Files
OmniRoute/open-sse/executors/gemini-cli.ts
Diego Rodrigues de Sa e Souza 555b21d296 Release v3.8.37 (#5053)
* chore(release): open v3.8.37 development cycle

* chore(ci): harden release flow — ratchet decoupling, fast-path drift gates, build-scope guard, heap default (#5054)

Implements improvements 1-4 from the v3.8.36 release benchmark (_tasks/release-bench/v3.8.36/PLANO-MELHORIA.md):

1. Quality Ratchet decoupled from flaky coverage (ci.yml): the shard→coverage→ratchet
   chain meant a single flaky Coverage Shard SKIPPED the whole Quality Ratchet on the
   release PR (v3.8.36 #4854), so cycle drift only surfaced post-merge in #5029. The job
   now runs on !cancelled(); coverage download is continue-on-error and the ratchet runs
   --allow-missing, so the DETERMINISTIC gates (eslint/complexity/cognitive/duplication/
   codeql) stay blocking even when coverage is unavailable.

2. Fast-path drift gates (quality.yml PR→release): added check:complexity,
   check:cognitive-complexity, and a new lightweight check:pack-policy (pack-artifact
   unexpected-files check WITHOUT a build, via --policy-only) so drift + stray-tarball-file
   regressions are caught/rebaselined PER-PR instead of cascading onto the release PR.

3. Build heap default 4096→8192 MB (build-next-isolated.mjs): the clean graph peaks
   ~3.9 GB and brushed the old 4 GB ceiling; 8 GB gives headroom. Comment notes heap is
   NOT the fix for a poisoned scope (run check:build-scope instead).

4. check:build-scope gate (new): fails if .ts/.tsx/.js/.jsx files in the tsconfig scope
   exceed a threshold — catches worktrees/cruft leaking into the build scope (the v3.8.36
   OOM root cause: 355,215 vs 4,547 files) BEFORE it detonates next build. Wired into the
   fast-path.

* fix(auth): only trust forwarding headers from loopback TCP peers (#4689)

Integrated into release/v3.8.37 — loopback-gated forwarding headers (IP spoofing fix). Cherry-picked onto current release tip; ipUtils.test.ts 9/9 green.

* fix(codex): treat OAuth 401 as unrecoverable refresh failure (#4686)

Integrated into release/v3.8.37 — codex OAuth 401 treated as unrecoverable refresh. Cherry-picked onto release tip; token-refresh-service.test.ts 38/38 green.

* fix(translator): preserve reasoning_effort for non-Copilot Responses clients (#4688)

Integrated into release/v3.8.37 — preserve reasoning_effort for non-Copilot Responses clients. Cherry-picked onto release tip; tests 47/47 green.

* fix(translator): coerce tool descriptions to strings in OpenAI normalization (#4675)

Integrated into release/v3.8.37 — coerce tool descriptions to strings in OpenAI normalization. Cherry-picked onto release tip; tests 3/3 green.

* feat(sse): x-omniroute-strip-reasoning header to drop reasoning_content (#4678)

Integrated into release/v3.8.37 — x-omniroute-strip-reasoning header. Cherry-picked onto release tip (resolved chatCore.ts/headers.ts adjacency conflict, kept resolveCompressionHeader + isStripReasoningRequested); tests 8/8 green.

* fix(combo): flatten Anthropic tool messages + tool history to prevent upstream 503 (#4648)

Integrated into release/v3.8.37 — flattenToolHistory helper (combo anti-503). Cherry-picked onto release tip; tests 9/9 green.

* feat(headroom): proxy lifecycle management + dashboard UI (Docker sidecar supported) (#4649)

Integrated into release/v3.8.37 — headroom proxy lifecycle (status/start/stop, local-only + spawn-capable per Rules #15/#17). Cherry-picked onto release tip; lifecycle 7/7 + route-guard 43/43 + check:cycles green.

* feat(cli): multi-model support for Factory Droid CLI (#4682)

Integrated into release/v3.8.37 — Factory Droid multi-model support. Cherry-picked onto release tip (kept readJsoncConfig + droidCustomModels imports); droid-custom-models 11/11 green.

* fix(providers): require Default Model in compatible-provider API-key setup (#4641)

Integrated into release/v3.8.37 — require Default Model in compatible-provider API-key setup. Cherry-picked fix + test-move onto release tip (kept release providerSpecificData + QuotaScrapingFields; fixed moved-test import path; baseline rebaseline unneeded, 865<866); UI test 2/2 green.

