forge CLI pivot: rewrite gitea+gitlab clients to tea/glab transports

Replace raw fetch transport with CLI subprocess calls:
- Gitea: spawn 'tea api --include' with GITEA_SERVER_TOKEN env var
- GitLab: spawn 'glab api --include' with GITLAB_TOKEN env var

Binary paths env-overridable (TEA_BIN / GLAB_BIN).
8s request timeout via AbortSignal on spawned process.
ETag cache and rate-limit cooldown dropped (tradeoff documented).
Pagination via --paginate for list endpoints.
Tests mock child_process.spawn instead of globalThis.fetch.
This commit is contained in:
2026-09-05 14:06:04 +00:00
parent aa0e3a222c
commit 77e7dd1127
6 changed files with 558 additions and 878 deletions
+115 -220
View File
@@ -1,197 +1,147 @@
import { spawn } from 'child_process';
import { getGiteaAuth } from './auth.js';
import { getProviderApiBaseUrl } from '../git-providers/config.js';
import { getEffectiveProviderApiBaseUrl } from '../git-providers/project-config.js';
// Per-request timeout for every Gitea call. Self-hosted instances can hang
// under load; bounding each request lets the caller fail fast and serve
// cached/last-known state instead of holding a socket open.
const TEA_BIN = process.env.TEA_BIN || '/home/user/.local/bin/tea';
const REQUEST_TIMEOUT_MS = 8000;
const timeoutFetch = (url, options = {}) => {
// Respect a caller-provided signal if present; otherwise attach our timeout.
if (options.signal) {
return fetch(url, options);
}
return fetch(url, { ...options, signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS) });
};
// Conditional-request cache for GET calls: Gitea serves 304 Not Modified for
// matching If-None-Match, so polling unchanged issues/PRs stays cheap. Keyed by
// token+URL so different identities never share responses.
const ETAG_CACHE_MAX_ENTRIES = 300;
const etagCache = new Map();
const rememberEtag = (key, etag, body, headers) => {
etagCache.delete(key);
etagCache.set(key, { etag, body, headers });
if (etagCache.size > ETAG_CACHE_MAX_ENTRIES) {
const oldest = etagCache.keys().next().value;
if (oldest !== undefined) {
etagCache.delete(oldest);
}
}
};
const createConditionalFetch = (token) => async (url, options = {}) => {
const method = (options.method || 'GET').toUpperCase();
if (method !== 'GET') {
return timeoutFetch(url, options);
}
const cacheKey = `${token}\n${url}`;
const cached = etagCache.get(cacheKey);
const headers = { ...(options.headers || {}) };
if (cached?.etag) {
headers['if-none-match'] = cached.etag;
}
const response = await timeoutFetch(url, { ...options, headers });
if (response.status === 304 && cached) {
// Touch for LRU and replay the cached success response.
rememberEtag(cacheKey, cached.etag, cached.body, cached.headers);
return new Response(cached.body, { status: 200, headers: cached.headers });
}
if (response.ok) {
const etag = response.headers.get('etag');
if (etag) {
const body = await response.arrayBuffer();
rememberEtag(cacheKey, etag, body, response.headers);
return new Response(body, { status: response.status, headers: response.headers });
}
}
return response;
};
// ---- Own rate-limit cooldown (deliberately NOT shared with github/gitlab) ----
const MAX_COOLDOWN_MS = 15 * 60 * 1000;
const DEFAULT_COOLDOWN_MS = 60 * 1000;
let rateLimitedUntil = 0;
const headerValue = (headers, name) => {
if (!headers) return undefined;
if (typeof headers.get === 'function') return headers.get(name);
return headers[name];
};
// NOTE: ETag conditional-GET cache and rate-limit cooldown have been dropped
// with the pivot to CLI transports. Each call spawns a fresh `tea` process,
// so there is no persistent connection to attach conditional headers to, and
// rate-limit state cannot be shared across invocations. The tradeoff is higher
// latency per call (process spawn overhead) and no 304 short-circuit, but
// simpler state management and no module-level mutable cache. Callers that
// relied on `isGiteaRateLimited()` will always see false (no cooldown active).
/**
* Record a cooldown after a Gitea 429. Accepts a fetch Response or any object
* carrying headers, honoring `Retry-After` (seconds) or `X-RateLimit-Reset`
* (Unix seconds) when present.
* Spawn a CLI binary and return { stdout, stderr, exitCode }.
* Rejects if the process does not finish within REQUEST_TIMEOUT_MS.
