DSH ⇄ ChatGPT Fusion Gateway

A minimal MCP bridge that keeps DeepSeek Harness (DSH) as the single agent and runtime, and uses ChatGPT Web only as a stateless, single-turn reasoning provider.

DSH owns agents, tools, permissions, sessions, compaction, subagents, workflows, goals, and real file/shell execution. The gateway simply relays one immutable DSH context snapshot to a ChatGPT "Brain" conversation, and relays back exactly one decision — a final answer, one action proposal, or a bounded read-only batch. The Gateway never executes the proposed action itself.

User
 ↓
DSH UI
 ↓
DSH Native Agent Loop  →  Context / Compaction
 ↓
ChatGPT Fusion LLM Adapter
 ↓
Fusion Reasoning Gateway          ← this repository
 ↓
ChatGPT Brain Worker              (a dedicated ChatGPT Web conversation + optional wake extension)
 ↓
final | action_proposal | readonly_batch
 ↓
Gateway validation
 ↓
DSH Permissions / ToolRuntime  →  DSH Native Tool  →  Next DSH Tick

Features

• Strict one-result reasoning contract — every request carries one complete, immutable snapshot; the Brain returns exactly one terminal result (final, action_proposal, or a bounded readonly_batch of read/glob/grep).

• Recoverable delivery — requests are journaled to disk; a crashed Gateway restores in-flight claims instead of losing them. Retries are idempotent via snapshot + nonce.

• Multi-worker failover — several Brain conversations can back the same model route; a stalled claim is released and re-delivered to a healthy standby with a rotated nonce.

• OAuth 2.0 (loopback) — optional authorization for the public ChatGPT connector (/brain-mcp), with rotating access/refresh tokens.

• Cross-platform — the gateway and Node daemon run on Linux, macOS, and Windows.

• Optional browser wake — a Chrome extension (browser-autowake/) polls loopback queue metadata and wakes a route-bound Brain tab when work is waiting.

Related MCP server: chatgpt-mcp

Requirements

• Node.js 18+ (uses the global fetch).

• git (and, on Linux, your configured command user) for the workspace tools.

DSH itself must be available for the DSH-native agent loop and tool runtime; the gateway detects the DSH-bundled node_modules automatically, or you can point it there with the DSH_NODE_MODULES environment variable.

Quick start

git clone <your-repo-url>
cd dsh-chatgpt-mcp
npm install

Put real credentials into config.json and model-bindings.json (all values ship as REDACTED_* placeholders), then start:

npm run daemon:start     # background, supervised with auto-restart
npm run daemon:status

…or run in the foreground:

npm start                # node server.strict.cjs

The gateway binds http://127.0.0.1:7688. Detailed architecture and invariants live in ARCHITECTURE.md.

Configuration

config.json

ownerPassword, staticToken, and the binding keys are required and must be filled in before use. On macOS/Windows there is no portable setuid helper, so sandboxed commands run as the current user and commandUser is ignored.

model-bindings.json

Maps each modelId + reasoningEffort route to a bindingKey. The bindingKey is the secret shared with the matching brain-worker-*.txt; keep both in sync.

Service control

Works identically on Linux, macOS, and Windows:

node scripts/daemon.mjs start      # start detached supervisor + server
node scripts/daemon.mjs stop
node scripts/daemon.mjs restart
node scripts/daemon.mjs status

These are also exposed as npm run daemon:start|stop|restart|status. Linux-only screen-based scripts remain in scripts/ (control.sh, public_quick.sh, …) for Cloudflare quick-tunnel and the headless Chrome worker pools.

Reasoning protocol (v4)

The gateways speaks a strict protocol to the ChatGPT "Brain":

fusion_wait_for_reasoning_request  →  one immutable snapshot + data-only action catalog
fusion_submit_reasoning_result     →  exactly one {final} | {action_proposal} | {readonly_batch}
fusion_fail_reasoning_request      →  terminal failure for the assigned snapshot

actionCatalog is data only — a description of the DSH native tools, never installed as ChatGPT/MCP tools. An action_proposal is a recommendation that DSH later executes under its own permission/sandbox policy; the Gateway does not run it.

The public endpoint is GET|POST /brain-mcp. The former /mcp and /fusion/tasks… routes return HTTP 410 (disabled).

Permissions

Uses the native three-level DSH model — read-only, workspace-write, danger-full-access — with workspace-write as the default.

Repository layout

server.strict.cjs        active gateway (strict v4; no legacy /mcp)
server.cjs               full gateway incl. workspace file/shell tools
config.json              runtime configuration (secrets are placeholders)
model-bindings.json      modelId → bindingKey map
scripts/daemon.mjs       cross-platform daemon (start/stop/restart/status)
scripts/*.sh             Linux deployment helpers (screen, cloudflared, Chrome workers)
brain-worker-*.txt       per-route prompts for dedicated Brain conversations
browser-autowake/        optional Chrome extension that wakes Brain tabs
tests/                   selftests and headless integration workers
state/  logs/            runtime artifacts (gitignored)

Testing

npm install
node tests/reasoning-contract-v4-selftest.mjs   # primary protocol test
# …plus oauth, worker-recovery, failover, context-overflow, payload-limit selftests

See the ## Verification section of the implementation notes for the full matrix.

Security notes

• Bind to 127.0.0.1 by default; only the quick-tunnel script exposes a public URL.

• Do not commit config.json or model-bindings.json with real secrets — this repository ships placeholders for that reason.

• /bin/setpriv is used on Linux to drop sandboxed commands to commandUser; on other platforms commands run as the current user.

DSH ⇄ ChatGPT Fusion is a community bridge and is not affiliated with or endorsed by OpenAI or DeepSeek. Use of ChatGPT is subject to OpenAI's terms of service.