RIPP

PROJECT 001 / EXPERIMENTAL LOCAL ALPHA

RIPP

Recursive Improvement Planning & Processing

A runtime for the work around the code.

RIPP helps a technical agent keep a long task bounded, evidence-led, and recoverable. It is built for Codex first. A Codex plugin is the intended product path; today’s installable experience is a local alpha sidecar.

THE USE CASE

One repository task. Many chances to drift.

Ask Codex to trace a bug across a repository, repair it, run the right checks, and leave a result another person can review. As the task grows, the agent must keep the finish condition clear and avoid reopening settled questions.

RIPP starts with a Run Charter: observable endpoint, acceptance checks, allowed scope, resource envelope, validation reserve, and authority boundary. It helps choose the next useful step and preserves the evidence behind important decisions.

HOW IT WORKS

Different work gets different machinery.

System 1 and System 2 are engineering roles in RIPP, not claims that software models think like people. The role depends on the operation, not the model brand.

SYSTEM 2 / EXPAND

Generate and do

Codex proposes candidate explanations, explores plans, writes code, and handles open-ended recovery.

SYSTEM 1 / JUDGE

Compare supplied choices

Jev is the first adapter for scoring stated options or evidence. It does not make tests pass or decide user permissions.

TOOLS / VERIFY

Establish facts

Read files, run tests, validate outputs, and report observed results. Facts come from the tool surface.

For example: Codex proposes three likely causes of a failing test; Jev ranks those supplied candidates; a file inspection or test checks the leading explanation.

THE CONTROL LOOP

A run you can inspect and resume.

Charter → Expand → Scope/admit → Discriminate → Act → Observe → Compress → Commit → Evaluate

A deterministic scope gate checks the expenditure case before routing or child work. Decision Records keep evidence addresses and reopen conditions. The runtime can stop, replan, or bypass its own orchestration when the overhead is unlikely to pay back.

EVIDENCE / MEASURED ONCE

Promising signal. Narrow claim.

RAW CODEX

50.70s

3 / 3 frozen tests passed

PROVISIONAL RIPP

38.71s

3 / 3 frozen tests passed

One paired run on the frozen alpha slug fixture showed about 23.7% less wall time in the provisional RIPP arm. Both arms used fresh equivalent fixture copies and the same acceptance checks.

The Jev checkpoint was prompt-injected because nested MCP registration was unavailable. Codex API cost was not reported. The result does not establish a general speed or cost advantage. Matched repetitions with complete planning, inference, tool, storage, retry, and validation overhead are still needed.

A later campaign was prepared but did not run because credential and smoke gates remained unresolved.

GET STARTED / LOCAL ALPHA

Try the current sidecar.

The Codex plugin package is still the product direction, so there is no plugin install command yet. From a local checkout of the repository, install the alpha package with:

uv tool install --editable .
  • On Windows, use ripp-credential store for the credential-provider flow.
  • On Linux desktop, use Secret Service. Headless Linux and CI need explicit process-scoped environment injection.
  • Follow the maintained repository guide for host-specific MCP setup. Keep credentials out of project files and prompts.
  • For a first run, choose one bounded coding task, define what passing means, set scope and budget, then inspect tests, resource records, and Decision Records.

Repository access may be limited while RIPP is in alpha.

CURRENT LIMITS

Built foundations. Open questions.

The alpha contains provider-neutral System 1 tooling, Run Charters, resource ledgers, Decision Records, deterministic routing and scope admission, and checkpointed recovery. Generalized scheduling, automatic context eviction, and savings on every task are not alpha guarantees.

Follow the work: GitHub · X / @danegerbravo