drug-discovery-prompts/upstream/mims-harvard-ToolUniverse/skills/devtu-code-optimization/references/code-patterns.md

title, task, lineage_type, upstream_source, upstream_sha, imported_at, prompt_class, upstream_changes, author, validated

title	task	lineage_type	upstream_source	upstream_sha	imported_at	prompt_class	upstream_changes	author	validated
Code Patterns Reference		import	https://github.com/mims-harvard/ToolUniverse/blob/3038dcbe/skills/devtu-code-optimization/references/code-patterns.md	3038dcbe	2026-06-30	prompt	accepted	upstream	false

Code Patterns Reference

Reusable implementation patterns for ToolUniverse tool development.

Schema Patterns

return_schema with oneOf (required)

{
  "return_schema": {
    "oneOf": [
      {
        "type": "object",
        "properties": {
          "data": {"type": "object"},
          "metadata": {"type": "object"}
        }
      },
      {
        "type": "object",
        "properties": {
          "error": {"type": "string"}
        }
      }
    ]
  }
}

Nullable fields

{"type": ["array", "null"]}
{"type": ["string", "null"]}

API Call Patterns

Client-Side Filter (when API ignores params)

results = api_call(base_params_only)
if interaction_types:
    results = [r for r in results if r.get("type") in interaction_types]
if sources:
    results = [r for r in results if r.get("source") in sources]

Fallback Lookup

precise = api_call(geneSymbol=gene_symbol)
if not precise:
    precise = api_call(name=gene_symbol)

Client-Side Pagination (when API ignores size/page)

all_items = api_call()
start = page * size
return all_items[start:start + size]

PostgREST Join

url = f"{base}/recommendation?select=*,drug(name)&genesymbol={_postgrest_eq(gene)}"

Output Patterns

Normalization Disclosure

_norm_parts = []
if original != normalized:
    _norm_parts.append(f"'{original}' → '{normalized}' (reason)")
if _norm_parts:
    result["normalization_note"] = "Auto-normalized: " + "; ".join(_norm_parts)

Truncation at Top Level

response = {"status": "success", "data": data[:limit]}
if len(data) > limit:
    response["truncated"] = True
    response["truncation_note"] = (
        f"Returning {limit} of {len(data)}. "
        f"Pass max_results={len(data)} for full data."
    )

No-Data vs Bad-Query

if count == 0 and query:
    result["hint"] = f"No results for '{query}'. Try a broader term or check spelling."
elif count == 0:
    result["hint"] = "No data available for this entity."

Hosted Model-API Tools (NVIDIA NIM-style)

Async poll host = the invocation host

Poll a 202 job-status on the SAME gateway you POSTed to. NVCF biology NIMs invoke and poll on health.api.nvidia.com; integrate.api.nvidia.com serves only the OpenAI-compatible LLM endpoints and has no /v1/status route.

host = urlparse(self.base_url).netloc          # e.g. health.api.nvidia.com
poll_url = f"https://{host}/v1/status/{req_id}"

Route-existence probe (find/verify hosted endpoints)

Plain-text 404 page not found = route does NOT exist; a structured {"status":404,...} (or 400/422/200) = route exists. Use the live API to confirm a model is hosted and to find the right slug before wrapping it.

Unwrap JSON envelopes around the "raw" payload

Some endpoints return {"pdbs": ["...ATOM..."]} even when response_type is pdb. Unwrap to the inner value so the field matches the schema (real PDB, not a JSON blob).

HTTP 200 with an inner failure

A 200 can carry {"status": "failed", ...} (e.g. DiffDock with an unreadable ligand). Surface it as an error — but only on explicit failed/error/errored; an inner status:"success" must stay a success (don't over-match).

404 "not found for account" ≠ wrong path

A gated/unprovisioned model returns a 404 whose body says "Not found for account". Report "model not available for your account" rather than "endpoint not found".

Retry a longer poll window before declaring "broken"

A heavy async job can return 504 / nvcf-status: errored simply because NVCF-POLL-SECONDS was shorter than its runtime. Only a persistent 400 DEGRADED/error across retries is a real outage. 5xx bodies are often empty — surface nvcf-status / nvcf-reqid from the headers instead.

Model-variant selection via templated endpoint

Expose multiple hosted sizes through one tool: a {placeholder} in the endpoint + fields.path_params default, filled from the request arg (sanitized slug) and stripped from the request body.

# endpoint "arc/{model}/generate", path_params {"model": "evo2-40b"}
value = args.get(key) or default
if not re.fullmatch(r"[A-Za-z0-9._-]+", str(value)):  # no path injection
    value = default
endpoint = endpoint.replace("{" + key + "}", value)

R-subprocess Tools

Run a script file, not Rscript -e <string>

Rscript -e collapses one backslash level before R parses it, so a regex literal like sub("\\..*", ...) becomes sub("\..*", ...) and R aborts with '\.' is an unrecognized escape. Write the script to a temp .R file and run Rscript <file> (parsed verbatim); always remove the temp file, incl. on timeout.

tmp = tempfile.NamedTemporaryFile(mode="w", suffix=".R", delete=False)
try:
    tmp.write(r_script); tmp.close()
    return subprocess.run(["Rscript", tmp.name], capture_output=True, text=True, timeout=t)
finally:
    try: os.unlink(tmp.name)
    except OSError: pass

try/except Indentation (Critical)

# CORRECT
try:
    resp = requests.get(url)
    data = resp.json()
except Exception as e:
    return {"status": "error", "error": str(e)}

# WRONG — SyntaxError
try:
    resp = requests.get(url)
if resp.ok:        # ← same indent as try: → OUTSIDE try block
    data = resp.json()
except Exception:  # ← Python: "try without except"
    pass

Every try: must have except: at the exact same indentation level.

Multi-Word Search Hint

if result["count"] == 0 and name_q and " " in str(name_q):
    first_word = str(name_q).split()[0]
    result["multi_word_hint"] = (
        f"Search may not match multi-word phrases like '{name_q}'. "
        f"Try a single keyword: name='{first_word}'."
    )