Each chunk is built by a cheap Flash model, then reviewed by a different model family before it can merge.
Cheap workers carry each chunk, an independent reviewer checks it, and the state on disk lets you resume any session.
Every chunk is checked, then reviewed by a different model family, then accepted. Defects surface inside the run, not on your screen.
45 blocking defects independent review caught before merge, across the two runs that built this site.
A top-tier model sets the plan and integrates; Flash workers carry each chunk, so implementation runs at worker rates.
7.82 US dollars estimated worker and reviewer spend for both runs, 156 model calls. A local estimate, not a billed figure; the coordinating model is not in it.
The run lives on disk. Clear the session, kill the process, come back tomorrow: resume picks up where the work stopped.
Split a feature into independently deliverable chunks, write acceptance criteria that can be checked, and judge results at chunk boundaries.
This site was rebuilt through Amaleh's own runs: 156 model calls, 45 blocking defects caught by review before merge, an estimated 7.82 US dollars of worker and reviewer spend.
The figures were read from the run records on 2026-09-23. The cost is pi's local estimate of the worker and reviewer calls, not a billed figure.
The coordinator holds the direction. Everything inside a chunk runs without it and comes back as one outcome to inspect and integrate.
Brass marks the coordinator sending chunks out to the routed worker models. Violet marks the bounded questions workers put to Jev. Teal marks a chunk reviewed by another family and accepted back.
The expensive model, kept to minimal turns. Clarifies intent, chunks work, defines acceptance criteria and integrates results. One delegate call per chunk.
The routed Flash families (DeepSeek, GLM, MiMo and Stealth) each own a chunk end to end: implementation, checks and repair cycles. Uncertain inside the chunk, they consult Jev directly instead of escalating.
Always the other model family, reading only. It verifies each chunk with structured coverage and routes findings back into the worker's repair loop rather than to the coordinator.
A cheap decision model answering bounded either/or questions for workers and the coordinator. TypeScript code, not any model, enforces dependencies, ownership, checks and review coverage.
Clone the repository and run both commands from the checkout root. Invoke /amaleh with your task afterwards; the skill handles the workflow and resume steps.
git clone https://github.com/mhamri/amaleh
cd amaleh
bun amaleh/scripts/run.ts doctor
bun amaleh/scripts/run.ts installWhat install actually does
Links, never copies. Install links the canonical skill directory into the default Claude and Codex skill directories, so the checkout stays the single source of truth.
Refuses conflicting targets. A conflicting destination is refused rather than overwritten, so an existing installation is never silently replaced. Keep the checkout in place afterwards.
doctor checks local setup before anything network-facing; the startup preflight covers network connectivity separately.
Fallback launcher. Node 24 or newer runs the erasable TypeScript directly when Bun is absent, so the same commands work without a second toolchain.
Configure pi with OpenRouter, or provide OPENROUTER_API_KEY through your environment. The runtime reads existing credential configuration and never prints or writes the key into artifacts.
Amaleh is built by Mohammad Hossein Amri, a software engineer in Kuala Lumpur, Malaysia. He has over 13 years in the industry, works at menumiz (AU), and writes C#/.NET, TypeScript and cloud applications. He built Amaleh to direct cheap models instead of typing every edit with an expensive one.
The name is the Persian word عمله, pronounced Ah-mah-leh, meaning workers or laborers.
Five documentation pages take the workflow apart, and the case study proves each claim with measured figures from the runs that built this site.
From the problem Amaleh solves through installation, the workflow, review and recovery, to a command reference drawn from the skill's own reference files.
Read the documentationEach claim this page makes, paired with a figure from the run records: 156 model calls, 45 blocking defects caught by review, an estimated 7.82 US dollars of worker and reviewer spend, and the honest limits beside them.
See the case studyThe skill's reference documents ship unchanged with this build, where every claim on this site is written out in full.
Browse the sourcesThe skill, its source and its issues are on GitHub.
Install once, invoke /amaleh with your task, and get back work each chunk of which was checked and independently reviewed.
git clone https://github.com/mhamri/amaleh