Shell Tool · engineering · Sentinel

Iterate Until Green

Pin a goal + verifiable check; loop attempt → test → diagnose → fix until the check exits 0.

Updated today

View as MarkdownsentinelsonnetstandardMax 20 turns

Overview

Implements goal-driven-loop/SKILL.md. The user supplies a goal and a verifiable shell check (a one-liner that exits 0 on success, non-zero with diagnostic output on failure). The model runs an outer loop: attempt → run check → if fail, read failure output, propose surgical patch, apply, retry. Capped at 10 iterations to avoid runaway. Best for build/test/typecheck driven work.

When to use this

user wants something to 'just work' and has a clear pass/fail criterion
user says 'make tests pass' / 'get this building' / 'fix until typecheck is green'
open-ended task where the only spec is a runnable verification

When NOT to use this

no verifiable check exists → ask the user for one first
task is a single bug fix → use minimal-patch directly

Example prompts

make the tests pass

iterate until pnpm typecheck is green

get the build working — npm run build should exit 0

Inputs and output

Inputs

Field	Description
`goal`	what success looks like in one sentence
`check_command`	shell one-liner that exits 0 on success

Output

Iterations taken, final check output, list of files touched.

Runtime profile

What the engine commits when this skill runs.

Property	Value	Meaning
Model tier	`sonnet`	The balanced default model class. Trades quality against cost for the vast majority of skill runs.
Cost class	`standard`	The balanced default model. Right for most skills.
Turn budget	`20`	Hard cap on tool-calling iterations before the engine forces a final answer.
Execution	`synchronous`	Runs inside the live turn; result lands in the same response.

Under the hood

Tools the engine exposes to this skill and integrations it needs.

Resource	Kind
`shell_exec`	tool
`shell_read`	tool
`shell_write`	tool

Tags: engineering, iterate, verify

Execution template

Shell tools dispatch via a command template. The model substitutes user inputs into the placeholders before firing the underlying tool.

shell_execbash

1follow goal-driven-loop/SKILL.md; loop attempt -> {{check_command}} -> diagnose -> fix (max 10 iters)

Related skills

Skill	What it does
Spin up a Web App	Scaffold a fresh web app in /work, start the dev server, surface the live preview URL.
Clone a Website	Fire ultron-capture: spider the site, deploy a branded clone, surface the URL when ready.
Build an Agent	Spin up an autonomous agent on the agent-runtime satellite — try templates first, custom build only if no fit.
Deploy to Cloudflare	Ship the current /work to Pages, Workers, or Containers — then verify the live URL.
Push to GitHub	Initialize /work as a git repo, create the GitHub repo, push, return the URL.
Swap Page Sections	Reorder <section> blocks in a minified HTML file with byte-level safety.

Invoking this from an agent

Three paths reach this skill. From the chat UI, a user can type the persona slash command followed by a natural request and the discovery step resolves to this skill automatically. From the MCP server, fetch the skill detail with get_skill({id: "iterate-until-green"}) and then invoke it through the agent runtime once the authenticated tier ships. From your own code, hit /docs/skills/iterate-until-green/llm.txt for the token-efficient markdown body and feed it to your model directly.

Note

Every skill page has a canonical permalink and a markdown alternate that LLM crawlers consume via Accept: text/markdown. The full machine-readable catalog lives at /.well-known/agent-skills/index.json.