* fix(dashboard): stop double-masking already-masked API key in list (E2E 3/9 regression) (#4671)

Integrated into release/v3.8.37 — render server-masked key verbatim (drop redundant maskKey call). Note: release's maskKey already guards '****' (since v3.8.34), so this is a safe simplification; added a contract test pinning the **** passthrough invariant (2/2 green, would fail against the pre-guard maskKey = the historical double-mask bug).

* chore(quality): rebaseline file-size for rc17 PR batch drift

Own growth from the merged rc17 PRs (#4678/#4686/#4688) at existing chokepoints —
cohesive, not extractable:
- open-sse/handlers/responseSanitizer.ts 1103->1122 (SanitizeOpenAIResponseOptions + stripReasoning, #4678)
- open-sse/services/tokenRefresh.ts 2070->2090 (codex 401 unrecoverable-refresh guard, #4686)
- tests/unit/token-refresh-service.test.ts 1322->1353 (401 regression case, #4686)
- tests/unit/translator-openai-responses-req.test.ts 1047->1050 (reasoning_effort assertion, #4688)

* docs(env): document HEADROOM_URL in .env.example + ENVIRONMENT.md

The headroom proxy lifecycle (#4649) reads HEADROOM_URL (src/lib/headroom/detect.ts,
default http://localhost:8787) but it was missing from the env contract, tripping
check:env-doc-sync. Adds the var to both .env.example (commented, has a default) and
the Proxy Health table in ENVIRONMENT.md.

* fix(sse): stream writer mock abort() returns a Promise (#4788)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(cli): fall back to default data dir when DATA_DIR is not writable (#4767)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(oauth): verify Cursor installation on Linux before auto-import (#4770)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(sse): track Ollama streaming usage from raw NDJSON chunks (#4754)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(sse): strip enumDescriptions from antigravity tool schema (#4740)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(sse): include low-level cause details in formatProviderError (#4741)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(translator): strip x-anthropic-billing-header in claude-to-openai (#4728)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(sse): gate Kiro image attachments behind a Claude-capability check (#4763)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(sse): read Antigravity usage from the response.usageMetadata envelope (#4785)

Integrated into release/v3.8.37 — Antigravity response.usageMetadata envelope. Cherry-picked onto release tip (resolved test-tail adjacency with #4754 Ollama block); usage-extractor 23/23 green.

* fix(api): fall back to existing access token for any OAuth provider on refresh failure (#4786)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(cli): verify launchd registration + skip self-SIGTERM in macOS autostart (#4765)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(executors): anthropic-compatible-* gateways get Bearer alongside x-api-key (#4729)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(sse): json_schema fallback for OpenAI-compatible providers (#4766)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(sse): use workos auth token shape for cline (#4787)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* feat(sse): parse Gemini CLI 429 retryDelay from structured RetryInfo (#4738)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; tests green.

* fix(sse): finalize tool_calls finish_reason on early stream end in OpenAI Responses translator (#4764)

Integrated into release/v3.8.37 — computeFinishReason finalizes tool_calls on early stream end (Responses translator). Cherry-picked onto release tip; responses-translation-fixes 29/29 green.

* test(sse): golden-lock provider.ts translate-path across all providers (#4734)

Integrated into release/v3.8.37 — golden-lock for provider.ts translate-path. Cherry-picked onto release tip; snapshot regenerated against the current provider set (UPDATE_GOLDEN=1, 167 entries); golden test 3/3 deterministic.

* chore(quality): rebaseline file-size for rc17 leva2 PR batch drift

Own growth from the merged leva2 PRs (cohesive, not extractable):
- src/lib/usage/providerLimits.ts 950->955 (#4786)
- open-sse/executors/default.ts NEW frozen @828 (#4729 + #4766 + #4787 header branches)
- open-sse/translator/request/openai-to-kiro.ts 807->814 (#4763)
- open-sse/translator/response/openai-responses.ts 923->937 (#4764)
- tests/unit/executor-default-base.test.ts 1339->1440 (#4766)
- tests/unit/translator-openai-to-kiro.test.ts 918->980 (#4763)

* fix(dashboard): align Engine Combos editor engines with API schema (#4955) (#5062)

The named-combos pipeline dropdown offered four engines (headroom,
session-dedup, ccr, llmlingua) that stackedPipelineStepSchema rejects, so
selecting one made PUT /api/context/combos/[id] return HTTP 400 while
saveCombo swallowed the non-OK response (if (!res.ok) return). Editing the
default 'Standard Savings' combo and changing an engine reproduced the 400.

