pjandro/OmniRoute
1
0
Fork 0
mirror of https://github.com/diegosouzapw/OmniRoute.git synced 2026-08-11 09:42:15 +03:00
Code Issues Packages Projects Releases Wiki Activity

fix(sse): extract Antigravity output-cap resolution to leaf module

Browse Source 
check:file-size froze open-sse/executors/antigravity.ts at 1528 lines;
the new per-model output-cap resolver pushed it to 1573. Move
MAX_ANTIGRAVITY_OUTPUT_TOKENS and resolveAntigravityOutputCap into a
new leaf module (antigravityOutputCap.ts) and re-export the constant
from antigravity.ts for existing test imports. No behavior change —
antigravity.ts is now 1525 lines (under the 1528 cap), all 48
executor-antigravity + antigravity-per-model-output-cap +
copilot-agent-antigravity-parity tests pass unchanged.

Co-authored-by: HouMinXi <HouMinXi@users.noreply.github.com>
This commit is contained in:
diegosouzapw
2026-08-05 19:58:42 -03:00
parent 4891c58c8c
commit fba7e7e409
2 changed files with 57 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
open-sse/executors/antigravity.ts
View File

@@ -22,7 +22,11 @@ import {
import { persistCreditBalance, getAllPersistedCreditBalances } from "@/lib/db/creditBalance";
import { setConnectionRateLimitUntil } from "@/lib/db/providers";
import { getMitmAlias } from "@/lib/db/models";
import { getExplicitModelOutputCap } from "@/lib/modelCapabilities";
import {
MAX_ANTIGRAVITY_OUTPUT_TOKENS,
resolveAntigravityOutputCap,
} from "./antigravityOutputCap.ts";
export { MAX_ANTIGRAVITY_OUTPUT_TOKENS } from "./antigravityOutputCap.ts";
import { ensureAntigravityProjectAssigned } from "../services/antigravityProjectBootstrap.ts";
import { persistDiscoveredAntigravityProjectId } from "../services/antigravityProjectPersist.ts";
import {
@@ -279,57 +283,6 @@ async function cleanModelName(model: string, modelIdOverride?: string): Promise<
return clean;
}
/**
* Fallback ceiling on `generationConfig.maxOutputTokens` for Antigravity
* Cloud Code, used when the model is unknown to the catalogue.
*
* Ports decolua/9router#779 (lukmanfauzie): VS Code GitHub Copilot Chat in
* Agent mode regularly requests 32K–65K output tokens, which the Antigravity
* backend rejects with HTTP 400 "Invalid Argument". 16384 was the ceiling
* confirmed safe at the time, via successful 200 OK runs with
* claude-sonnet-4-6 and gemini-pro-agent across both Ask and Agent modes.
*
* Both of those models are catalogue-known today, so neither one reaches this
* constant anymore: they get their own declared limit via
* `resolveAntigravityOutputCap` (65536 and 65535, respectively). The higher
* limit holds against the live upstream. A gemini-3.6-flash-high request came
* back with completion_tokens 16754 and finish_reason "stop", which exceeds
* 16384 on its own and so cannot be an artifact of thinking-token accounting.
*
* Note also that #779 was reported against Copilot Chat in Agent mode, a path
* that does not reach this executor, so 16384 arrived with that port rather
* than from a limit measured here. Beware of re-deriving it from a running
* instance: the clamp below rewrites maxOutputTokens before the request
* leaves, so a build still carrying a low constant measures its own clamp and
* reports it as an upstream ceiling.
*/
export const MAX_ANTIGRAVITY_OUTPUT_TOKENS = 16384;
/**
* The output ceiling this specific model accepts, or the conservative
* fallback above when the id is not in the catalogue.
*
* The declared limits are not uniform: most Antigravity models publish
* 65535 or 65536, but gpt-oss-120b-medium publishes 32768. A single global
* ceiling either starves the first group or lets an oversized request
* through to the second, so the number has to come from the model.
*/
function resolveAntigravityOutputCap(modelId: string | null | undefined): number {
const id = typeof modelId === "string" ? modelId.trim() : "";
if (!id) return MAX_ANTIGRAVITY_OUTPUT_TOKENS;
try {
const declared = getExplicitModelOutputCap({ provider: "antigravity", model: id });
return typeof declared === "number" && Number.isFinite(declared) && declared > 0
? declared
: MAX_ANTIGRAVITY_OUTPUT_TOKENS;
} catch {
// DB not available (build phase, transient error) -- fall through to the
// conservative fallback, the same guard cleanModelName uses above for
// its own MITM alias lookup.
return MAX_ANTIGRAVITY_OUTPUT_TOKENS;
}
}
function applyAntigravityGenerationDefaults(
request: Record<string, unknown>,
modelId?: string | null

52
open-sse/executors/antigravityOutputCap.ts Normal file
View File

@@ -0,0 +1,52 @@
import { getExplicitModelOutputCap } from "@/lib/modelCapabilities";
/**
* Fallback ceiling on `generationConfig.maxOutputTokens` for Antigravity
* Cloud Code, used when the model is unknown to the catalogue.
*
* Ports decolua/9router#779 (lukmanfauzie): VS Code GitHub Copilot Chat in
* Agent mode regularly requests 32K–65K output tokens, which the Antigravity
* backend rejects with HTTP 400 "Invalid Argument". 16384 was the ceiling
* confirmed safe at the time, via successful 200 OK runs with
* claude-sonnet-4-6 and gemini-pro-agent across both Ask and Agent modes.
*
* Both of those models are catalogue-known today, so neither one reaches this
* constant anymore: they get their own declared limit via
* `resolveAntigravityOutputCap` (65536 and 65535, respectively). The higher
* limit holds against the live upstream. A gemini-3.6-flash-high request came
* back with completion_tokens 16754 and finish_reason "stop", which exceeds
* 16384 on its own and so cannot be an artifact of thinking-token accounting.
*
* Note also that #779 was reported against Copilot Chat in Agent mode, a path
* that does not reach this executor, so 16384 arrived with that port rather
* than from a limit measured here. Beware of re-deriving it from a running
* instance: the clamp below rewrites maxOutputTokens before the request
* leaves, so a build still carrying a low constant measures its own clamp and
* reports it as an upstream ceiling.
*/
export const MAX_ANTIGRAVITY_OUTPUT_TOKENS = 16384;
/**
* The output ceiling this specific model accepts, or the conservative
* fallback above when the id is not in the catalogue.
*
* The declared limits are not uniform: most Antigravity models publish
* 65535 or 65536, but gpt-oss-120b-medium publishes 32768. A single global
* ceiling either starves the first group or lets an oversized request
* through to the second, so the number has to come from the model.
*/
export function resolveAntigravityOutputCap(modelId: string | null | undefined): number {
const id = typeof modelId === "string" ? modelId.trim() : "";
if (!id) return MAX_ANTIGRAVITY_OUTPUT_TOKENS;
try {
const declared = getExplicitModelOutputCap({ provider: "antigravity", model: id });
return typeof declared === "number" && Number.isFinite(declared) && declared > 0
? declared
: MAX_ANTIGRAVITY_OUTPUT_TOKENS;
} catch {
// DB not available (build phase, transient error) -- fall through to the
// conservative fallback, the same guard cleanModelName uses above for
// its own MITM alias lookup.
return MAX_ANTIGRAVITY_OUTPUT_TOKENS;
}
}