- Home
- Skills
- Agents & Orchestration
- Codex x OpenCode MCP: Multi-Harness Delegation Bridge
More screenshots
Works with the AI Agents you already use
Codex x OpenCode MCP: Multi-Harness Delegation Bridge
Codex keeps the plan and the acceptance call. Six harnesses do the work. One validated contract between them.
$14
Codex x OpenCode MCP: Multi-Harness Delegation Bridge
Example session with this skill installed
Delegate this to the Command Code harness: our NDJSON reader drops the final result line whenever the CLI exits while stdout is still flushing. Fix it in C:\dev\billing-api and prove it with a test that emits a result line and exits immediately.
- Read your context and instructions
- Compiled the codex x opencode
Summary
Fixed the dropped result line in the Command Code adapter. Stream parsing is now incremental and drains stdout on exit, so the final result frame is captured even when the CLI terminates mid-flush. Added a regression test that emits one result line and exits at once.
| File | Change |
|---|---|
src/harness/commandcode.ts | Parse NDJSON per line, drain remaining buffer on close, keep the last result frame |
test/harness-commandcode.test.ts | New case: result frame followed by immediate exit |
{
"task_id": "task_01J9XK4M",
"status": "complete",
"summary": "Command Code adapter keeps the final NDJSON result frame when the CLI exits during a flush.",
"what_changed": [
"Moved stdout parsing to an incremental line reader",
"Drained the remaining buffer on process close"
],
"changed_files": ["src/harness/commandcode.ts", "test/harness-commandcode.test.ts"],
"artifacts": [".codex-x-opencode-mcp/tasks/task_01J9XK4M/stdout.log"],
"validation": {
"typecheck": "pass",
"lint": "not_run",
"tests": "pass",
"build": "pass",
"custom_checks": ["regression test fails on the old parser, passes on the new one"]
},
"evidence": ["tsx --test test/harness-commandcode.test.ts: 13 pass, 0 fail"],
"risks": ["A CLI that writes partial lines without a trailing newline would still be dropped"],
"assumptions": ["The CLI always writes the result frame as one complete line"],
"unresolved_issues": [],
"decisions_required": [],
"agents_used": ["implementer", "tester", "reviewer"],
"models_used": ["harness standard tier", "harness strong tier"],
"tools_used": ["terminal", "edit", "read"]
}
The reviewer role read the diff independently and found no further change.
Next steps
- Run your own suite with
npm testbefore merging. - If your CLI writes partial lines, raise the stdout cap in
limits.
Connects securely to your tools. The creator never sees your data.
About this skill
What this is
Codex decides what to build and whether the result is acceptable. This bridge hands the approved brief to a real coding agent, then returns a result Codex can check before it accepts anything.
A prompt hands off text. This hands off a task: the workspace path is verified, the brief is schema-checked, the result contract is validated, failed results are withheld from acceptance, and corrections resume the same session where the harness supports it. Where it does not, the bridge replays a compressed transcript instead of losing context.
Six execution harnesses, one config
| Harness | Notes |
|---|---|
| OpenCode 1.18.32 | Default. Validated end to end. |
| Codex CLI | Native JSON schema output, session resume. |
| Claude Code | Task tool subagents, JSON schema output. |
| Command Code | -p headless mode, NDJSON results, session resume. |
| Cursor CLI | Replays a compressed transcript between cycles. |
| Copilot CLI | Replays a compressed transcript between cycles. |
Switch harnesses by editing harness.config.json. A deployment routed to a CLI harness
does not need OpenCode installed at all.
Ten orchestration roles on every harness
Every cycle prompt carries the same vendor-neutral protocol: intake, plan, plan review, bounded delegation, smallest change, real verification, independent diff review, iterate until green, stop only for an orchestrator decision, report only evidence.
The ten roles are orchestrator, intake, planner, plan-reviewer, implementer,
researcher, reviewer, tester, advisor, and inspector. Rename any of them in
config if your team uses different words.
Each role binds to a model tier of the harness that actually runs the task. The bridge reads that harness's own model block, so a harness with subagents spawns them on models it really offers, and switching harness changes the model ids without touching the protocol.
The seven tools
harness_execute, harness_status, harness_result, harness_inspect,
harness_correct, harness_decision, harness_stop. The older opencode_* names stay
registered as deprecated aliases, so existing setup keeps working.
What you get
- The bridge source in TypeScript, with 217 automated tests and a strict typecheck.
- A Windows installer that copies, builds, installs the skill files, and rewrites only its
own block in
config.toml, keeping a backup next to it. - A skill file for your orchestrator, plus setup, troubleshooting, and three end-to-end usage walkthroughs.
- Result contract JSON schemas and a per-task cost cap.
- MIT license. Read the source before you trust it with your repo.
Requirements
Windows 10 or 11, Node.js 22 or newer, and at least one harness CLI installed and authenticated by you. Harness CLIs are never bundled. OpenCode, Claude Code, Cursor, and Copilot are proprietary and governed by their vendors' terms. Codex CLI is Apache-2.0.
What it does not do
It does not install, license, or authenticate a harness for you. It does not make a single-agent harness into a multi-agent one; those harnesses run the same ten-role sequence themselves. macOS and Linux are not validated in this release.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Windows only. Requires Node 22+ and pre-authenticated CLI harnesses. No support for macOS or Linux. Pinned to OpenCode 1.18.32 behavior.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 day ago
- Software
Node.js
Git
Harness CLIs are not bundled and each needs your own account. OpenCode is the default and the only harness validated end to end; the other five are covered by automated tests.
Permissions
Terminal and file access are required because the bridge spawns your harness CLI inside the task workspace and reads the resulting diff. Network access comes only from that CLI reaching its own provider. The bridge itself sends no telemetry.