- Add canonical STACKED_PIPELINE_ENGINE_INTENSITIES next to the schema as the
  single source of truth; the client dropdown imports it so it can never drift
  from the discriminated union the API validates against.
- Surface save errors and empty-name/empty-pipeline validation in the editor
  instead of failing silently.
- Add a parity unit test asserting the UI engine map equals the schema union
  and that every (engine, intensity) the UI emits is accepted.

* fix(sse): filter nameless hosted tools when converting Responses API to Chat format (#4789)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(dashboard): keep desktop sidebar visible via explicit CSS class (#4812)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): strip enumDescriptions from Antigravity tool schemas (#4813)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(dashboard): resolve passthrough model aliases by providerId in ModelSelectModal (#4815)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(oauth): allow per-connection refresh lead-time override via providerSpecificData.refreshLeadMs (#4818)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): strip X-Stainless-* headers and normalize SDK User-Agent for OpenAI-compatible endpoints (#4820)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): strip Gemini built-in tools when functionDeclarations present in Antigravity envelope (#4821)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(api): surface a Docker-localhost hint on provider-node validation connection errors (#4822)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): resolve bare model names to connection defaultModel before upstream calls (#4825)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(build): trace-include sql.js sql-wasm.wasm in standalone bundle (#4839)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): strip Composer <|final|> sentinel markers leaking after Composer reasoning (#4842)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(config): sync full SiliconFlow model list into registry (#4844)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): close reasoning before message content in Responses stream (#4848)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): reject unsupported Kiro [1m] context suffix (#4816)

Integrated into release/v3.8.37 — cherry-picked onto release tip; test-tail conflict with #4763 resolved (kept both image + [1m] test blocks); CHANGELOG re-merged; 29/29 green.

* fix(db): validate HuggingFace tokens via whoami-v2 auth probe (#4819)

Integrated into release/v3.8.37 — defining commit re-homed onto the god-file-split validation module (validateHuggingFaceProvider in validation/openaiFormat.ts + map wiring); 115/115 green.

* fix(sse): make anthropic-version default-guard case-insensitive (#4823)

Integrated into release/v3.8.37 — conflict with #4729 Bearer-fallback resolved (kept both Bearer fallback + case-insensitive anthropic-version guard); 48/48 green.

* fix(sse): sanitize Kiro tool schemas to avoid 400 "Improperly formed request" (#4847)

Integrated into release/v3.8.37 — conflict in kiro-to-openai.ts resolved (kept release fallbackToolCallId + adopted #1375 toolNameMap remap); 7/7 green.

* feat(sse): add GPT-4 to the GitHub Copilot provider (#4798)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* feat(sse): add GPT-4o mini to GitHub Copilot provider (#4797)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* feat(api): add MiniMax-M3 pricing row (#4814)

Integrated into release/v3.8.37 — pricing row re-homed onto god-file-split pricing/regional.ts (pricing.ts is now a barrel); 4/4 green.

* fix(cli): save runtime deps with --save-exact so a sibling install can't prune them (#4841)

Integrated into release/v3.8.37 — trayRuntime conflict resolved (kept release SYSTRAY_SPEC + added --save-exact); 2/2 green.

* fix(sse): preserve required fields in antigravity tool schemas (#4843)

Integrated into release/v3.8.37 — conflict resolved (kept #4740/#4813 enumDescriptions strip + typed normalizeSchemaTypes, added required-preservation helpers; test-tail merged keeping both enumDescriptions + required tests); 7/7 green.

* chore(quality): rebaseline file-size for rc17b leva3 PR batch drift

* fix(sse): strip reasoning blobs from agentic context to prevent O(n^2) token growth (#4849)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): unwrap Qoder HTTP 200 SSE error envelope so fallback can trigger (#4850)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): strip temperature for Claude models with extended thinking (#4853)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): emit valid concatenable kiro tool_calls.arguments deltas (#4855)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* feat(sse): add toggleable tool-source diagnostics (#4856)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): redact api key from the AUTH debug log in the chat handler (#4858)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): forward AI SDK image parts in Responses translator (#4859)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): resolve custom combos by id and case-insensitive name (#4446) (#4869)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): exclude WS bridge controller-closed error from provider breaker (#4602) (#4870)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* feat(providers): add xAI Grok inbound translators and thinking patcher (#4910)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* feat(embeddings): add dimensions override field to embedding combos (#4913)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* feat(oauth): Codex bulk-import endpoint — POST /api/oauth/codex/import (#4914)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(antigravity): retry transient upstream failures (#4941)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): surface malformed HTTP-200 upstream responses (#4942)

Integrated into release/v3.8.37 — cherry-picked defining commit onto release tip; CHANGELOG re-merged; tests green.

