PITHOS Workbench Continuity Layer

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for PITHOS Workbench Continuity Layer

Inspiration

Long-running AI work has a continuity problem: new tasks may have the model capability to help but not the exact decisions, tests, permissions, and active state needed for the next answer. Rebuilding context is slow, while dumping an entire archive into a prompt can introduce stale, irrelevant, or restricted material.

What it does

PITHOS Workbench Continuity Layer treats continuity as a bounded function over a durable file tree. A command is deduplicated, assigned to an explicit route, filtered by freshness and permission scope, reduced to the smallest sufficient record set, and bound before the answer is written. The answer retains inspectable provenance, and duplicate requests reuse the existing process-local mount.

The portable demo uses a fictional Northstar release review. It mounts exactly two records: the current release decision and latest test result. It excludes a stale decision, a restricted incident, and an unrelated project record for three distinct reasons. Eleven behavioral assertions and a separate package verifier pass.

How I built it

The zero-install reference implementation uses synthetic JSON records and route contracts, Windows PowerShell 5.1 built-ins for the tested router and verifiers, static HTML/CSS/JavaScript for the visual workflow, SHA-256 manifests for source integrity, and generated JSON and Markdown receipts for inspectable output.

The router rejects stale records, denied scopes, and route mismatches before selecting the newest eligible record for each required type. It binds only that smallest sufficient set and caches the completed mount in-process by normalized route, scope, and prompt.

The portable reference implementation was built and tested in Codex using GPT-5.6 Sol at max effort. The durable PITHOS file tree is the continuity architecture; models are replaceable execution layers.

Challenges

The main challenges were separating pre-existing PITHOS from qualifying new work, producing a runnable demo without exposing the private archive, protecting a resource-constrained laptop from new toolchain load, and proving bounded retrieval through behavior and receipts rather than a conceptual slide deck.

Accomplishments

The Build Week extension converts continuity contracts into a runnable zero-install reference implementation with smallest-sufficient retrieval, explicit exclusions, provenance-bound answers, duplicate-request reuse, deterministic verification, and a sanitized public package.

What I learned

Continuity is safer when it is an inspectable routing and permission function rather than a raw memory dump. Eligibility evidence is stronger when timestamps, manifests, limitations, and tests are designed into the build.

What's next

Test the package on a second physical host before claiming cross-host continuity, then extend the same bounded retrieval contracts into richer low-hand-use and embodied interfaces.

PITHOS is a pre-existing project. Judges are asked to evaluate only the post-cutoff Workbench Continuity Layer extension described above.

Built With

• gpt-5.6