11 Commits

Author SHA1 Message Date
temeddix dcacac18e8 Make bot review setup self-contained (#11)
Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-14 08:49:49 +00:00
temeddix 70d71aa4fd Inline comments (#10)
A finding about one line now lands on that line of the diff: the agent writes anchors as JSON, and the review posts them with the body in one call. Bad entries fail the run, and an anchor Gitea refuses falls back to posting the body alone.

Verified with deno fmt, lint and check.

Reviewed-on: #10
Co-authored-by: bot <temeddix@gmail.com>
Co-committed-by: bot <temeddix@gmail.com>
2026-09-14 00:11:35 +00:00
temeddix 2b51793c92 Review on request (#9)
A pull request comment that asks for a review now posts a real review, not a plain comment: the review file decides, the event no longer does.

Verified with deno fmt, lint and check.

Reviewed-on: #9
Co-authored-by: bot <temeddix@gmail.com>
Co-committed-by: bot <temeddix@gmail.com>
2026-09-13 23:48:04 +00:00
temeddix 4ed8b48b7a Verdict words (#8)
The review file starts with `Approved` or `Changes requested` instead of `Yes`/`No`/`With fixes`. The match is still the whole trimmed first line; `run.ts` prepends , 🛑, or 💬 when posting, so the mark never takes part in the match.

Verified: `deno fmt`, `deno lint`, `deno check` pass; a scratch run of the matcher maps `Approved` → APPROVED, `Changes requested` → REQUEST_CHANGES, and `Approved, mostly`, ` Approved`, `Yes` → COMMENT.
Reviewed-on: #8
Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-13 17:21:34 +00:00
temeddix 0d109ebec8 Review file (#7)
The reviewer writes its review to a file whose first line is the verdict, instead of relying on a narration-free final message. Sonnet put `Yes` after three paragraphs of narration on memona #936 (review 595), which the fail-closed verdict posted as a plain comment.

Reviewed-on: #7
Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-13 16:54:54 +00:00
temeddix a2bc4c9541 Model input (#6)
Optional `model` input. Defaults move to the mid tiers, `claude-sonnet-5` and `gpt-5.6-terra`, since a run follows a fixed template plus the project's checks; a workflow can still pass a bigger model.

Reviewed-on: #6
Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-13 16:48:13 +00:00
temeddix 75d3593376 Author replies (#5)
The reviewer drops a finding the author has answered in the PR comments as intended or a false alarm, once the code or docs make that clear. Pairs with memona's merge-branch gate loop.

Reviewed-on: #5
Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-13 15:35:17 +00:00
temeddix 439f2b4e77 Checked-out head (#4)
Checked-out head (#4)

Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-13 15:06:27 +00:00
temeddix f0506adfea Fail-closed verdict (#3)
Fail-closed verdict (#3)

Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-13 14:57:01 +00:00
temeddix b8d0093ef2 Superpowers review (#2)
Superpowers review (#2)

Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-13 14:30:29 +00:00
temeddix 8c2041bbc5 Split tokens (#1)
The agent and the bot are now two accounts, and the runner is one Deno script.

- `author-token` (was `gitea-token`): commits, pushes, and opens pull requests; the agent sees it as `GITEA_TOKEN`, and commits use that account's login and email.
- `reviewer-token`: posts comments and reviews as the bot; withheld from the agent's environment so it can never approve as the bot.
- A `pull_request` run posts a review instead of a comment: `REQUEST_CHANGES` when the response mentions `@bot`, `COMMENT` when the run failed, `APPROVED` otherwise. Gitea refuses self-approval, so the two accounts must differ.
- The reviewer checks the whole repository against `commons/code-rules`, which is fetched and embedded in the prompt, and requests changes for violations even when the diff did not cause them.
- `run.ts` replaces the three shell scripts plus `jq`, `envsubst`, and `ansifilter`; only `deno` is added to the install step, per the rules' Deno-over-Node policy. A `.gitea` workflow runs `deno fmt`, `lint`, and `check`.

Callers must rename `gitea-token` and add `reviewer-token` (`write:issue` and `write:repository` scopes). A rejection stays until the bot reviews again, so callers that want it lifted after a fix should trigger on `pull_request: [opened, synchronize]`.

