A local-first, agent-agnostic code-review skill for Asterinas, with a benchmark
that measures its recall. It reviews code against the persona-keyed Coding
Guidelines and writes one Markdown review file, making both objective (bug) and
subjective (guideline-grounded) calls.
Two review modes (scripts/resolve_target.sh), both anchored at HEAD:
- `diff <base>` — the commit series merge-base(base,HEAD)..HEAD: each commit's
message and diff, so the review covers per-commit intent and commit hygiene.
- `files <path[:lines] ...>` — the working-tree contents of the named files.
The raw argument string is self-tokenized (double quotes, line ranges) so the
parse is deterministic regardless of how the skill is triggered.
Execution model (SKILL.md, spec/execution_model.md): a thin orchestrator
activates the five personas by path, fans out one isolated pass per persona
(scripts/build_pass_prompt.sh assembles a cache-ordered prompt — stable contract
and guideline first, the review input last), then deterministically assembles
(scripts/assemble_review.sh), verifies, consolidates, and summarizes into the
review file. `--per-persona-context=auto|yes|no` trades the fan-out (best recall)
for a cheaper combined pass.
Benchmark (benchmark/): one queryable problems.yaml, each problem pinned to a
source commit and checked by validate_problem_yaml.sh; run.sh reconstructs each
problem in an isolated worktree (fetching the commit if absent), reviews
cheap-first with escalation to the fan-out on a miss, and grades recall — without
leaking the answers to the reviewer (overlay_skill.sh overlays the current skill
but excludes the suite, the worktree path is opaque, and the defects go only to
the grader).
Design rationale lives in spec/ (motivation, coding_guidelines, interface,
execution_model, benchmark, related_work), with every significant decision argued
in place. The deterministic scripts and the problem schema are covered by
model-free tests in tests/.
7a765ce649