- Home
- Skills
- Agents & Orchestration
- Orkestra: Guarded, Test-First Pipeline for Coding Agents
Orkestra: Guarded, Test-First Pipeline for Coding Agents
Runs every coding request through the same local steps: plan, test-first build, gates, review and an evidence report. Blocks commit, push, deploy and secret reads unless you ask.
Free
Works with the AI Agents you already use
Orkestra: Guarded, Test-First Pipeline for Coding Agents
Example session with this skill installed
/orkestra:run add CSV export to the order list
- Read your context and instructions
- Compiled the orkestra
Summary
Orkestra opened a tracked run, sized it, backed up the uncommitted work, and printed the one next step. The evidence report starts empty and fills in as the run moves.
$ orkestra start "add CSV export to the order list"
run started: 001-add-csv-export-to-the-order-list in ~/demo
size: M (inferred from request; override with --size XS|S|M|L)
baseline: git, no commits yet · uncommitted files: 1
NOTE: the working tree already had uncommitted changes. They are backed up and will never be attributed to this run.
next action: Classify the request (size XS/S/M/L, project, risk) into request.md → `orkestra phase next`.
suggested classification: Type: feature · Size: M · Risk: medium
$ orkestra watch --once
phase intake › discover › clarify › plan › approve › implement › verify › review › deliver › done
status ● intake phase open · run state updated 0s ago
$ orkestra next
next [intake]: classify the request: append `## Classification` to request.md
why: request.md has no classification yet
| Evidence field | What it records |
|---|---|
| Gates | Baseline vs last verify, per test command |
| Tasks | Red (failing) and green (passing) exit codes |
| Test integrity | Deleted tests, added skip/only/xfail |
| Review triage | Fixed, deferred, rejected findings |
| Commit/push | "not done" unless you asked |
Next steps
- Let the agent continue: it follows
orkestra nextuntil the report is written. - Run
orkestra report evidenceat any time to see the machine-written proof.
Connects securely to your tools. The creator never sees your data.
About this skill
What it does
Orkestra walks your coding agent through the same fixed phases on every request: intake, discover, clarify, plan, approve, implement, verify, review, deliver. A local CLI keeps the state and hooks enforce the rules, so the process does not depend on the model remembering it.
Problems it solves
- You ask for a change and then read a long diff to find out what the agent touched, skipped or guessed.
- Agents commit, push or deploy when you did not ask.
- "Tests pass" with no proof, or tests quietly weakened to go green.
- Agents writing into the wrong project or reading secret files.
How it works
- Hook guards: git commit, push, merge, reset and stash are refused unless your current message asks for them. Reading or writing
.env*, private keys and service-account JSON is refused. Deploys,ssh, dependency installs outside the approved plan and writes outside the run's project are refused during a run. - Test-first evidence: each task records a failing test (red) and then a passing one (green). The test cannot change between the two, and deleted tests or added skip/only/xfail are flagged.
- No-regression gates: your project's test commands run before and after the change; a regression blocks delivery.
- Agent team: read-only scouts, an architect, a plan critic, a TDD implementer, a reviewer, an adversarial tester and a verifier, each with a fixed role.
- Cost-aware: small requests take a short path with no scouts, no plan critic and no per-task review. Cheap models do mechanical work and the strongest model is kept for planning and the final review.
- Live view:
orkestra watchandorkestra nextshow the phase, the activity and the one next step.
Who it is for
Developers who use Claude Code, Cursor, GitHub Copilot CLI or Pi and want the same disciplined process every time, with proof instead of promises.
Free and open source
MIT license, Python standard library only, no telemetry, nothing leaves your machine except what your model provider's own client sends.
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.
Guards are filters on tool calls, not a sandbox: the agent still runs as you, on your files. Shell guards read commands without running them, so obfuscated commands can slip through (the activity log shows them afterwards). Claude Code and Cursor hooks fail open if the hook crashes; Copilot fails closed. The proof is only as strong as your project's tests: a project without tests has no regression net. Windows is less tested than macOS and Linux.
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
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed today
- Free with an Agensi account
- Accounts
GitHub
- Software
Python 3.9+
GitOrkestra CLI (free, github.com/soft-ai-solutions/orkestra)Claude CodeCursorGitHub Copilot
- Optional
- Telegram bot (remote control through the bridge)
Install the Orkestra CLI once; the skill shows the steps and never runs the installer without your yes. On Windows use python install.py.
Permissions
Allowed Hosts
File Scopes
Terminal: runs the orkestra CLI and your project's own test commands. Read/Write Files: the agent edits your project; run state is kept in .orkestra/ and ~/.orkestra/. Network: only git clone of the open-source repo during install. Hooks refuse secret-file reads, commits, pushes and deploys unless you ask.