Skip to content

fix(openai): collapse callOpenAI messages into a single user content (Devvit HTTP-plugin 400 root-cause fix) #32

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

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

fix(openai): collapse callOpenAI messages into a single user content (Devvit HTTP-plugin 400 root-cause fix) #32

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
52 changes: 38 additions & 14 deletions src/server/index.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1271,22 +1271,46 @@ async function callOpenAI(
console.warn('[vibe-mod] callOpenAI: settings.get(openaiModel) threw — using default:', describeErr(err));
}

// Single user message containing system instructions, few-shot examples, and
// the user rule (+ optional clarification), all inline.
//
// Why a single message: Devvit's HTTP plugin reliably trips on `chat/completions`
// bodies that combine (a) ≥ ~7 KB total size, (b) multiple messages, and
// (c) JSON-escape sequences from nested `JSON.stringify` of few-shot
// `assistant` content. Direct (laptop → OpenAI) POSTs of the *identical* body
// return HTTP 200; the same body via Devvit's `fetch` returns HTTP 400
// "We could not parse the JSON body". Probe v3 confirmed transit-safe shapes:
// (b) tiny single user 121 B 200, (d) tiny + response_format 200, (e) tiny +
// gpt-5.x family params 200, (f) 6 KB ASCII single user 200, (c) 7 KB 10
// messages 400. Inlining everything as one user message keeps us on the (f)
// shape (single user content) — independently verified transit-safe at 5610 B
// and now extended to ~7 KB of structured instructions + examples.
//
// Why this preserves prompt fidelity: gpt-5.4-mini treats leading user-message
// instructions identically to a system role for the purposes of JSON-mode
// compilation; `response_format: { type: 'json_object' }` still enforces strict
// JSON output. Few-shot examples are inlined in `INPUT → OUTPUT` blocks so the
// 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);
const compositeContent = [
'=== SYSTEM INSTRUCTIONS ===',
VIBE_MOD_SYSTEM_PROMPT,
'',
'=== EXAMPLES ===',
...FEW_SHOT_EXAMPLES.map(
(ex, i) => `--- Example ${i + 1} ---\nINPUT:\n${ex.user}\n\nOUTPUT:\n${JSON.stringify(ex.assistant)}`,
),
'',
'=== TASK ===',
`INPUT:\n${userRule.slice(0, 1000)}`,
...(clarif ? ['', `CLARIFICATION:\n${clarif}`] : []),
'',
'OUTPUT:',
].join('\n\n');
const messages: Array<{ role: 'system' | 'user' | 'assistant'; content: string }> = [
{ role: 'system', content: VIBE_MOD_SYSTEM_PROMPT },
{ role: 'user', content: compositeContent },
];
for (const ex of FEW_SHOT_EXAMPLES) {
messages.push({ role: 'user', content: ex.user });
messages.push({ role: 'assistant', content: JSON.stringify(ex.assistant) });
}
// Original rule (capped — the form layer doesn't bound length; keep the prompt
// budget predictable and limit the prompt-injection surface).
messages.push({ role: 'user', content: userRule.slice(0, 1000) });
// Clarification answer as a separate turn (audit FIND-11: no string concat into
// the original user content), also length-capped (gap-analysis SEC-03).
const clarif = clarificationAnswer?.trim().slice(0, 500);
if (clarif) {
messages.push({ role: 'user', content: `Clarification: ${clarif}` });
}

// Build the body, then escape every non-ASCII character as a JSON \uXXXX
// sequence. OpenAI returned HTTP 400 "We could not parse the JSON body of
Expand Down