Instruction file imported from mac-reichelt/tuning-coach (.github/instructions/llm-workflow-payload-limits.instructions.md). Copyright stays with the author.

GitHub Models API payload limits in workflows

Symptom

curl: (22) The requested URL returned error: 413 from https://models.inference.ai.azure.com/chat/completions even with payloads that look small in raw bytes.

Root cause

GitHub Models free tier has an ~8K-token input limit (~32KB of UTF-8 text). Above that, the endpoint hard-rejects with 413. Newline-heavy --rawfile content explodes via JSON escaping (\n), so the encoded body is ~30% bigger than the raw file. Empirical thresholds:

Pattern that works

1. Cap MAX_DIFF_BYTES at 25000 (not 60000 as initially set)
2. Send only: diff (capped), changed-files list, agent prompt file
3. Do NOT include the full repo file corpus
4. Add a belt-and-suspenders pre-curl check on wc -c < /tmp/req.json: if assembled body > 90 KB, skip the LLM call and emit a distinct SKIPPED (budget-exceeded) verdict that completes the required check without implying model approval. The example below uses APPROVE for backwards-compatibility with workflows whose downstream steps already gate on APPROVE/UPDATES_NEEDED; for new workflows prefer a dedicated SKIPPED verdict whose summary makes it explicit that the model review did not run and humans must verify manually.

REQ_BYTES=$(wc -c < /tmp/req.json)
if [ "$REQ_BYTES" -gt 90000 ]; then
  echo "::warning::Request body ${REQ_BYTES} bytes exceeds LLM budget; skipping inline review"
  jq -n --argjson n "$REQ_BYTES" '{verdict:"APPROVE",summary:("PR diff plus context (" + ($n|tostring) + " bytes) exceeds LLM budget. Inline LLM review did NOT run; please verify manually."),patches:[]}' > /tmp/verdict.json
  # ... emit outputs and exit 0
fi

Reference

• See tuning-coach's tech-writer.yml (introduced in PR #100).
• The same fix needs propagation to the existing agent-review.yml and devops-review.yml. If a dedicated security-review.yml workflow is added in the future (tracked by tuning-coach#101), apply the same payload-budget guard there as well.

Real incident

2026-04-26: rolling out the agent pipeline to 4 sister repos via greenfield bootstrap PRs (~5600 LOC each, 230 KB diffs). Every LLM review workflow returned 413 simultaneously. Bootstrap PRs were admin-merged because they're inherently too big for LLM review. Filed tuning-coach#101 to propagate the budget fix to the other review workflows.