*/
export function noteGiteaRateLimit(error) {
const headers = error?.headers;
let retryMs = null;
const retryAfter = headerValue(headers, 'retry-after');
if (retryAfter !== undefined && retryAfter !== null) {
const secs = Number(retryAfter);
if (Number.isFinite(secs) && secs > 0) retryMs = secs * 1000;
function spawnCli(bin, args, env, timeoutMs = REQUEST_TIMEOUT_MS) {
return new Promise((resolve, reject) => {
const child = spawn(bin, args, {
env: { ...process.env, ...env },
stdio: ['ignore', 'pipe', 'pipe'],
});
let stdout = '';
let stderr = '';
child.stdout.on('data', (chunk) => { stdout += chunk; });
child.stderr.on('data', (chunk) => { stderr += chunk; });
const timer = setTimeout(() => {
child.kill('SIGKILL');
reject(new Error(`CLI ${bin} timed out after ${timeoutMs}ms`));
}, timeoutMs);
child.on('close', (code) => {
clearTimeout(timer);
resolve({ stdout, stderr, exitCode: code ?? 1 });
});
child.on('error', (err) => {
clearTimeout(timer);
reject(err);
});
});
}
/**
* Run a `tea api` call and parse the response envelope.
*
* `tea api --include` outputs:
* <status line: HTTP/1.1 200 OK>
* <headers, one per line>
* <empty line>
* <JSON body>
*
* Without `--include`, stdout is just the JSON body on success.
*/
async function teaApiCall(endpoint, { method = 'GET', body, raw, paginate, teaBin, token }) {
const args = ['api', '--include'];
if (method !== 'GET') args.push('-X', method);
if (paginate) args.push('--paginate');
if (raw) args.push('--header', 'Accept: text/plain');
if (body !== undefined) args.push('--header', 'Content-Type: application/json', '-d', JSON.stringify(body));
args.push(endpoint);
let result;
try {
result = await spawnCli(teaBin, args, { GITEA_SERVER_TOKEN: token });
} catch (err) {
return { status: 500, headers: {}, data: null, page: null, error: err.message };
}
if (retryMs === null) {
// Gitea sends `X-RateLimit-Reset`; check the generic name too for robustness.
const reset = headerValue(headers, 'x-ratelimit-reset') ?? headerValue(headers, 'ratelimit-reset');
if (reset !== undefined && reset !== null) {
const delta = Number(reset) * 1000 - Date.now();
if (Number.isFinite(delta) && delta > 0) retryMs = delta;
const { stdout, stderr, exitCode } = result;
if (exitCode !== 0 && !stdout.trim()) {
return { status: 500, headers: {}, data: null, page: null, error: stderr.trim() || `tea exited with code ${exitCode}` };
}
// Parse --include output: status line, headers, blank line, body.
const lines = stdout.split('\n');
let status = 200;
const headers = {};
let bodyStart = 0;
const statusMatch = lines[0]?.match(/HTTP\/\S+\s+(\d+)/);
if (statusMatch) {
status = Number(statusMatch[1]);
for (let i = 1; i < lines.length; i++) {
if (lines[i].trim() === '') {
bodyStart = i + 1;
break;
}
const colonIdx = lines[i].indexOf(':');
if (colonIdx > 0) {
headers[lines[i].slice(0, colonIdx).trim().toLowerCase()] = lines[i].slice(colonIdx + 1).trim();
}
}
}
if (retryMs === null) retryMs = DEFAULT_COOLDOWN_MS;
retryMs = Math.min(retryMs, MAX_COOLDOWN_MS);
const until = Date.now() + retryMs;
if (until > rateLimitedUntil) {
rateLimitedUntil = until;
console.warn(`[gitea] rate limited — pausing Gitea calls for ~${Math.round(retryMs / 1000)}s`);
const bodyText = lines.slice(bodyStart).join('\n').trim();
if (!bodyText) {
return { status, headers, data: null, page: null };
}
if (raw) {
return { status, headers, data: bodyText, page: null };
}
try {
return { status, headers, data: JSON.parse(bodyText), page: null };
} catch {
return { status, headers, data: bodyText, page: null };
}
}
export function isGiteaRateLimited() {
return Date.now() < rateLimitedUntil;
}
// ---- Rate-limit helpers (no-ops with CLI transport) ----
export function noteGiteaRateLimit() { /* no-op: CLI processes are stateless */ }
export function isGiteaRateLimited() { return false; }
// ---- Response helpers ----
const joinApiUrl = (baseUrl, path) => {
const base = String(baseUrl || '').replace(/\/+$/, '');
// Build the relative API path for tea CLI. tea resolves the base URL from its
// own login config, so we pass only the /api/v1/... portion.
const apiPath = (path) => {
const p = typeof path === 'string' && path ? (path.startsWith('/') ? path : `/${path}`) : '';
return `${base}/api/v1${p}`;
};
const headersToObject = (headers) => {
const out = {};
if (!headers) return out;
if (typeof headers.forEach === 'function') {
headers.forEach((value, key) => {
out[key] = value;
});
} else if (typeof headers === 'object') {
for (const [key, value] of Object.entries(headers)) {
out[key] = value;
}
}
return out;
};
const parsePageInfo = (headers) => {
const get = (name) => {
const value = headerValue(headers, name);
return typeof value === 'string' ? value : '';
};
// Gitea paginates list endpoints via the `Link` header (rel="next") and
// reports the total via `X-Total-Count`.
