seahi/baoyu-skills
1
0
Fork 0
mirror of https://github.com/JimLiu/baoyu-skills.git synced 2026-07-12 05:51:44 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(baoyu-image-gen): require real image_gen evidence in codex-imagegen

Browse Source 
Stop Codex from satisfying a request by copying an unrelated pre-existing
image from generated_images instead of generating a new one. Verification
now requires a PNG in this thread's generated_images dir (or an image_gen
stream item, kept as a forward-compatible signal), and the spawned-agent
instruction forbids reading or reusing history images before image_gen is
called. Removes the findCpToTarget fallback that allowed the shortcut.

Validated against a real codex exec run: image_gen leaves no stream item,
so the filesystem check is the load-bearing signal; locked in via a
condensed real-stream regression fixture.

Closes #185
This commit is contained in:
Jim Liu 宝玉
2026-06-18 14:13:43 -05:00
parent 441ca307a6
commit 51b31aa050
6 changed files with 105 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/baoyu-codex-imagegen/src/validator.ts
+10 -9
View File
@@ -3,6 +3,7 @@ import { homedir } from "node:os";
import path from "node:path";
import { GenError } from "./types.ts";
import type { ToolCall } from "./types.ts";
import { hasImageGenInvocation } from "./parser.ts";
const PNG_MAGIC = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a]);
@@ -23,15 +24,15 @@ export async function verifyImageGenWasInvoked(threadId: string | null): Promise
}
}
export function findCpToTarget(toolCalls: ToolCall[], target: string): boolean {
return toolCalls.some(
(tc) =>
tc.tool === "shell" &&
typeof tc.command === "string" &&
(tc.command.includes(target) || tc.command.includes(path.basename(target))) &&
/\b(cp|mv|cat)\b/.test(tc.command) &&
tc.command.includes("generated_images"),
);
// Real evidence that image_gen ran in THIS thread. Codex's image_gen tool does
// not surface as a stream item, so a successful run shows only reasoning/shell/
// agent_message — `dirHasImage` (a PNG in this thread's generated_images dir) is
// what proves it. The stream check is kept as a forward-compatible signal in
// case a future Codex version emits the item. The #185 shortcut (copying an
// unrelated history image, which lives under a different thread id) yields
// neither, so it is correctly rejected.
export function hasImageGenEvidence(toolCalls: ToolCall[], dirHasImage: boolean): boolean {
return dirHasImage || hasImageGenInvocation(toolCalls);
}
export async function verifyOutput(outputPath: string): Promise<{ bytes: number }> {