• 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). 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" (default) or "mcp_managed"
• approvalPolicy (string, optional): "strict", "balanced", or "permissive"
• mcpServers (string[], optional): MCP server names tracked for approvals (Grok manages its own MCP config via grok mcp)
• 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
• bestOfN (integer, optional): Run the task N ways in parallel and pick the best (headless only)
• check (boolean, optional): Append a self-verification loop (headless only)
• 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) — targets an isolated leader process, e.g. a local/branch Grok build
• nativeWorktree (boolean|string, optional): Grok's own --worktree flag (true → bare, string → 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)
• 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-build",
  "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 live for 30 days by default; 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 job-store 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
dedupWindowMs = 3600000
acknowledgeEphemeral = false                # required to enable async tools with memory backend

# 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

Backends:

• sqlite (default) — durable, file-backed. Safe for single-instance deployments.
• postgres — durable PostgreSQL-backed async job, dedup, orphan recovery, HTTP job, and validation receipt storage. 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.

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

• LLM_GATEWAY_LOGS_DB / LLM_GATEWAY_JOBS_DB — none selects backend = "none"; any other value selects backend = "sqlite" with that path.
• 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 keeps its own (longer) [persistence].retentionDays retention, 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, all gateway data 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 files:

• ~/.llm-cli-gateway/logs.db (async jobs + flight recorder)
• ~/.llm-cli-gateway/sessions.json (CLI sessions)
• ~/.llm-cli-gateway/config.toml (resolved config)

This is usually what you want — session_list from repo A shows 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. SQLite WAL mode makes concurrent access from multiple gateway subprocesses safe.

If you instead want per-project isolation (e.g. unrelated repos shouldn't share session lists or risk false dedup hits), point each project at its own config file. 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 to its own SQLite file; sessions, jobs, and dedup state are scoped to the repo. Other repos keep using the global default. llm_process_health.persistence.sources.configFile lets an agent confirm which config it's 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 persistence isolation"
spec              = "https://github.com/verivus-oss/llm-cli-gateway#per-project-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 async job store, sessions, dedup state, and flight recorder are scoped to THIS repository instead of the per-user default at ~/.llm-cli-gateway/."
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      = ["PerProjectIsolationComplete"]

[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. If set, the legacy env var will override the new config and the deprecation warning will fire at every gateway boot. The agent reports this as a finding and asks the operator to unset 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:isolation-verification","OUT:per-project-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 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.

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). 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. Emitted as --agent <name>. Defaults to auto-approve in programmatic mode.
• allowedTools (string[], optional): One --enabled-tools <tool> flag per entry (allow-list only).
• disallowedTools (string[], optional): Accepted for parity with the other providers; ignored at the CLI boundary with a logged warning.
• 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" (default) or "mcp_managed"
• approvalPolicy (string, optional): "strict", "balanced", or "permissive"
• mcpServers (string[], optional): MCP server names tracked for approvals (Vibe manages its own MCP config via vibe mcp)
• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ)
• 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): 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

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). 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), smart (also auto-runs actions a fast model judges safe), dangerous (auto-approves all). Omit to use Devin's headless default
• 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 cwd (--continue)
• createNewSession (boolean, optional): Force a new session
• optimizePrompt / optimizeResponse (boolean, optional): Token-efficiency optimisation, 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

cursor_request

Run a Cursor Agent CLI request synchronously. Defaults to 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] and [acp.providers.cursor].runtime_enabled are enabled; current ACP routing accepts only prompt/model/session inputs and rejects CLI-only options such as mode, workspace, sandbox, force, and trust.

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 enabled in gateway config and rejects unsupported Cursor CLI-only controls instead of dropping them. 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): Cursor workspace path or name (--workspace); remote HTTP/OAuth callers must pass a registered workspace alias, while local stdio callers may pass paths
• addDir (string[], optional): Additional workspace roots (one --add-dir per entry); remote HTTP/OAuth callers must use registered workspace roots
• sessionId (string, optional): Cursor chat/session ID to resume (--resume <id>). The gw-* id minted for a brand-new gateway session is not resumable through sessionId; continue with resumeLatest: true
• resumeLatest (boolean, optional): Resume the most recent Cursor chat (--continue)
• createNewSession (boolean, optional): Force a new session
• approvalStrategy (string, optional): "legacy" (default) or "mcp_managed"; under MCP-managed approval, high-impact Cursor flags (force, trust, or sandbox: "disabled") are denied unless bypass approval is explicitly allowed
• approvalPolicy (string, optional): "strict", "balanced", or "permissive" override for MCP-managed approval
• optimizePrompt / optimizeResponse (boolean, optional): Token-efficiency optimisation, 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

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 approval strategy fields as their sync variants:

• approvalStrategy: "legacy" (default) or "mcp_managed"
• approvalPolicy: "strict"|"balanced"|"permissive" override
• mcpServers: Names of requested MCP servers, resolved against the gateway's local registry / Codex MCP config
• claude_request_async also supports strictMcpConfig and fails fast when requested servers are unavailable

llm_job_status

Return lifecycle status (queued, running, completed, failed, canceled, orphaned) and metadata for an async job.

llm_job_result

Return captured stdout/stderr for an async job (with configurable max chars per stream).

llm_job_cancel

Cancel a running async job.

approval_list

List recent 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", or "mistral"

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

llm_process_health

Report gateway process health: async-job manager state plus the resolved persistence block (backend, dbPath, config sources). Use it to confirm which config file and SQLite paths the gateway is actually running under.

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 JS