const linkHeader = get('link');
const relNextMatch = linkHeader.match(/<([^>]+)>\s*;\s*rel="next"/);
const totalRaw = get('x-total-count');
const total = totalRaw ? Number(totalRaw) : null;
const parsed = {
page: null,
next: null,
total: total !== null && Number.isFinite(total) ? total : null,
hasMore: Boolean(relNextMatch),
};
if (relNextMatch) {
parsed.nextUrl = relNextMatch[1];
}
return parsed;
};
const parseData = async (response, raw) => {
const text = await response.text();
if (raw) {
return text;
}
if (!text) {
return null;
}
try {
return JSON.parse(text);
} catch {
return null;
}
return `/api/v1${p}`;
};
/**
* Create a raw-fetch Gitea/Forgejo REST v1 client. `request` never throws for
* HTTP error statuses — it returns `{ status, headers, data, page }` so callers
* can branch on status codes. On 429 it also sets `error: 'Gitea rate limited'`
* and records a module-level cooldown.
* Create a CLI-backed Gitea/Forgejo REST v1 client. Spawns `tea api` for each
* request. `request` never throws for HTTP error statuses — it returns
* `{ status, headers, data, page }` so callers can branch on status codes.
*/
export function createGiteaClient({ token, baseUrl }) {
const effectiveBaseUrl = typeof baseUrl === 'string' ? baseUrl.trim().replace(/\/+$/, '') : '';
const teaBin = process.env.TEA_BIN || '/home/user/.local/bin/tea';
const request = async (path, options = {}) => {
const method = (typeof options.method === 'string' ? options.method : 'GET').toUpperCase();
const query = options.query && typeof options.query === 'object' ? options.query : {};
const body = options.body;
const callerSignal = options.signal;
const raw = options.raw === true;
if (isGiteaRateLimited()) {
return { status: 429, headers: {}, data: null, page: null, error: 'Gitea rate limited' };
}
let url = joinApiUrl(effectiveBaseUrl, path);
// Build the relative endpoint path with query params baked in.
let endpoint = apiPath(path);
const qs = new URLSearchParams();
let hasQuery = false;
for (const [key, value] of Object.entries(query)) {
@@ -200,61 +150,11 @@ export function createGiteaClient({ token, baseUrl }) {
hasQuery = true;
}
if (hasQuery) {
url += `${url.includes('?') ? '&' : '?'}${qs.toString()}`;
endpoint += `${endpoint.includes('?') ? '&' : '?'}${qs.toString()}`;
}
// Gitea/Forgejo PAT auth: `Authorization: token <pat>`.
const headers = {
Authorization: `token ${token}`,
accept: raw ? 'text/plain' : 'application/json',
};
const fetchOptions = {
method,
headers,
redirect: 'manual',
};
if (body !== undefined) {
headers['content-type'] = 'application/json';
fetchOptions.body = JSON.stringify(body);
}
if (callerSignal) {
fetchOptions.signal = callerSignal;
}
const conditionalFetch = createConditionalFetch(token);
let response = await conditionalFetch(url, fetchOptions);
// Follow redirects (301/302/308) exactly once. Gitea serves them for moved
// repos/users; a manual redirect keeps our Authorization header across the hop.
// Only follow same-origin redirects to avoid leaking the token to a different host.
let redirects = 0;
const baseHost = new URL(url).host;
while (
(response.status === 301 || response.status === 302 || response.status === 308)
&& headerValue(response.headers, 'location')
&& redirects < 1
) {
const location = headerValue(response.headers, 'location');
const nextUrl = new URL(location, url).toString();
if (new URL(nextUrl).host !== baseHost) break;
response = await conditionalFetch(nextUrl, fetchOptions);
redirects += 1;
}
const result = {
status: response.status,
headers: headersToObject(response.headers),
data: await parseData(response, raw),
page: parsePageInfo(response.headers),
};
if (response.status === 429) {
noteGiteaRateLimit(response);
result.error = 'Gitea rate limited';
}
return result;
const paginate = method === 'GET' && hasQuery;
return teaApiCall(endpoint, { method, body, raw, paginate, teaBin, token });
};
return {
@@ -302,8 +202,6 @@ export function createGiteaClient({ token, baseUrl }) {
request(`/repos/${owner}/${repo}/pulls/${number}/merge`, { method: 'POST', body }),
branches: (owner, repo, params = {}) =>
request(`/repos/${owner}/${repo}/branches`, { query: params }),
// Assignable users (collaborators with role access + org members) are the
// mention/assign candidate set; Gitea mirrors the GitHub assignees route.
assignees: (owner, repo, params = {}) =>
request(`/repos/${owner}/${repo}/assignees`, { query: params }),
tags: (owner, repo, params = {}) =>
@@ -320,9 +218,6 @@ export function getGiteaClientOrNull(directory) {
let baseUrl = auth.baseUrl;
if (directory) {
const effectiveBaseUrl = getEffectiveProviderApiBaseUrl('gitea', directory);
// Only a per-project override replaces the account's base URL; without one
// the effective value is just the global default, which stored accounts
// (an explicit baseUrl is required) already outrank.
if (effectiveBaseUrl !== null && effectiveBaseUrl !== getProviderApiBaseUrl('gitea')) {
baseUrl = effectiveBaseUrl;
}