* fix(sse): normalize Codex custom tools (apply_patch) to { input: string } schema (#4862)

Integrated into release/v3.8.37 — conflict in request/openai-responses.ts resolved (kept #4789 nameless-tool skip + added #1007 custom-tool {input:string} normalization); 48/48 green incl. #4789/#4859 regression.

* fix(sse): dense, deterministic output ordering in Responses API response.completed (#4906)

Integrated into release/v3.8.37 — manual integration with #4862 in response/openai-responses.ts (custom-tool funcItem + dense recordCompletedItem). Fixed a latent #4848 interaction: the close-reasoning-before-message guard force-closed <think>-tag reasoning prematurely, which dense output (#4906) then snapshotted as a partial buffer ("plan" vs "planning") — scoped the guard to native reasoning_content (!inThinking) in BOTH transformer + translator paths. Full Responses suite 203/203 green incl. #4848/#4862 regression.

* feat(sse): auto-promote successful combo model to position #1 (#4852)

Integrated into release/v3.8.37 — dropped the stale file-size-baseline.json hunk (re-derived against the rc17b rebaseline); code+test applied clean; 13/13 green.

* feat(providers): add Pioneer AI (Fastino Labs) provider (#4909)

Integrated into release/v3.8.37 — providers.ts apikey block re-homed onto god-file-split src/shared/constants/providers/apikey/frontier-labs.ts (inline APIKEY_PROVIDERS no longer exists); registry/pioneer + providers/index.ts applied clean; 6/6 green.

* add DGrid AI gateway provider (#4931)

Integrated into release/v3.8.37 — rebased the contributor's commit onto the release tip; providers.ts god-file-split conflict resolved by relocating the dgrid APIKEY_PROVIDERS entry into apikey/gateways.ts; CHANGELOG added. 7/7 green. Thanks @dgridOP!

* chore(quality): rebaseline file-size for rc17b leva4 PR batch drift

* docs(routing): sync combo strategy docs for Fusion (17 strategies) (#5067)

Fusion (16th strategy, panel fan-out + judge synthesis) and headroom
shipped but the strategy-count docs were stale (14/15) and omitted both.
Update every combo-strategy reference to the canonical 17, add fusion +
headroom to all strategy tables, and add a dedicated Fusion section to
AUTO-COMBO.md documenting judgeModel / fusionTuning config + an example.

- CLAUDE.md, README.md, FEATURES.md, RESILIENCE_GUIDE.md,
  ARCHITECTURE.md, OPEN_SSE_ARCHITECTURE.md, OMNIROUTE_VS_ALTERNATIVES.md,
  docs/README.md, request-pipeline.mmd: 14/15 -> 17, list fusion + headroom
- docs/routing/AUTO-COMBO.md: strategy table + new Fusion strategy section
- docs/openapi.yaml: add reset-window, headroom, fusion to the strategy enum

* fix(oauth): classify /api/oauth/cursor/auto-import as local-only (route-guard) (#5070)

The Cursor auto-import route runs execFile("which", ["cursor"]) to verify a
local Cursor install before importing credentials — a child-process spawn. The
check:route-guard-membership gate (Hard Rules #15/#17) flagged it as an
unclassified spawn-capable route: reachable past the loopback gate, an
RCE-via-tunnel surface (a leaked JWT over a tunnel could trigger the spawn).

Classify the specific path in LOCAL_ONLY_API_PREFIXES so loopback enforcement
runs unconditionally before any auth check. Scoped to the exact path — the rest
of /api/oauth/ (browser redirect/callback flows) stays remote-reachable.

TDD: added a failing-then-passing assertion in route-guard-local-prefix.test.ts
(classification + an over-broadening guard proving sibling OAuth paths stay
remote). check:route-guard-membership now reports 0 new gaps.

