llmwrap

Usage guide for llmwrap with eight distinct integration patterns. Each pattern appears as a pair: wrap_llm_call (decorator) and wrap_llm_line (inline) implement the same behavior; only prompt binding differs. Additional permutations (OpenRouter-only snippets, naming variants, LangChain-style flows) remain in tests/testing_different_interfaces.py.

This document intentionally covers API usage only. Internal algorithm details are not disclosed.

Table of Contents

• Install
• Public APIs
• Config Fields
  • Workflow graph example (linear chain)
• Distinct example pairs
  • 1. Chat Completions
  • 2. Responses API
  • 3. Multipart assistant content
  • 4. Tool calls with passthrough
  • 5. Nested wrapped calls
  • 6. Custom response extractor
  • 7. Non-reconstructable model output
  • 8. Tool agent with wrapped LLM tools
• Console output
• Notes
• License

Install

pip install arbis-llmwrap

Public APIs

from llmwrap import wrap_llm_call, wrap_llm_line, openai_sdk_result_text

Config Fields

Core Mental Model

For each wrapped LLM call, you are describing where this call sits in your orchestration graph.

• workflow_group_name = top-most business/process container
• Workflow_name = a workflow inside that group
• agent_name = the exact executing step/agent inside that workflow
• Run_Id = one end-to-end execution id tying all related calls together

So for one user request that traverses many agents, all calls share the same Run_Id, but agent_name changes per step.

Golden rule: one Run_Id = one user query/request. Reuse it across all hops for that query; allocate a new one for the next query.

Shared Fields (wrap_llm_call and wrap_llm_line)

company_name: str (Required)

What it is: Organization identifier.
Why it matters: High-level tenant/org partition for ingest and analytics.
How to fill: Stable canonical company name, e.g. "ArbisAI".
Best practice: Keep consistent casing/spelling across all services.

project_name: str (Required)

What it is: Product/application name inside the company.
Why it matters: Separates telemetry for different apps under same company.
How to fill: Stable app id, e.g. "Arbis-Decorator" or "support-copilot".
Best practice: Do not use random env-specific names unless intentional.

agent_name: str (Required)

What it is: The current executing agent/step label for this call.
Why it matters: This is the node identifier in your execution chain.
How to fill: Step-specific name, e.g. "Agent-1", "Retriever", "Writer".
Best practice: Human-readable and unique enough within workflow.

Run_Id: str | None (Optional, strongly recommended)

What it is: Correlation id for one complete run/user query/session.
Why it matters: Lets you reconstruct the full path across agents/workflows.
How to fill: Generate once at run start (uuid4) and reuse on every call in that run.
Example: All A1 -> A2 -> B3 -> B4 calls carry same Run_Id.

Workflow_name: str | None (Optional, recommended)

What it is: Workflow containing this agent.
Why it matters: Groups agent nodes into workflow-level segments.
How to fill: Use logical workflow id, e.g. "A", "B", "triage_flow".
Best practice: Stable naming so workflow-level analytics remain clean.

workflow_group_name: str | None (Optional, recommended)

What it is: Top-level grouping for one or more workflows.
Why it matters: Organizes related workflows under a single umbrella.
How to fill: Product/process group, e.g. "Workflow Group 1" or "customer_support".
Best practice: This should typically remain constant while Workflow_name may change within the run.

agent_parent: list[str] | None (Optional)

What it is: Upstream agent lineage for current agent call.
Why it matters: Captures agent-to-agent dependency edges.
How to fill:

• Root/first step: None
• Next step consuming Agent-1 output: ["Agent-1"]
• For deeper nesting, include lineage order if you track ancestry.

Best practice: At minimum include immediate parent when there is one.

Workflow_parent_name: list[str] | None (Optional)

What it is: Parent workflow lineage when transitioning across workflows.
Why it matters: Captures workflow-level dependency (A -> B, etc.).
How to fill:

• Within same workflow: often None
• First step in child workflow B after A: ["A"]

Best practice: Set it at workflow boundary transitions for clear graph reconstruction.

metadata: dict | None (Optional)

What it is: Arbitrary JSON-serializable context for user/request/business metadata.
Why it matters: Adds business observability and audit dimensions.
How to fill: Include only safe, needed keys. Typical:

• user_id
• role
• tenant_id
• request_id
• session_id
• channel

Best practice:

• Avoid secrets/PII unless policy permits
• Keep schema consistent across calls
• You can enrich this per step if needed, but keep core keys stable per run

secret_key: str (Required)

What it is: Wrapper authentication/authorization key used by ingest path.
Why it matters: Required for secure wrapper operations.
How to fill: From env var (WRAP_SECRET_KEY), never hardcoded.
Best practice: Rotate regularly and keep out of logs.

max_tries: int = 1 (Optional)

What it is: Retry budget for wrapped execution pipeline.
Why it matters: Improves resilience against transient errors.
How to fill: Integer >= 1; often 1 to 3.
Best practice: Use higher values only where retries are safe and expected.

response_extractor: Callable[[Any], str] | None (Optional)

What it is: Custom function that extracts answer text from raw model output.
Why it matters: Needed when return shape is custom/non-standard.
How to fill: Provide function returning str, e.g. lambda obj: obj["raw_text"].
Best practice: Pair with merge/writeback settings if preserving object shape.

prompt_json_pointer: str | None (Optional)

What it is: RFC 6901 pointer to prompt field inside structured prompt payload.
Why it matters: Wrap only one field instead of whole JSON payload.
How to fill: Example "/messages/0/content" or "/query".
Best practice: Validate pointer path exists in your payload schema.

