CodeAFSpecialists

Specialists

Build with /senior-dev, audit code with /sec and review GitHub pull requests with /review. Purpose-built specialists, built into CodeAF.

Build, audit and review. Built-in specialists, each with an engine for its own job.
  • /senior-dev implement the fix with tests
  • /sec changes since main
  • /review owner/repo#123

CodeAF carries purpose-built specialist harnesses. senior-dev implements a coding change. sec audits local code for security problems. review reads a GitHub pull request and reports findings. These specialists are built in on macOS and Linux, with no extra specialist installation.

Choose the specialist

CommandGive itGet backDefault ceiling
/senior-dev <brief>A settled coding change and what done meansA submitted change, discovered build/test results and a task branch in git$10 / 3 hours
/sec or /sec changes since mainA local folder, or eligible changed codeSecurity findings, evidence and suggested fixes; Markdown, JSON and SARIF reports$5 / 4 hours
/review owner/repo#123A GitHub pull request and optional focusFindings, blocking status and coverage; Markdown and JSON reports$5 / 2 hours

Conversation limits can lower those ceilings. Results depend on the scope, models and limits: inspect what ran and what remained incomplete. No specialist guarantees that every bug or vulnerability will be found.

Original projects and benchmark evidence

The original PR-AF and SEC-AF projects are imported into CodeAF as /review and /sec. Use the CodeAF commands above; the standalone projects are not extra services you must install.

PR-AF publishes Martian Code-Review-Bench results, including its models and evaluation scope. That evidence belongs to the original project's runs, rather than a fresh measurement of CodeAF's imported /review. The DeepSWE evidence below measures senior-dev. No security benchmark result is claimed here.

Try it in CodeAF

Where

In a conversation on your repository; review can also take an explicit GitHub pull-request URL.

Type

Choose /senior-dev with a coding brief, /sec for an audit, or /review with a pull request. Press Enter to start it at once.

You see

A task wearing the specialist's name. Open it to follow its stages and spending or to stop it. Its ending returns to the chat.

How a specialist fits into CodeAF

The chat can propose a specialist in plain words too. The coding brief can carry the context you worked out together; sec's brief names scope and depth, while review's names the pull request and focus. Read the proposal card for its brief and limits before it starts.

Every model call goes through CodeAF, counts toward your spending and is held to the run's ceiling. Each specialist has its own process and output. senior-dev can receive steering until it submits; sec and review cannot read messages mid-run. Stop and restart those with a new scope or focus.

Security and review use read-only file tools and run no tests, builds, apps or exploits. They leave your code unchanged. senior-dev builds the change and checks the frozen tree; its ending says when no build or tests could be found. Posting a review, fixing findings and merging a branch are separate actions.

Inside the coding specialist

The following figures and demo show senior-dev, the coding example. sec's recon, hunt and prove phases and review's planning, challenge and coverage stages are described on their own pages.

a generalista checked resulta specialistintakea pinned checka freezeindependent verificationa checked resulta generalista checked resulta specialistintakea pinned checka freezeindependent verificationa checked result
Guided autonomy: the same field of possible paths, searched two ways.

The coding harness wraps one model session in stages that retain the brief, pin a check, implement, submit and verify. The model makes coding decisions; the harness runs checks on what was submitted.

On the DeepSWE study, senior-dev solved the most issues of ten harnesses on the same model, at the lowest cost per solved task. Those results describe the coding specialist and that benchmark; they do not measure sec or review. See the coding workflow and benchmark.

Things to build

  • A coding change: settle the brief in the chat, then hand it to senior-dev. Hand a job to a specialist
  • An audit followed by a fix: run /sec changes since main, inspect the findings and coverage, then ask for selected findings to be fixed with senior-dev.
  • A focused pull-request review: push and open the pull request, then run /review owner/repo#123 focus on the retry logic. Inspect the report before asking to post it.
  • A local test gate: a team member reads the branch and runs tests after senior-dev builds it. This reviewer member is a normal chat worker, distinct from the GitHub-only /review specialist. Review every change

Save a way of working

For work you repeat, ask CodeAF to design a saved harness, such as "make a harness for triaging flaky tests". It shows a design card for you to review and approve. Open /subharness to find and run the saved program; its intake card asks for the input it needs before you start it. Your saved programs are separate from the built-in senior-dev, sec and review specialists.

Go further

Coming soon: describe the org you want and CodeAF builds it: teams, managers, budgets, standing orders.