Verified with a fake `claude` binary against this PR in an isolated `HOME`: git author configured from the token, prompt rendered with rules and comment history, events streamed, reviewer token absent from the agent's environment, review posted (then deleted).

Reviewed-on: #1
Co-authored-by: Danny Kim <temeddix@gmail.com>
Co-committed-by: Danny Kim <temeddix@gmail.com>
2026-09-13 14:15:23 +00:00
5 changed files with 261 additions and 58 deletions
+11
View File
@@ -20,10 +20,20 @@ inputs:
bot-token:
description: API key or token for the selected bot
required: false
model:
description: >-
Model for the selected bot. Defaults to `claude-sonnet-5` or
`gpt-5.6-terra`, the mid tiers, which cover reviews and fixes.
required: false
runs:
using: composite
steps:
# The agent works on the event's commit with full history, as the author.
- uses: actions/checkout@v4
with:
fetch-depth: 0
token: ${{ inputs.author-token }}
# This step assumes this is `node:24-bookworm` container.
- name: Install dependencies
shell: bash
@@ -35,6 +45,7 @@ runs:
ACTION_PATH: ${{ gitea.action_path }}
BOT_TYPE: ${{ inputs.bot-type }}
BOT_TOKEN: ${{ inputs.bot-token }}
MODEL: ${{ inputs.model }}
GITEA_API_URL: ${{ gitea.api_url }}
GITEA_REPOSITORY: ${{ gitea.repository }}
GITEA_TOKEN: ${{ inputs.author-token }}
+12
View File
@@ -0,0 +1,12 @@
export async function retryInvalidToken<
T extends { status: { success: boolean }; error: string },
>(attempt: () => Promise<T>, relogin: () => Promise<void>): Promise<T> {
let result = await attempt();
if (
!result.status.success && result.error.includes("invalid_refresh_token")
) {
await relogin();
result = await attempt();
}
return result;
}
+44
View File
@@ -0,0 +1,44 @@
import { assertEquals } from "jsr:@std/assert@1";
import { retryInvalidToken } from "./auth.ts";
Deno.test("retries once after an invalid refresh token", async () => {
let attempts = 0;
let relogins = 0;
const result = await retryInvalidToken(
() =>
Promise.resolve({
status: { success: false },
error: attempts++ === 0 ? "invalid_refresh_token" : "still failed",
}),
() => {
relogins++;
return Promise.resolve();
},
);
assertEquals({ attempts, relogins, error: result.error }, {
attempts: 2,
relogins: 1,
error: "still failed",
});
});
Deno.test("does not retry an unrelated failure", async () => {
let attempts = 0;
let relogins = 0;
await retryInvalidToken(
() => {
attempts++;
return Promise.resolve({
status: { success: false },
error: "rate limited",
});
},
() => {
relogins++;
return Promise.resolve();
},
);
assertEquals({ attempts, relogins }, { attempts: 1, relogins: 0 });
});
+43 -9
View File
@@ -13,15 +13,49 @@ response ends, so background monitors, scheduled wake-ups, and queued tasks
never resume. Never promise future action and never claim to be waiting on a
notification.
For a `pull_request` event, review the PR without changing code. Your final
response is posted as a pull request review from the bot account: it requests
changes when it mentions `@bot` and approves otherwise. So include `@bot` with
instructions to fix the findings exactly when changes are needed, and never
mention `@bot` when the PR is ready. For UI changes, check that the result is
aligned, clean, and pixel-perfect, and that included screenshots prove the
intended result was achieved. Also check the whole repository, not only the
diff, against the code rules at the end of this prompt, and request changes for
every violation you find even when the diff did not cause it.
The repository is checked out in the working directory at the event's commit,
the head of the pull request when there is one, with full history and the
author's push credentials. Read the code there, run its checks and tests when
they bear on the task, and push from there.
For a `pull_request` event, and for a comment on a pull request that asks you to
review it, review the PR without changing code against its base and head. Run
the project's checks on the head and treat a failure as at least Important.
Check the whole repository against the code rules at the end of this prompt, not