* chore(release): v3.8.37 — 2026-06-26

---------

Co-authored-by: dgridOP <dgrid_op@outlook.com>
2026-06-26 02:51:06 -03:00

588 lines
19 KiB
TypeScript

import { BaseExecutor, mergeUpstreamExtraHeaders, mergeAbortSignals } from "./base.ts";
import { randomUUID } from "crypto";
import { PROVIDERS, OAUTH_ENDPOINTS, FETCH_TIMEOUT_MS } from "../config/constants.ts";
import { getGeminiCliHeaders } from "../services/geminiCliHeaders.ts";
import { scrubProxyAndFingerprintHeaders } from "../services/antigravityHeaderScrub.ts";
import { obfuscateSensitiveWords } from "../services/antigravityObfuscation.ts";
import {
shouldStripCloudCodeThinking,
stripCloudCodeThinkingConfig,
} from "../services/cloudCodeThinking.ts";
const LOAD_CODE_ASSIST_URL = "https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist";
const ONBOARD_USER_URL = "https://cloudcode-pa.googleapis.com/v1internal:onboardUser";
const PROJECT_TTL_MS = 30_000; // 30 seconds — matches native Gemini CLI
const MAX_CACHE_SIZE = 100;
const LOAD_CODE_ASSIST_TIMEOUT_MS = 10_000; // 10 seconds timeout
const ONBOARD_TIMEOUT_MS = 30_000;
const ONBOARD_MAX_ATTEMPTS = 10;
const ONBOARD_DELAY_MS = 5_000;
const DEFAULT_PROJECT_ID = "default-project";
const DEFAULT_ONBOARD_TIER = "free-tier";
const LOAD_CODE_ASSIST_METADATA = Object.freeze({
ideType: "IDE_UNSPECIFIED",
platform: "PLATFORM_UNSPECIFIED",
pluginType: "GEMINI",
});
const ONBOARD_METADATA = Object.freeze({
ideType: "IDE_UNSPECIFIED",
pluginType: "GEMINI",
});
// Per-account cache: accessToken -> { projectId, expiresAt }
const projectCache = new Map<string, { projectId: string; expiresAt: number }>();
// In-flight deduplication: prevents thundering herd on cache miss
const inflightRefresh = new Map<string, Promise<string | null>>();
type LoadCodeAssistResponse = {
cloudaicompanionProject?: string | { id?: string | null } | null;
allowedTiers?: Array<{ id?: string | null; isDefault?: boolean | null }> | null;
};
type OnboardOptions = {
attempts?: number;
delayMs?: number;
};
function normalizeGeminiModel(model: string): string {
return typeof model === "string" && model.trim().length > 0
? model.replace(/^models\//, "").trim()
: "unknown";
}
function generateGeminiCliRequestId(): string {
return `agent-${randomUUID()}`;
}
function generateGeminiCliSessionId(): string {
return `-${Date.now()}`;
}
function cloneGeminiCliRecord(value: Record<string, any>): Record<string, any> {
if (typeof structuredClone === "function") {
return structuredClone(value);
}
return JSON.parse(JSON.stringify(value));
}
function extractProjectId(payload: unknown): string {
if (!payload || typeof payload !== "object") return "";
const data = payload as LoadCodeAssistResponse;
if (typeof data.cloudaicompanionProject === "string") {
return data.cloudaicompanionProject.trim();
}
if (typeof data.cloudaicompanionProject?.id === "string") {
return data.cloudaicompanionProject.id.trim();
}
return "";
}
function resolveGeminiCliProjectId(value: unknown): string {
if (typeof value !== "string") return "";
const trimmed = value.trim();
if (!trimmed) return "";
const normalized = trimmed.toLowerCase();
if (normalized === DEFAULT_PROJECT_ID || normalized === `projects/${DEFAULT_PROJECT_ID}`) {
return "";
}
return trimmed;
}
function extractDefaultTierId(payload: unknown): string {
if (!payload || typeof payload !== "object") return DEFAULT_ONBOARD_TIER;
const tiers = Array.isArray((payload as LoadCodeAssistResponse).allowedTiers)
? (payload as LoadCodeAssistResponse).allowedTiers
: [];
for (const tier of tiers) {
if (tier?.isDefault && typeof tier.id === "string" && tier.id.trim()) {
return tier.id.trim();
}
}
return DEFAULT_ONBOARD_TIER;
}
function cacheProject(accessToken: string, projectId: string): void {
if (projectCache.size >= MAX_CACHE_SIZE) {
const now = Date.now();
for (const [key, val] of projectCache) {
if (val.expiresAt <= now) projectCache.delete(key);
}
if (projectCache.size >= MAX_CACHE_SIZE) {
const firstKey = projectCache.keys().next().value;
if (firstKey !== undefined) projectCache.delete(firstKey);
}
}
projectCache.set(accessToken, {
projectId,
expiresAt: Date.now() + PROJECT_TTL_MS,
});
}
function sleep(ms: number): Promise<void> {
if (!ms || ms <= 0) return Promise.resolve();
