kres

Kernel code RESearch agent — an LLM-driven multi-agent REPL for reviewing, auditing, and finding bugs in C source trees. The Linux kernel is the primary target; any large C codebase with source-level tooling works too.

kres introduction

kres splits the job of reviewing code across a number of cooperating agents:

• fast scopes the work, picks the code to look at, and emits a structured brief for deeper analysis.
• main fetches that code via MCP tools, grep, read, git — treating code navigation as a first-class tool-call surface rather than free-form text manipulation.
• slow runs the deep analysis with a prepared context and previous findings in hand, so the expensive model's tokens go to bug-hunting rather than chasing files.
• todo dedups follow-up questions, reprioritises, and keeps a running list across turns so a single prompt can drive 30+ tasks without losing coverage.
• merger folds each task's findings into a cumulative, deduplicated bug list; old findings get invalidated when a later one supersedes them.

The results of every turn are used to reprioritize the todo list, and identify additional context needed for the next round.

See docs/agents.md for the task flow and docs/review-template.md for the parallel-lens review.

Quick start

1. Build:

   cargo build --release
   

2. Populate ~/.kres/ from shipped configs:

   ./setup.sh --fast-key $FAST_API_KEY --slow-key $SLOW_API_KEY
   

   Each key arg accepts a literal API key or a path to a key file. setup.sh --help lists every option — model picks (--slow, --model), --semcode PATH, --review-prompts PATH, --overwrite, and more. The shipped defaults use claude-opus-4-7 for the slow agent and claude-sonnet-4-6 for the fast / main / todo roles; ~/.kres/settings.json is the single source of truth for model selection.

3. Run a review from a kernel tree:

   cd linux
   kres --results review --prompt 'review: fs/btrfs/ctree.c' --turns 2
   kres --summary-markdown --results review
   # review/summary.md now has your results
   

   --prompt 'review: X' invokes the embedded review template — a five-lens parallel audit over the target. --results DIR keeps the run's artifacts under DIR/ (findings.json, report.md, summary.txt). --turns 2 stops after two completed tasks; see docs/turns-and-follow.md for the other stop modes.

Two optional integrations are worth wiring up while you're here: semcode-mcp for whole-program code navigation and the kernel review-prompts repo for subsystem knowledge. Both are configured via setup.sh flags — see docs/configuration.md for details.

Exporting findings

You can export results into either text or markdown:

• docs/summary.md — /summary, kres --summary, and the summary output format.

But these scans can produce a lot of results, and churning through a giant text file isn't the easiest way to walk them. You can also dump them into a one-dir-per-finding format:

• docs/exporting.md — kres --export DIR:

Further reading

• docs/agents.md — fast / main / slow / todo / merger flow and how follow-up tasks drive larger reviews.
• docs/review-template.md — the parallel-lens review flow behind --prompt "review:".
• docs/coding-tasks.md — reproducer and fix generation (code_output, code_edits, bash verify).
• docs/turns-and-follow.md — when kres decides a non-interactive run is done.
• docs/action-allowlist.md — which non-MCP tools the main agent can dispatch and how to change that.
• docs/configuration.md — ~/.kres/ layout, model selection, system-prompt overrides, semcode MCP integration, and kernel review-prompts setup.
• docs/commands.md — slash-command templates (/review, /summary, operator-authored additions).
• docs/cli.md — every CLI flag and REPL command.
• docs/development.md — workspace layout, build / test / lint, pre-commit hook.