akapen 0.4.0 → 0.5.1

package/README.md

@@ -1,32 +1,28 @@
 # akapen
 
-A toolkit for deep, multi-stage code review. A change is reviewed by a different model than the one that wrote it, in stages: context research, reviewers with different lenses (logic, boundaries, regressions, paths, security, and a generalist), a critic that looks for what they missed, and a verifier that checks every finding against the code before it reaches you. You get findings to fix, with the evidence behind each, or LGTM.
+Deep, multi-stage code review by a different model than the one that wrote the change: context research, reviewers with different lenses, a critic, and a verifier that checks every finding against the code. You get findings to fix, each with its evidence, or LGTM.
 
 ## Reviewers
 
-| `--reviewer` | Runs on                                                                                                             | Needs                                    |
-| ------------ | ------------------------------------------------------------------------------------------------------------------- | ---------------------------------------- |
-| `codex`      | Codex, through your ChatGPT subscription                                                                            | the `codex` CLI, signed in               |
-| `claude`     | Claude Code, through your Claude subscription                                                                       | Claude Code, signed in                   |
-| `api`        | akapen's own smart agent harness, purpose-built for the best possible code review, on the OpenAI and Anthropic APIs | `OPENAI_API_KEY` and `ANTHROPIC_API_KEY` |
+| `--reviewer` | Runs on                                                      | Needs                                    |
+| ------------ | ------------------------------------------------------------ | ---------------------------------------- |
+| `codex`      | Codex, through your ChatGPT subscription                     | the `codex` CLI, signed in               |
+| `claude`     | Claude Code, through your Claude subscription                | Claude Code, signed in                   |
+| `api`        | akapen's own agent harness, on the OpenAI and Anthropic APIs | `OPENAI_API_KEY` and `ANTHROPIC_API_KEY` |
 
-All three need Node.js 24 or later. With `api`, a review of a few hundred changed lines costs about $3–9 and a re-review after fixes about $1–2. Each result reports what it cost.
-
-Reviewers read the repository and can run small scripts inside a sandbox. They never change your code. The sandbox has been verified on macOS only; on Linux, Claude's sandbox needs `bubblewrap` and `socat`, and the `api` reviewer's `bash` needs what `codex sandbox` needs (without it, the review goes on without `bash`).
+All need Node.js 24 or later. With `api`, a review of a few hundred changed lines costs about $3–9, a re-review after fixes about $1–2. Reviewers never change your code; they can run small scripts in a sandbox (verified on macOS).
 
 ## Use it from your coding agent
 
-```bash
-npx akapen setup
-```
+In your repository, ask your coding agent (Claude Code or Codex):
 
-`npx` reads the `.npmrc` of the directory it runs in, so run this in your own repository (to install the skill there) or in an empty directory (to install it for your user), not in a checkout of someone else's branch.
+> Run `npx akapen setup project` and set up akapen by following it.
 
-This installs the review skill for Claude Code and/or Codex. Then ask your agent to run akapen (or "赤ペン入れて"). It runs the review, checks each finding itself, fixes what holds up, and re-reviews until LGTM, five re-reviews at most unless you say otherwise.
+Then ask it to run akapen (or "赤ペン入れて"). It reviews, checks each finding itself, fixes what holds up, and re-reviews until LGTM.
 
 ## Run it yourself
 
-Run it from an empty directory and point it at the repository with `-C`. Run inside the repository, `npx` would read the reviewed branch's `.npmrc` (`script-shell`, `registry`), which runs that branch's code before any sandbox.
+Run it from an empty directory and point it at the repository with `-C`, so `npx` does not read the reviewed branch's `.npmrc`:
 
 ```bash
 R="$PWD"; OUT="$R/.git/akapen/last.json"; mkdir -p "$(dirname "$OUT")"
@@ -34,10 +30,4 @@ R="$PWD"; OUT="$R/.git/akapen/last.json"; mkdir -p "$(dirname "$OUT")"
 (cd "$(mktemp -d)" && npx akapen -C "$R" --reviewer api --previous "$OUT" --out "$OUT")   # a re-review after fixes
 ```
 
-`(cd "$(mktemp -d)" && npx akapen --help)` lists the scopes and options.
-
-## Repository settings
-
-A repository can add `.akapen/knowledge.md` (rules and terms reviewers should know), `.akapen/config.json` (models, limits, docs MCP servers) and `.akapen/kinds.json` (file kinds such as translations and generated files). They are read from the base branch, so a pull request cannot change how it is reviewed.
-
-Settings in `.akapen/config.json` that widen what a review may read, run or send (the sandbox, MCP servers, web search for stages) apply only once you trust that exact file: `akapen trust --show`, then `akapen trust`.
+`npx akapen --help` lists the options, and `npx akapen validate -C "$R"` checks the settings before a review.