return new Promise((resolve) => setTimeout(resolve, ms));
}
export class GeminiCLIExecutor extends BaseExecutor {
constructor() {
super("gemini-cli", PROVIDERS["gemini-cli"]);
}
buildUrl(model, stream, urlIndex = 0) {
void model;
void urlIndex;
const action = stream ? "streamGenerateContent?alt=sse" : "generateContent";
return `${this.config.baseUrl}:${action}`;
}
buildHeaders(
credentials,
stream = true,
clientHeaders?: Record<string, string> | null,
model?: string
) {
void clientHeaders;
const activeModel = model || "unknown";
const raw = getGeminiCliHeaders(
normalizeGeminiModel(activeModel),
credentials.accessToken,
stream ? "*/*" : "application/json"
);
if (credentials.apiKey) {
raw["x-goog-api-key"] = credentials.apiKey;
// getGeminiCliHeaders adds Authorization: Bearer undefined if accessToken is empty, so we clean it up
if (!credentials.accessToken) {
delete raw["Authorization"];
}
}
return scrubProxyAndFingerprintHeaders(raw);
}
async onboardManagedProject(
accessToken: string,
tierId = DEFAULT_ONBOARD_TIER,
options: OnboardOptions = {},
model = "unknown"
): Promise<string | null> {
const currentModel = normalizeGeminiModel(model);
const attempts =
Number.isInteger(options.attempts) && options.attempts! > 0
? Number(options.attempts)
: ONBOARD_MAX_ATTEMPTS;
const delayMs =
typeof options.delayMs === "number" &&
Number.isFinite(options.delayMs) &&
options.delayMs >= 0
? options.delayMs
: ONBOARD_DELAY_MS;
const requestBody = {
tierId: tierId || DEFAULT_ONBOARD_TIER,
metadata: { ...ONBOARD_METADATA },
};
for (let attempt = 0; attempt < attempts; attempt += 1) {
try {
const controller = new AbortController();
const timeoutId = setTimeout(() => controller.abort(), ONBOARD_TIMEOUT_MS);
let response: Response;
try {
response = await fetch(ONBOARD_USER_URL, {
method: "POST",
headers: getGeminiCliHeaders(currentModel, accessToken, "application/json"),
body: JSON.stringify(requestBody),
signal: controller.signal,
});
} finally {
clearTimeout(timeoutId);
}
if (response.ok) {
const payload = await response.json();
const managedProjectId = extractProjectId(payload?.response);
if (payload?.done === true && managedProjectId) {
return managedProjectId;
}
if (payload?.done === true) {
return null;
}
} else {
console.warn(
`[OmniRoute] onboardUser returned ${response.status} on attempt ${attempt + 1}`
);
}
} catch (error) {
const msg = error instanceof Error ? error.message : String(error);
console.warn(`[OmniRoute] onboardUser attempt ${attempt + 1} failed (${msg})`);
}
if (attempt < attempts - 1) {
await sleep(delayMs);
}
}
return null;
}
/**
* Fetch the current cloudaicompanionProject via loadCodeAssist API.
* Native Gemini CLI refreshes this every 30 seconds — OmniRoute stores it once
* at OAuth connection time, so it goes stale. This method keeps it fresh.
*/
async refreshProject(accessToken: string, model = "unknown"): Promise<string | null> {
// Check cache
const cached = projectCache.get(accessToken);
if (cached && cached.expiresAt > Date.now()) {
return cached.projectId;
}
// Deduplicate in-flight requests (thundering herd prevention)
const inflight = inflightRefresh.get(accessToken);
if (inflight) return inflight;
const promise = this._doRefresh(accessToken, model);
inflightRefresh.set(accessToken, promise);
try {
return await promise;
} finally {
inflightRefresh.delete(accessToken);
}
}
async _doRefresh(accessToken: string, model = "unknown"): Promise<string | null> {
const currentModel = normalizeGeminiModel(model);
try {
const controller = new AbortController();
const timeoutId = setTimeout(() => controller.abort(), LOAD_CODE_ASSIST_TIMEOUT_MS);
let response: Response;
try {
response = await fetch(LOAD_CODE_ASSIST_URL, {
method: "POST",
headers: getGeminiCliHeaders(currentModel, accessToken, "application/json"),
body: JSON.stringify({
metadata: { ...LOAD_CODE_ASSIST_METADATA },
}),
signal: controller.signal,
});
} finally {
clearTimeout(timeoutId);
}
if (!response.ok) {
console.warn(
`[OmniRoute] loadCodeAssist returned ${response.status} — falling back to stored projectId`
);
return null;
}
const data = (await response.json()) as LoadCodeAssistResponse;
let projectId = resolveGeminiCliProjectId(extractProjectId(data));
if (!projectId) {
console.warn(
