Skip to content
This repository was archived by the owner on Jun 19, 2026. It is now read-only.

fix(openai): Uint8Array body + shrink few-shot to 1 example (Devvit 400 round 5) #36

Merged
ComBba merged 1 commit into from
May 13, 2026
Merged

fix(openai): Uint8Array body + shrink few-shot to 1 example (Devvit 400 round 5) #36

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
51 changes: 35 additions & 16 deletions src/server/index.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1293,27 +1293,33 @@ async function callOpenAI(
// model still learns the rule-compilation pattern. Length caps preserved
// (rule ≤ 1000, clarification ≤ 500) for prompt-injection surface control.
const clarif = clarificationAnswer?.trim().slice(0, 500);
// Flatten to a single line of ASCII without `\n` so the wire body has zero
// newline-escape sequences. PR #34 dropped reasoning_effort + verbosity but
// production v0.0.35 still hit 400; the only remaining variable separating
// our body from probe(f) (the 6 KB single-user 200 OK shape) is escape-char
// density (`\n` and `\"`). Collapsing whitespace eliminates the `\n` family
// entirely. The few-shot OUTPUT examples still produce `\"` from the outer
// JSON.stringify wrapping our content string -- if 400 persists after this,
// we'll need to express few-shot output without quoted keys.
// Round-5 strategy: ship a minimal-shape body that exactly matches probe(f)
// (the only production-verified 200-OK shape at ~6 KB):
// - single user message
// - no `\n` in content (use ' ' between sections)
// - few-shot truncated to 1 example so total body fits well under 5610 B
// (probe(f)'s exact size). PR #32-#35 all kept 4 examples; 400 persisted.
// - {model, messages, max_completion_tokens, response_format} only
// Plus: send the body as a Uint8Array rather than a string, so Devvit's
// HTTP plugin transports raw bytes without re-encoding the JSON wrapper.
const collapse = (s: string) => s.replace(/\s+/g, ' ').trim();
const exampleLine = (ex: (typeof FEW_SHOT_EXAMPLES)[number], i: number) =>
`EXAMPLE ${i + 1} INPUT: ${collapse(ex.user)} OUTPUT: ${JSON.stringify(ex.assistant)}`;
const SHORT_PROMPT_CHARS = 3500; // hard cap on system prompt; rest of body stays small
const systemShort = collapse(VIBE_MOD_SYSTEM_PROMPT).slice(0, SHORT_PROMPT_CHARS);

@gemini-code-assist gemini-code-assist Bot May 13, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The silent truncation of the system prompt at 3500 characters is risky. If the prompt grows in the future, this could cut off critical instructions (such as the clarification mode logic or the JSON schema definition) at the end of the string. This would lead to unpredictable LLM behavior or parsing errors. Consider adding a check to log a warning or throw an error if the collapsed prompt exceeds this limit, rather than silently slicing it.

const firstExample = FEW_SHOT_EXAMPLES[0];
const exampleText = firstExample
? `EXAMPLE INPUT: ${collapse(firstExample.user)} EXAMPLE OUTPUT: ${JSON.stringify(firstExample.assistant)}`
: '';
const compositeContent = [
'SYSTEM INSTRUCTIONS:',
collapse(VIBE_MOD_SYSTEM_PROMPT),
'EXAMPLES:',
...FEW_SHOT_EXAMPLES.map(exampleLine),
systemShort,
exampleText,
'TASK INPUT:',
collapse(userRule.slice(0, 1000)),
...(clarif ? ['TASK CLARIFICATION:', collapse(clarif)] : []),
'OUTPUT (strict JSON only, no prose):',
].join(' ');
]
.filter(Boolean)
.join(' ');
const messages: Array<{ role: 'system' | 'user' | 'assistant'; content: string }> = [
{ role: 'user', content: compositeContent },
];
Expand Down Expand Up @@ -1345,10 +1351,23 @@ async function callOpenAI(
// body as pure 7-bit ASCII bytes, sidestepping the UTF-8-on-the-wire bug.
const asciiSafeBody = rawBody.replace(/[€-￿]/g, (c) => '\\u' + c.charCodeAt(0).toString(16).padStart(4, '0'));

// Round 5: send body as Uint8Array with explicit Content-Length. String
// bodies travel through Devvit's HTTP plugin as a JS string that the plugin
// re-encodes to UTF-8 before the wire write; PR #32-#35 history shows
// large stringified-JSON bodies corrupt somewhere in that re-encode path.
// Uint8Array skips it: bytes are already final, plugin just streams them.
// Body is already pure ASCII so TextEncoder().encode produces exactly
// asciiSafeBody.length bytes (one byte per char).
const bodyBytes = new TextEncoder().encode(asciiSafeBody);
console.log('[vibe-mod] callOpenAI: body bytes =', bodyBytes.byteLength);
const resp = await fetch('https://api.openai.com/v1/chat/completions', {
method: 'POST',
headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${apiKey}` },
body: asciiSafeBody,
headers: {
'Content-Type': 'application/json',
'Content-Length': String(bodyBytes.byteLength),

@gemini-code-assist gemini-code-assist Bot May 13, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Manually setting the Content-Length header is generally discouraged when using fetch. In many standard-compliant environments (like browsers or Node.js with undici), Content-Length is a forbidden header name and attempting to set it manually will throw a TypeError. While this may be a necessary workaround for the current Devvit HTTP plugin issues, it could cause failures if the environment becomes more standard-compliant or when running in different execution contexts (e.g., unit tests). Since bodyBytes is a Uint8Array, most fetch implementations will automatically calculate and set the correct Content-Length.

Authorization: `Bearer ${apiKey}`,
},
body: bodyBytes,
});

if (!resp.ok) {
Expand Down