Select the task mode
Implementation context, diff review, architecture mapping, or focused test selection.
In three paired end-to-end runs, Speeder reduced average input tokens by 34% and completion time by 19%, while both variants passed every quality check. It works by starting Codex with the relevant files, callers, risks, and tests.
› Review the rejected RPC handling
↳ Codex Speeder: narrowed this task to 3 relevant files and 1 focused test
Speeder does not replace Codex. It prepares a bounded, evidence-backed starting point; Codex performs the implementation, review, or analysis.
Implementation context, diff review, architecture mapping, or focused test selection.
Resolve symbols, callers, configuration, tests, and risk boundaries without reading the repository broadly.
Codex receives the evidence packet, continues the normal task, and reuses unchanged results when possible.
Three paired, end-to-end Codex CLI runs on a 553-file repository. Each pair used the same prompt, source snapshot, task, permissions, and quality criteria.
153,323 → 100,599 average
50.8s → 41.1s average
20 → 14 across three runs
Equal to Codex alone
Cold isolated comparison using Codex CLI 0.146.1, GPT-5.6 Terra, medium reasoning. This is a scoped cross-file navigation result—not a universal speed or cost guarantee.
Read the full methodologyRepository analysis
Runs on your machineSpeeder reads the Git repository you point it at and writes a local content-hash cache. The analysis engine does not contact a network provider or send repository content to the publisher.
Best suited to unfamiliar paths, cross-file changes, code review, and architecture work in Python or TypeScript repositories. Trivial known-file edits bypass structural discovery.
Secure checkout by Stripe · GitHub delivery
It reduces broad repository discovery before implementation, review, architecture analysis, and test selection. Speeder traces the direct evidence for a task and gives that map to Codex as a starting point.
No. The plugin includes a local analysis engine and platform binaries. The skill routes Codex into the appropriate mode; the engine performs deterministic repository analysis and returns a compact evidence packet.
No. It does not rewrite or compress source code. It identifies relevant files, direct callers, configuration, risk boundaries, and focused tests so Codex can avoid unnecessary discovery.
Speeder runs locally in Codex CLI. Its repository analysis does not contact a network provider or send your source code to the publisher.
Python and TypeScript Git repositories on macOS, Linux, and Windows. TypeScript analysis requires Node.js; Python analysis requires ripgrep.
A trivial edit with an exact file and validation already known does not need structural discovery. Speeder is designed to reuse the known file list instead of adding another repository scan.
No. The published result is a scoped benchmark for one cross-file TypeScript navigation task. It demonstrates that Speeder can reduce discovery work under those conditions; it is not a universal speed, token, or cost guarantee.
Checkout collects your GitHub username. We invite that account to the private plugin repository, where installation and update instructions are available.
For tasks where repository navigation is part of the work