> ## Documentation Index > Fetch the complete documentation index at: https://docs.prysm1.com/llms.txt > Use this file to discover all available pages before exploring further. # Code > Generate code with a self-correcting, multi-model loop — a coder writes it, a separate critic reviews it, the coder repairs against the feedback — and get a PrysmProof v2. **POST** `https://api.prysm1.com/v2/code` · Requires authentication **Code Mode** writes code, reviews it with a *separate* critic model, and repairs against that feedback until the review passes or a budget/iteration cap is hit ([Reflexion](https://arxiv.org/abs/2303.11366)-style). In `depth` policy several diverse coders run in parallel and the best candidate is kept. The response carries the files, the final review verdict, and a [PrysmProof v2](/concepts/prysmproof) that binds the task to the code it produced. You pick the objective with a **policy**; PRYSM picks the coder and the critic (or you force them). See [How Code Mode works](/concepts/code-mode) for the full model. This endpoint lives under **`/v2`**, not `/v1`. The SDKs target it automatically with `client.code(...)`. **PRYSM never executes the generated code** — the default critic is a pure LLM reviewer. ## Authorization Your secret key as a bearer token: `Bearer prysm_sk_...` ## Body The coding task in natural language, e.g. *"Write a thread-safe LRU cache in Python with tests"*. Target language (`python`, `typescript`, `rust`, …). Auto-detected from the task if omitted. The objective dial: `efficiency` (one cheap coder), `depth` (cross several diverse coders in parallel and keep the best), or `balanced`. Force the model that writes the code. Ignored if it isn't a known catalog model. Force the critic model that reviews the code. Ignored if it isn't a known catalog model. Total iterations: the first generation plus up to `max_iters - 1` repairs. Set `false` for a single shot — one coder, one draft, no critic loop. Maximum tokens per underlying model call. Sampling temperature passed to the underlying models. Lower is steadier for code. A soft budget hint, in USD. The repair loop stops early once the estimated spend approaches this cap. Include the per-stage execution trace in `prysm.stages`. Set `false` for a leaner response. ## Response Unique code-run id, e.g. `prysm-a1b2c3d4`. Always `code`. Unix timestamp (seconds). The task that was solved. The target language (detected or forced). The policy that ran: `efficiency`, `balanced`, or `depth`. Plain-English explanation of how the run was carried out. Whether the final review passed. How many generate→review cycles ran. The generated files. Relative file path, e.g. `main.py`. The file's language. The full file contents. Length of `content` in bytes. The final critic verdict, or `null` when `review` was `false`. Whether the critic passed the code. The critic's score, `0`–`1`. Concrete issues the critic raised (empty when clean). The critic model that produced the verdict. Aggregate token usage across every model call. Total input tokens. Total output tokens. Sum of the two. The Code Mode extension block. The policy that ran. The target language. Whether the final review passed. The final review score, `0`–`1`. Blended confidence across the loop, `0`–`1`. How many generate→review cycles ran. Every coder that wrote a candidate. The critic model. Every model that contributed (coders + critic). Wall-clock latency in milliseconds. Total cost across all model calls, USD. Always `true` — v2 cost is estimated from text length (\~4 chars/token). A verifiable [PrysmProof v2](/concepts/prysmproof) binding the task to the code. Verify it later via [`GET /v1/proof/{request_id}`](/api-reference/proof). The id to verify against. ISO-8601 UTC timestamp. SHA-256 over the task, files, and stages, e.g. `sha256:a1b2c3d4e5f6...`. The policy that ran. The target language. Whether the final review passed. Every model that contributed. Blended confidence across the loop. Always `true`. Per-stage execution trace (present when `include_trace` is `true`). Stages run as `plan` → `generate:0` → `review:0` → `generate:1` → `review:1` … Each stage has a `name`, a `results` array with one entry per model call (model id, ok, token counts, latency, cost, confidence, role — the raw text is omitted), and a `detail` object. A skipped critic (`review: false`) leaves `review:0` with an empty `results` array. ## Errors | Status | `error` | Meaning | | ------ | ----------------------- | ------------------------------------------------------ | | `400` | `no_task` | `task` was empty or whitespace. | | `401` | — | Missing or invalid API key. | | `502` | `no_code_produced` | Keys are configured but no model produced usable code. | | `502` | `code_error` | The engine raised while generating or reviewing. | | `503` | `no_provider_available` | No provider API keys are configured on the server. | ```python Python theme={null} from prysm import Prysm client = Prysm() r = client.code( "Write a Python function to add two numbers", policy="balanced", ) for f in r["files"]: print(f["path"]); print(f["content"]) print(r["passed"], r["prysm"]["proof"]["proof_hash"]) ``` ```typescript Node theme={null} import { Prysm } from "@prysmai/sdk"; const client = new Prysm(); const r = await client.code( "Write a Python function to add two numbers", { policy: "balanced" }, ); for (const f of r.files) { console.log(f.path); console.log(f.content); } console.log(r.passed, r.prysm.proof.proof_hash); ``` ```bash cURL theme={null} curl https://api.prysm1.com/v2/code \ -H "Authorization: Bearer $PRYSM_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "task": "Write a Python function to add two numbers", "policy": "balanced" }' ``` ```json 200 theme={null} { "id": "prysm-a1b2c3d4", "object": "code", "created": 1767312000, "task": "Write a Python function to add two numbers", "language": "python", "policy": "balanced", "reason": "balanced policy: one mid-tier coder, reviewed and repaired by a reasoning critic.", "passed": true, "iterations": 1, "files": [ { "path": "main.py", "language": "python", "content": "def add(a, b):\n \"\"\"Return the sum of two numbers.\"\"\"\n return a + b\n", "bytes": 71 } ], "review": { "passed": true, "score": 0.92, "issues": [], "reviewer": "deepseek-v4-pro" }, "usage": { "prompt_tokens": 86, "completion_tokens": 120, "total_tokens": 206 }, "prysm": { "code": { "policy": "balanced", "language": "python", "passed": true, "score": 0.92, "confidence": 0.9, "iterations": 1, "coder_models": ["qwen-2.5-72b"], "reviewer_model": "deepseek-v4-pro", "models_used": ["qwen-2.5-72b", "deepseek-v4-pro"], "latency_ms": 3180 }, "cost": { "total_usd": 0.000412, "estimated": true }, "proof": { "request_id": "a1b2c3d4-...-...", "timestamp": "2026-06-03T00:00:00+00:00", "proof_hash": "sha256:7f3a9c2b1e4d8a06", "policy": "balanced", "language": "python", "passed": true, "models_used": ["qwen-2.5-72b", "deepseek-v4-pro"], "confidence": 0.9, "verifiable": true }, "stages": [ { "name": "plan", "results": [], "detail": { "language": "python", "policy": "balanced", "coder_models": ["qwen-2.5-72b"], "reviewer_model": "deepseek-v4-pro" } }, { "name": "generate:0", "results": [ { "model": "qwen-2.5-72b", "ok": true, "in_tokens": 42, "out_tokens": 78, "latency_ms": 1620, "cost_usd": 0.00021, "confidence": 0.8, "role": "coder" } ], "detail": { "iteration": 0 } }, { "name": "review:0", "results": [ { "model": "deepseek-v4-pro", "ok": true, "in_tokens": 44, "out_tokens": 42, "latency_ms": 1560, "cost_usd": 0.0002, "confidence": 0.92, "role": "reviewer" } ], "detail": { "passed": true, "score": 0.92 } } ] } } ```