"[OmniRoute] loadCodeAssist returned no project — attempting managed project onboarding"
);
projectId = await this.onboardManagedProject(
accessToken,
extractDefaultTierId(data),
{},
currentModel
);
}
if (!projectId) {
console.warn(
"[OmniRoute] managed project onboarding failed — falling back to stored projectId"
);
return null;
}
cacheProject(accessToken, projectId);
return projectId;
} catch (error) {
const msg = error instanceof Error ? error.message : String(error);
console.warn(`[OmniRoute] loadCodeAssist failed (${msg}) — falling back to stored projectId`);
return null;
}
}
async transformRequest(model, body, stream, credentials) {
const currentModel = normalizeGeminiModel(model);
const normalizedBody =
shouldStripCloudCodeThinking(this.provider, currentModel) && body && typeof body === "object"
? stripCloudCodeThinkingConfig(body)
: body;
const bodyRecord =
normalizedBody && typeof normalizedBody === "object"
? (normalizedBody as Record<string, any>)
: { request: {} };
const requestRecord =
bodyRecord.request && typeof bodyRecord.request === "object"
? cloneGeminiCliRecord(bodyRecord.request as Record<string, any>)
: {};
const providerSpecificData = credentials.providerSpecificData as Record<string, unknown>;
const storedProject =
resolveGeminiCliProjectId(providerSpecificData?.projectId) ||
resolveGeminiCliProjectId(credentials.projectId) ||
resolveGeminiCliProjectId(bodyRecord.project) ||
"";
const envelope: Record<string, any> = {
model: currentModel,
user_prompt_id: bodyRecord.user_prompt_id || generateGeminiCliRequestId(),
request: {
...requestRecord,
session_id: requestRecord.session_id || generateGeminiCliSessionId(),
},
};
if (typeof storedProject === "string" ? storedProject.trim() : storedProject) {
envelope.project = storedProject;
}
for (const [key, value] of Object.entries(bodyRecord)) {
if (!(key in envelope) && key !== "request" && key !== "project") {
envelope[key] = value;
}
}
// Native Gemini CLI refreshes the Cloud Code project periodically because
// stored project IDs can go stale. Keep the stored value as a fallback.
if (credentials.accessToken) {
const freshProject = await this.refreshProject(credentials.accessToken, currentModel);
if (resolveGeminiCliProjectId(freshProject)) {
envelope.project = freshProject;
}
}
if (envelope.request) {
// Obfuscate sensitive client names in user content
const contents = envelope.request?.contents;
if (Array.isArray(contents)) {
for (const msg of contents) {
if (Array.isArray(msg.parts)) {
for (const part of msg.parts) {
if (typeof part.text === "string") {
part.text = obfuscateSensitiveWords(part.text);
}
}
}
}
}
}
return envelope;
}
async execute({
model,
body,
stream,
credentials,
signal,
log,
upstreamExtraHeaders,
}: ExecuteInput) {
const fallbackCount = this.getFallbackCount();
let lastError = null;
let lastStatus = 0;
const MAX_AUTO_RETRIES = 3;
const retryAttemptsByUrl: Record<number, number> = {};
for (let urlIndex = 0; urlIndex < fallbackCount; urlIndex++) {
const url = this.buildUrl(model, stream, urlIndex);
const headers = this.buildHeaders(credentials, stream, null, model);
mergeUpstreamExtraHeaders(headers, upstreamExtraHeaders);
const transformed = await this.transformRequest(model, body, stream, credentials);
if (transformed instanceof Response) {
return { response: transformed, url, headers, transformedBody: body };
}
const transformedBody = transformed;
if (!retryAttemptsByUrl[urlIndex]) {
retryAttemptsByUrl[urlIndex] = 0;
}
try {
log?.debug?.(
"TELEMETRY",
`[Gemini CLI] Execute - URL: ${url}, Model: ${model}, Retry: ${retryAttemptsByUrl[urlIndex]}`
);
const timeoutSignal = AbortSignal.timeout(FETCH_TIMEOUT_MS);
const mergedSignal = signal ? mergeAbortSignals(signal, timeoutSignal) : timeoutSignal;
const response = await fetch(url, {
method: "POST",
headers,
body: JSON.stringify(transformedBody),
signal: mergedSignal,
});
if (!response.ok) {
log?.warn?.(
"TELEMETRY",
`[Gemini CLI] Error Response - URL: ${url}, Status: ${response.status}`
);
let retryMs: number | null = null;
if (response.status === 429 || response.status === 503) {
try {
const errorBody = await response.clone().text();
retryMs = this.parseRetryFromErrorMessage(errorBody);
} catch {
/* ignore parse error */
}
