RepoLanternEngineered with AELSHA™ ASTRA
Most AI tools ask you to trust the model. AELSHA™ ASTRA is how AELSHA™ builds AI agents so that you don't have to. Every step is gated, measured and shown to you.
What AELSHA™ ASTRA is
AELSHA™ ASTRA is AELSHA™'s engineering framework for taking AI agents to production and keeping them there. It starts from one idea: an agent is a software system first. It gets the same discipline as any system people depend on: a lifecycle with gates, tests that run on every change, measurements instead of impressions, and controls that hold even when a model misbehaves.
RepoLantern is an AI agent, and it is built this way. Everything below is RepoLantern's own engineering, and you can see each part of it in the app.
A lifecycle with gates
Every guide moves through the same stages, and each one has a gate it must pass before the next begins.
| Stage | What happens | The gate |
|---|---|---|
| Frame | The code is read, sized and priced | You approve the scope and the cost; nothing is spent before |
| Build | Files are summarised, folders combined, sections written | Each finished piece is saved, so a pause, a crash or a model switch never redoes or re-bills it |
| Evaluate | Every name, path and citation is checked against the code | A grounding score, and a rewrite of any section that names things that don't exist |
| Harden | Throughout: secrets removed, text aimed at AI tools flagged, spending capped | Repository text goes only to providers and models you agreed to, for guides and for questions alike |
| Release | The guide is published | Only to you, only with its score and the list of names it couldn't verify |
| Observe | Every model and tool call is recorded with tokens, time and cost | Your rating closes the loop |
Measured, not assumed
Every run reports five kinds of signal, and How it was made shows them for each guide:
- Telemetry: each model call and tool call, with tokens, time, retries and errors.
- Evaluation: the grounding score and what it checked.
- Quality and safety: secrets removed, text aimed at AI tools, sections repaired.
- Cost: spend by stage and by model, and what prompt caching and reused summaries saved.
- Outcome: whether readers found the guide and its answers helpful.
The operator sees the same signals across all guides in a control tower with a view for each role, alerts that name the playbook to follow, and error budgets for success rate and speed.
Code before judges
Where a check can be written as code, it is code, not another model's opinion. The checker itself is tested the way any safety system is:
- 12 golden guides from 5 real repositories in 4 languages reproduce exactly, with no false alarms.
- Planted mistakes: invented names planted into real guides are all caught.
- Blind rounds written by a separate reviewer: 94.3% and then 83.3% of planted names caught, and real names never flagged as invented.
- A regression in any of these stops a release. How we check every guide.
Readers close the loop: a guide of a public repository rated down becomes a candidate for the golden set, so what went wrong once is tested for ever after.
Built to keep going
Models run out, slow down and fail. RepoLantern plans for it:
- It paces itself to your model's limits, retries what is worth retrying, and stops calling a provider that keeps failing.
- When a quota runs out it pauses with everything finished kept, and you can continue on any other model or provider without paying again for finished work.
- Fallback models take over in the order you choose, each with its own spending cap.
- A question retried after a dropped connection is never asked or billed twice.
Open standards
Traces follow the OpenTelemetry conventions for AI agents, so you can load any guide's trace into your own tracing tool. Assistants connect over MCP and A2A, with tokens you can revoke (details).
You stay in control
RepoLantern acts on its own only for low-risk steps. Spending, consent, and the choice of model are always yours:
- It never spends before you approve, and never beyond your limit.
- Your key is never stored; it lives in your browser tab.
- A new provider or model never sees your code until you agree, whether it is writing the guide or answering a question.
- You can pause, switch, finish with what's done, or delete at any time.
The same lens on your code
AELSHA™ ASTRA also shapes what RepoLantern tells you about the repository you document. Readiness lists the production, cloud and agent signals found in your code, each with the files that show it: tests, CI, health checks, logging, cost control, guardrails, evaluation and human approval. These are automated signals, not an audit, and the guide says so (details).
See it for yourself
Start a guide and open How it works, live while it runs: the pipeline graph and every call as it happens. When it's done, open How it was made. AELSHA™ also helps teams build their own agents this way: get in touch.