only the diff; a violation is at least Important even when the diff did not
cause it. Write the complete review, and nothing else, to the file
`${REVIEW_PATH}`: it is posted verbatim as a pull request review from the bot
account, and your final response is not posted at all. The file's first line
must be exactly the verdict and nothing else: `Approved` when the head is ready
to merge, `Changes requested` otherwise. The mark in front of it is added when
posting, so write the words alone; any other first line is posted as a plain
comment, which wastes the run. Minor issues alone never block, and neither does
a finding the author has answered in the comment history below as intended or a
false alarm, once the code or docs make that clear. When the verdict is
`Changes requested`, the second line names what must change in one line,
addressed to the author; the author's own agent picks the fixes up, so never ask
`@bot` to make them. For UI changes, check that the result is aligned, clean,
and pixel-perfect, and that included screenshots prove the intended result was
achieved.
Every finding that belongs to one line of the diff goes on that line instead of
into the body. Write those to `${ANCHORS_PATH}` as a JSON array, each entry
`{"path": "<path from the repository root>", "line": <number>, "side": "new" |
"old", "body": "<the finding>"}`.
`side` is `new` for a line in the head file and `old` for one only in the base
file; `line` is that file's own line number, and it must be a line the diff
touches, or Gitea refuses the anchor. Write the file only when there is
something to anchor, and keep each body to the what, the why, and the how, with
no `file:line` prefix; the line carries that.
The review body must read at a glance: everything outside `<details>` blocks
totals under 512 bytes. Only core information stays visible: the verdict, the
summary line, and the section headings. Anything verbose goes into a `<details>`
block whose `<summary>` is a few words, such as the title of an issue with the
what, why, and how inside; the same for each strength, each recommendation, the
reasoning, and any compliance notes. A finding you anchored belongs there only
as its title, since its detail is on the line. Details blocks are top-level,
never inside a list item, because Gitea breaks them there.
For an `issue_comment` or `pull_request_review_comment` event, treat the `body`
in the triggering comment payload below as the user's exact instruction.
+151 -49
View File
@@ -3,6 +3,7 @@
// GITEA_TOKEN; everything this script posts goes through the reviewer token,
// so it appears as the bot account.
import { TextLineStream } from "jsr:@std/streams@1/text-line-stream";
import { retryInvalidToken } from "./auth.ts";
type GiteaUser = { login: string; email: string };
type GiteaComment = { user: GiteaUser; created_at: string; body: string };
@@ -20,6 +21,13 @@ const EVENT = env("EVENT_NAME");
const AUTHOR_TOKEN = env("GITEA_TOKEN");
const REVIEWER_TOKEN = env("REVIEWER_TOKEN");
const RULES_PATH = "repos/commons/code-rules/raw/README.md";
// The mid tiers: a run follows a fixed template and the project's checks.
const DEFAULT_MODELS: Record<string, string> = {
claude: "claude-sonnet-5",
codex: "gpt-5.6-terra",
};
const model = (bot: string): string =>
Deno.env.get("MODEL") || DEFAULT_MODELS[bot];
// The reviewer token is withheld so the agent cannot approve as the bot.
const { REVIEWER_TOKEN: _, ...agentEnv } = Deno.env.toObject();
@@ -52,29 +60,97 @@ async function postComment(body: string): Promise<void> {
});
}
// A pull request event is a review request, so the response becomes a review:
// changes are requested when the agent asked @bot to fix something, a failed
// run only comments, and anything else approves.
async function postResult(body: string): Promise<void> {
if (EVENT !== "pull_request") return postComment(body);
const event = body.includes("@bot")
? "REQUEST_CHANGES"
: body.startsWith("Bot failed:")
? "COMMENT"
: "APPROVED";
await gitea(REVIEWER_TOKEN, `repos/${REPO}/pulls/${INDEX}/reviews`, {
body: stripAnsi(body),
event,