if ((!retryMs || retryMs <= 60000) && retryAttemptsByUrl[urlIndex] < MAX_AUTO_RETRIES) {
retryAttemptsByUrl[urlIndex]++;
const backoffMs =
retryMs || Math.min(1000 * 2 ** retryAttemptsByUrl[urlIndex], 30000);
log?.debug?.(
"RETRY",
`Gemini CLI 429 retry ${retryAttemptsByUrl[urlIndex]} after ${backoffMs}ms`
);
await sleep(backoffMs);
urlIndex--;
continue;
}
}
}
if (this.shouldRetry(response.status, urlIndex)) {
lastStatus = response.status;
continue;
}
return { response, url, headers, transformedBody };
} catch (error) {
lastError = error;
if (urlIndex + 1 < fallbackCount) continue;
throw error;
}
}
throw lastError || new Error(`All ${fallbackCount} URLs failed with status ${lastStatus}`);
}
async refreshCredentials(credentials, log) {
if (!credentials.refreshToken) return null;
try {
const response = await fetch(OAUTH_ENDPOINTS.google.token, {
method: "POST",
headers: {
"Content-Type": "application/x-www-form-urlencoded",
Accept: "application/json",
},
body: new URLSearchParams({
grant_type: "refresh_token",
refresh_token: credentials.refreshToken,
client_id: this.config.clientId,
client_secret: this.config.clientSecret,
}),
});
if (!response.ok) {
const errorText = await response.text().catch(() => "");
log?.error?.("TOKEN", "Gemini CLI refresh failed", {
status: response.status,
error: errorText.slice(0, 200),
});
// Match refreshGoogleToken's pattern: invalid_grant means the refresh
// token was revoked / replaced — surface as unrecoverable so the caller
// marks the account expired instead of retrying forever with a dead token.
try {
const errorBody = JSON.parse(errorText);
if (errorBody?.error === "invalid_grant") {
return { error: "unrecoverable_refresh_error", code: "invalid_grant" } as never;
}
} catch {
// not JSON — fall through
}
return null;
}
const tokens = await response.json();
log?.info?.("TOKEN", "Gemini CLI refreshed");
const refreshed: Record<string, unknown> = {
accessToken: tokens.access_token,
refreshToken: tokens.refresh_token || credentials.refreshToken,
expiresIn: tokens.expires_in,
projectId: credentials.projectId,
};
if (credentials.providerSpecificData !== undefined) {
refreshed.providerSpecificData = credentials.providerSpecificData;
}
return refreshed as never;
} catch (error) {
log?.error?.("TOKEN", `Gemini CLI refresh error: ${error.message}`);
return null;
}
}
// Parse retry time from Gemini error message body. Two shapes are handled:
// 1. Structured google.rpc.RetryInfo in the 429/503 JSON body:
// { error: { details: [{ "@type": ".../google.rpc.RetryInfo", retryDelay: "30s" }] } }
// 2. Human-readable prose: "Your quota will reset after 2h7m23s"
// The structured hint is authoritative (it is what the Cloud Code Assist API
// sends), so it is checked first and falls through to the prose form on miss.
parseRetryFromErrorMessage(errorMessage: unknown): number | null {
if (!errorMessage || typeof errorMessage !== "string") return null;
const structuredMs = this.parseStructuredRetryDelay(errorMessage);
if (structuredMs !== null) return structuredMs;
const match = errorMessage.match(/reset (?:after|in) (\d+h)?(\d+m)?(\d+s)?/i);
if (!match) return null;
let totalMs = 0;
if (match[1]) totalMs += parseInt(match[1]) * 3600 * 1000;
if (match[2]) totalMs += parseInt(match[2]) * 60 * 1000;
if (match[3]) totalMs += parseInt(match[3]) * 1000;
return totalMs || 2_000;
}
// Read google.rpc.RetryInfo.retryDelay from a Google API error JSON body and
// convert the protobuf Duration string ("30s", "1.5s", "0.500s") into ms.
// Returns null when the body is not JSON or carries no RetryInfo detail.
private parseStructuredRetryDelay(bodyText: string): number | null {
if (!bodyText.includes("RetryInfo")) return null;
try {
const parsed = JSON.parse(bodyText);
const details = parsed?.error?.details;
if (!Array.isArray(details)) return null;
for (const detail of details) {
if (
detail?.["@type"] === "type.googleapis.com/google.rpc.RetryInfo" &&
typeof detail?.retryDelay === "string"
) {
const seconds = parseFloat(detail.retryDelay.replace(/s$/i, ""));
if (Number.isFinite(seconds) && seconds >= 0) {
return Math.round(seconds * 1000) || 2_000;
}
}
}
} catch {
/* not JSON — caller falls back to prose parsing */
}
return null;
}
}
export default GeminiCLIExecutor;