AmendScope

Inspiration

A correction made in one AI task can leave three other tasks acting on the old truth. The usual fix is to paste a long explanation everywhere and hope nothing unrelated gets changed. That creates exactly the context mud I wanted to remove.

AmendScope comes from a four-question lens I have used while working with Codex: what is visible, what was said, which source governs, and what action is actually allowed? The product turns that lens into a live correction controller.

What it does

Describe a changed decision in plain language or use the included synthetic correction. AmendScope identifies the affected Codex recipients, proves which work is intentionally untouched, previews exact edits, blocks an action that is now invalid, and requires a short-lived human approval before any update packet is produced. A separate process then reads the result back and verifies the expected post-state.

For a new correction, an already-installed local Ollama model can propose the affected/untouched map and draft edits. The proposal is validated and fully editable. Codex receives only the smallest approved context packet—not a muddy transcript—and no recipient is contacted automatically in the judge build.

The four proofs

1. Affected: show every task that truly depends on the changed decision.
2. Untouched: show what was considered and deliberately left alone.
3. Blocked: deny at least one action made invalid by the correction.
4. Verified: independently reproduce the expected result after approval.

How I used Codex and GPT-5.6

GPT-5.6 through Codex materially helped turn the four-question mental model into typed source records, dependency-selected scope, approval manifests, the correction engine, denial boundary, separate-process verifier, adversarial tests, editorial interface, and judge package. Codex also helped compress the workflow until the payoff was visible in seconds. I retained product, boundary, identity, claims, approval, and release decisions.

Challenges

The central challenge was making the system feel alive without making false claims or letting an LLM become the authority. The deterministic correction path is therefore inspectable. Optional local AI proposes; schemas and policy validate; the human edits and approves; the verifier checks. Another challenge was proving restraint, so intentionally untouched work is first-class evidence rather than an omission.

Accomplishments

• A complete synthetic correction transaction with three affected tasks, one intentionally untouched task, three edits, one denied action, zero emitted actions before approval, and four verifier checks.
• 28 passing focused tests plus one intentional environment-dependent skip.
• Short-lived approval and zero-effect denial behavior.
• Optional localhost-only AI drafting with no OpenAI API key or provider billing path.
• A Codex intake packet that carries only approved, task-minimal context.

What I learned

Context quality is not measured by how much text moves. It is measured by whether the right recipient gets the smallest sufficient truth, with its source and permission boundary intact.

What's next

Native Codex recipient adapters, richer dependency discovery, signed correction receipts, team policy packs, and longitudinal learning from approved—not merely generated—corrections.

Built With

• typescript