GLM-5.3

{
  "model": "glm-5.3",
  "input": {
    "model": "glm-5.3",
    "max_tokens": 4096,
    "messages": [
      {
        "role": "user",
        "content": "Review this architecture proposal and return the risks, tradeoffs, and recommended next steps."
      }
    ],
    "temperature": 1,
    "top_p": 1,
    "top_k": 0
  },
  "output": "A codebase you don't know is a risk, not a hurdle. The plan should burn down that risk in order: understand, run, trace, spike, then make the change small enough that \"production-ready\" is a property you can demonstrate, not a hope.\n\nHere are the first five steps, with the evidence you collect at each.\n\n**Step 1 — Get it building and the baseline green (half day)**\nRead the README, the manifest(s), the CI config, and any architecture docs. Boot the app locally, run the full test suite, run the linter/typechecker.\n*Evidence:* The exact commands that reproduce a clean build and a passing test suite from a fresh checkout (recorded output, not memory of it). A one-page map you wrote yourself: where the entry points are, the layers, the build/test toolchain, and the main risks to your feature (e.g. \"this touches cron jobs and payments\").\n\n**Step 2 — Trace the smallest end-to-end slice that resembles your feature (half day)**\nDon't start with your feature. Pick a neighboring one that already works and follow it from entry to persistence. Note the conventions along the way: how errors surface, how config is injected, how logging is done, how tests are written.\n*Evidence:* An annotated trace — file → function → what it does — for that slice. A written list of \"conventions I must follow\" (not intuitions; things you observed in real code). A named shortlist of the 3–5 files you will actually touch. If you can't produce this trace, you're not ready to write code.\n\n**Step 3 — Spike the core path (day 2)**\nBuild the ugly version: hardcoded values, no error handling, no tests. The purpose is to confirm the path you traced in step 2 is real and to surface what you didn't know you didn't know.\n*Evidence:* A working spike demonstrating the feature's central data path, alongside a list of every assumption the spike broke and what you corrected. That correction list is the most valuable document in this whole plan.\n\n**Step 4 — Write the contract before the code (half of day 2 / day 3)**\nOnce the spike proves the path, pin down what production needs: the inputs/outputs, the error cases, where it sits in the conventions from step 2. Then write the tests — they'll be red, but they're the specification.\n*Evidence:* A one-to-two-page design doc, an agreed interface/API shape (with the team if there is one), and a red test suite that encodes intended behavior. If you can't write the contract without consulting the code, you haven't finished step 2.\n\n**Step 5 — Implement in small, verified increments (days 3–4)**\nReplace the spike with the real thing in small commits, each one keeping the suite green, leaning on the existing patterns. Run lint/tests/typecheck per commit — CI, not just locally — and exercise the actual path against a real instance (staging or a local environment that isn't stubbed).\n*Evidence:* A branch with progressive commits, each green in CI; coverage on the new code; something that proves it works against reality (a test result, a log trace, a screenshot); and a review by at least one person who knows the codebase. The review counts as evidence — an unfamiliar codebase has tribal knowledge you cannot extract from the files alone.\n\nSteps 6+ would be the things that actually make it \"shipped\": a migration plan and its rollback, feature flagging, observability, the release and post-release verification. But the first five get you to a reviewed, green, working slice in staging — which is the point at which you can say \"this will work in production\" with evidence behind it, instead of a guess.",
  "metrics": {
    "input_tokens": 134,
    "output_tokens": 2354,
    "generated_in_seconds": 42.7
  },
  "example": true
}