StarShade Context Compiler

Compile large project histories into small, auditable context capsules—and fail closed when trust breaks.

Comment

Inspiration

Long-running Codex projects can accumulate large histories, plans, control files, and evidence. Repeatedly carrying every source into every task is wasteful, but ordinary summaries can hide whether the selected context is current or where each value came from.

I built StarShade Context Compiler to test a safer local approach: preserve complete sources, compile a smaller auditable working context, and fail closed to full rehydration whenever trust checks break.

What it does

  • L0 loads an accepted Context Capsule plus a cursor-bound Delta.
  • L1 adds only explicitly named, allowlisted, in-root sources.
  • L2 fully rehydrates an immutable source snapshot after stale cursors, conflicts, hash drift, malformed inputs, provenance failures, parity failures, or concurrent changes.
  • Every compiled field carries its source pointer, source hash, and value hash.
  • HTML and JSON reports explain the selected route, exact sources, escalation reasons, provenance, and payload bytes.

How I built it

I used Codex throughout planning, implementation, adversarial review, repair, testing, evidence preparation, and packaging. GPT-5.6 was used in the installed-plugin validation task to exercise the normal L0 route, allowlisted L1 route, and copied-fixture fail-closed L2 route.

The project is packaged as an installable Codex plugin with a deterministic local Python engine, fixed JSON schemas, SHA-256 manifests, two fictional fixtures, automated tests, and readable audit reports.

Verification

The installed plugin was tested with its bundled synthetic fixtures only:

  • Both scenario matrices produced 1 L0, 1 L1, and 12 L2 cases.
  • Fixture 1: 89,759 full canonical JSON UTF-8 bytes versus 4,728 L0 bytes.
  • Fixture 2: 53,888 full canonical JSON UTF-8 bytes versus 4,679 L0 bytes.
  • Required provenance verified: 13/13.
  • Independent routing-critical parity: true.
  • Tampering caused a truthful full L2 rehydration instead of accepting compact context.

These are fixture-specific canonical JSON UTF-8 payload-byte measurements. They are not claims about tokens, Codex weekly limits, quota, billing, latency, or semantic quality. Actual weekly usage-limit reduction remains unproven.

Privacy and scope

The Build Week MVP reads only its two included fictional fixtures. It does not access private project data, call cloud APIs, infer arbitrary project schemas, or publish telemetry.

Built With

Share this project:

Updates