// A review is posted whenever the agent wrote one, whether a review request or
// a comment asked for it. It comes through a file, because a final chat message
// picks up narration while a file's first line is written on purpose. That line
// is the verdict, matched whole; anything unexpected only comments, never
// approves. The mark in front is added here, so it is never part of the match.
const REVIEW_DIR = await Deno.makeTempDir();
const REVIEW_PATH = `${REVIEW_DIR}/review.md`;
const VERDICTS: Record<string, [event: string, mark: string]> = {
Approved: ["APPROVED", "✅"],
"Changes requested": ["REQUEST_CHANGES", "🛑"],
};
// A finding about one line is posted on that line of the diff rather than as
// `file:line` prose in the body. Those anchors come as JSON, so the file and
// line are structured instead of parsed back out of English; a malformed entry
// fails the run, because a silently dropped finding is worse than a red run.
const ANCHORS_PATH = `${REVIEW_DIR}/anchors.json`;
type Anchor = { path: string; line: number; side: "new" | "old"; body: string };
function parseAnchors(text: string): Anchor[] {
const entries: unknown = JSON.parse(text);
if (!Array.isArray(entries)) throw new Error(`${ANCHORS_PATH}: not an array`);
return entries.map((entry: unknown, index) => {
const at = `${ANCHORS_PATH}[${index}]`;
if (typeof entry !== "object" || entry === null) {
throw new Error(`${at}: not an object`);
}
const { path, line, side = "new", body } = entry as Record<string, unknown>;
if (typeof path !== "string" || path === "") {
throw new Error(`${at}.path: expected a repository path`);
}
if (typeof line !== "number" || !Number.isInteger(line) || line < 1) {
throw new Error(`${at}.line: expected a line number`);
}
if (side !== "new" && side !== "old") {
throw new Error(`${at}.side: expected "new" or "old"`);
}
if (typeof body !== "string" || body.trim() === "") {
throw new Error(`${at}.body: expected the finding`);
}
return { path, line, side, body };
});
}
async function readAnchors(): Promise<Anchor[]> {
const written = await Deno.readTextFile(ANCHORS_PATH).catch(() => null);
return written === null ? [] : parseAnchors(written);
}
async function postResult(body: string): Promise<void> {
const review = await Deno.readTextFile(REVIEW_PATH).catch(() => null);
if (review === null) {
if (EVENT === "pull_request") {
throw new Error(`no review was written to ${REVIEW_PATH}`);
}
return postComment(body);
}
const [verdict, ...rest] = review.split("\n");
const [event, mark] = VERDICTS[verdict.trim()] ?? ["COMMENT", "💬"];
const post = (anchors: Anchor[]) =>
gitea(REVIEWER_TOKEN, `repos/${REPO}/pulls/${INDEX}/reviews`, {
body: stripAnsi([`${mark} ${verdict.trim()}`, ...rest].join("\n")),
event,
comments: anchors.map(({ path, line, side, body }) => ({
path,
body: stripAnsi(body),
new_position: side === "new" ? line : 0,
old_position: side === "old" ? line : 0,
})),
});
const anchors = await readAnchors();
// Gitea rejects the whole review when an anchor names a line outside the
// diff, and a verdict that never lands blocks the pull request, so the body
// goes up alone rather than not at all.
await post(anchors).catch(async (error: Error) => {
if (anchors.length === 0) throw error;
console.error(`inline comments rejected: ${error.message}`);
await post([]);
});
}
async function run(command: string, args: string[]): Promise<void> {
const { success, code } = await new Deno.Command(command, { args }).output();
if (!success) throw new Error(`${command} ${args[0]} exited with ${code}`);
}
// Commits belong to the same account as the pull request they end up in.
async function configureGitAuthor(): Promise<void> {
const user: GiteaUser = await (await gitea(AUTHOR_TOKEN, "user")).json();
for (const [key, value] of [["name", user.login], ["email", user.email]]) {
await new Deno.Command("git", {
args: ["config", "--global", `user.${key}`, value],
}).output();
await run("git", ["config", "--global", `user.${key}`, value]);
}
}
@@ -93,6 +169,8 @@ async function renderPrompt(): Promise<string> {
GITEA_API_URL: API,
GITEA_REPOSITORY: REPO,
ISSUE_INDEX: INDEX,
REVIEW_PATH,
ANCHORS_PATH,
};
const template = await Deno.readTextFile(
new URL("prompt.md", import.meta.url),
@@ -106,17 +184,16 @@ async function renderPrompt(): Promise<string> {
async function runClaude(prompt: string): Promise<string> {
const token = Deno.env.get("BOT_TOKEN");
if (!token) {
await postComment(
throw new Error(
"Run `claude setup-token` locally and set the `bot-token` action input.",
);
Deno.exit(1);
}
const claude = new Deno.Command("claude", {
args: [
"--print",
"--dangerously-skip-permissions",
"--model",
"claude-fable-5",
model("claude"),
"--output-format",
"stream-json",
"--verbose",
@@ -127,7 +204,7 @@ async function runClaude(prompt: string): Promise<string> {
clearEnv: true,
stdout: "piped",
}).spawn();
let result = "Bot failed: no result";
let result: { result?: string; subtype: string } | undefined;
// Print events as they stream so the runner does not kill the job as a zombie.
const lines = claude.stdout
.pipeThrough(new TextDecoderStream())
@@ -139,12 +216,13 @@ async function runClaude(prompt: string): Promise<string> {
const text = part.thinking ?? part.text ?? part.name;
if (text) console.log(text);
}
if (event.type === "result") {
result = event.result ?? `Bot failed: ${event.subtype}`;
}
if (event.type === "result") result = event;
}
await claude.status;
return result;
if (result?.result === undefined) {
throw new Error(`claude ended with ${result?.subtype ?? "no result"}`);
}
return result.result;
}
// Posts the device code so a human can finish the login on the persisted home.
@@ -160,12 +238,17 @@ async function codexDeviceLogin(): Promise<void> {
new WritableStream({ write: (chunk) => void (shown += chunk) }),
);
const drained = Promise.all([collect(login.stdout), collect(login.stderr)]);
const status = login.status;
await new Promise((resolve) => setTimeout(resolve, 3000));
while (shown.trim() === "") {
const exited = await Promise.race([
status.then(() => true),
new Promise<boolean>((resolve) => setTimeout(() => resolve(false), 100)),
]);
if (exited) throw new Error("AI bot login produced no instructions");
}
await postComment(shown);
await Promise.all([login.status, drained]);
const status = await new Deno.Command("codex", { args: ["login", "status"] })
.output();
await postComment(new TextDecoder().decode(status.stdout));
await Promise.all([status, drained]);
}
async function runCodex(prompt: string): Promise<string> {
@@ -175,28 +258,47 @@ async function runCodex(prompt: string): Promise<string> {
.output();
if (!loggedIn.success) await codexDeviceLogin();
const file = await Deno.makeTempFile();
const status = await new Deno.Command("codex", {
args: [
"exec",
"--model",
"gpt-5.5",
"--dangerously-bypass-approvals-and-sandbox",
"--output-last-message",
file,
prompt,
],
env: agentEnv,
clearEnv: true,
stdout: "inherit",
stderr: "inherit",
}).spawn().status;
const attempt = async () => {
const codex = new Deno.Command("codex", {
args: [
"exec",
"--model",
model("codex"),
"--dangerously-bypass-approvals-and-sandbox",
"--output-last-message",
file,
prompt,
],
env: agentEnv,
clearEnv: true,
stdout: "inherit",
stderr: "piped",
}).spawn();
const decoder = new TextDecoder();
let error = "";
for await (const chunk of codex.stderr) {
await Deno.stderr.write(chunk);
error += decoder.decode(chunk, { stream: true });
}
error += decoder.decode();
return { status: await codex.status, error };
};
const { status } = await retryInvalidToken(attempt, async () => {
await run("codex", ["logout"]);
await codexDeviceLogin();
});
if (!status.success) throw new Error(`codex exited with ${status.code}`);
return await Deno.readTextFile(file);
}
await configureGitAuthor();
const prompt = await renderPrompt();
const result = env("BOT_TYPE") === "claude"
? await runClaude(prompt)
: await runCodex(prompt);
await postResult(result);
try {
await configureGitAuthor();
const prompt = await renderPrompt();
const result = env("BOT_TYPE") === "claude"
? await runClaude(prompt)
: await runCodex(prompt);
await postResult(result);
} catch (error) {
await postComment(`Bot failed: ${error}`);
throw error;
}