passthrough_when: Callable[[Any], bool] | None (Optional)

What it is: Predicate to bypass parse/ingest flow and return raw output unchanged.
Why it matters: Useful for tool-calls or intermediate SDK objects.
How to fill: Function returning True when output should pass through.
Example: detect tool_calls and skip final merge logic.

return_merger: Callable[[Any, str], Any] | None (Optional)

What it is: Hook to combine original output object with extracted answer text.
Why it matters: Gives full control over final response shape.
How to fill: (base_output, answer_text) -> desired_output.
Best practice: Use when default merge behavior does not match your API contract.

response_answer_json_pointer: str | None (Optional)

What it is: Explicit pointer for where extracted answer should be written back.
Why it matters: Deterministic answer placement in complex return objects.
How to fill: RFC 6901 path like "/choices/0/message/content".
Best practice: Prefer explicit pointer in non-standard/ambiguous structures.

wrap_llm_call-Specific Field

prompt_arg: str = "prompt" (Optional param with default, functionally required correctness)

What it is: Name of decorated function argument containing prompt payload.
Why it matters: Wrapper must know which argument to wrap.
How to fill: Match actual function signature, e.g. "messages", "question".
Best practice: Always set explicitly if your prompt arg is not named prompt.

wrap_llm_line-Specific Fields

llm_call: Callable[[Any], Any] (Required)

What it is: Callable that executes model request using wrapped prompt payload.
Why it matters: This is the actual invocation path for line-level API.
How to fill: lambda prompt: client.chat.completions.create(...) etc.
Best practice: Keep deterministic and side-effect-free except model call.

prompt: Any (Required)

What it is: Raw prompt input to be wrapped.
Why it matters: Source content entering wrapper pipeline.
How to fill:

• plain string prompt, or
• dict/JSON string when using prompt_json_pointer

Best practice: Ensure shape matches what llm_call expects after wrapping.

Practical Fill Pattern (for your hierarchy)

For one incoming user query:

1. Generate run_id once.
2. Keep workflow_group_name constant for entire orchestration.
3. Set Workflow_name based on current workflow segment.
4. Set agent_name for current step.
5. Set lineage:
   • first agent: agent_parent=None, Workflow_parent_name=None
   • downstream in same workflow: agent_parent=[prev_agent]
   • first step in new workflow B after A: Workflow_parent_name=["A"]
6. Attach metadata with user context (user_id, role, etc.).

Workflow graph example (linear chain)

This matches the topology and field usage in tests/test_linear_workflow_group_chain.py: four agents, two workflows inside one workflow group, with prompt text handed off along the chain.

Graph (who talks to whom)

Data flow: Agent-1 → Agent-2 → Agent-3 → Agent-4.

workflow_group_name = "Workflow Group 1"     ← same on every call

Workflow_name = "A"
  Agent-1   (root: no parents)
    ↓
  Agent-2   (agent_parent = ["Agent-1"])

Workflow_name = "B"   (starts after A’s last agent)
  Agent-3   (agent_parent = ["Agent-2"], Workflow_parent_name = ["A"])
    ↓
  Agent-4   (agent_parent = ["Agent-3"])

How each field builds the graph

• workflow_group_name — Set to "Workflow Group 1" (or your real group id) on all steps so ingest knows every call belongs to the same product/process bucket.
• Workflow_name — "A" for Agent-1 and Agent-2; "B" for Agent-3 and Agent-4. This splits the run into two workflow segments under that group.
• agent_name — The current node: "Agent-1" … "Agent-4". Must be distinct per step so each hop is identifiable.
• Run_Id — One id reused on all four wrap_llm_line calls so analytics can stitch them into a single user/query trace (one session = one Run_Id). In this repo, the id is allocated from the VeryTrace run-id API on WRAP_ARBIS_BASE_URL and then reused for every agent hop.
• agent_parent — Immediate upstream agent name(s). None only for the first agent. Agent-2 points at Agent-1; Agent-3 at Agent-2; Agent-4 at Agent-3. That encodes the agent-level edge list.
• Workflow_parent_name — Workflow-level lineage when you enter a new workflow that continues after another. Here workflow B follows A, so only Agent-3 (the first step in B) sets Workflow_parent_name=["A"]. Agent-4 stays in B with no new workflow transition, so it uses None (same as agents fully inside A).

Optional metadata — Use the same dict on each call for fields like user_id and role so every step in the run carries the same user context; you can add step-specific keys if needed.

Run_Id API (session id allocation)

In tests/test_linear_workflow_group_chain.py, the session id comes from the run-id API before the first agent call:

• Endpoint: POST {WRAP_ARBIS_BASE_URL}/api/unique-run-id
• Request JSON: {"secret_key": WRAP_SECRET_KEY, "company_name": WRAP_COMPANY_NAME}
• Response JSON: {"run_id": "..."}

Then that run_id is passed as Run_Id to Agent-1, Agent-2, Agent-3, and Agent-4. This is what links all hops into one session in the graph.

Step-by-step summary

Step	agent_name	Workflow_name	workflow_group_name	agent_parent	Workflow_parent_name
1	Agent-1	A	Workflow Group 1	None	None
2	Agent-2	A	Workflow Group 1	["Agent-1"]	None
3	Agent-3	B	Workflow Group 1	["Agent-2"]	["A"]
4	Agent-4	B	Workflow Group 1	["Agent-3"]	None

Code shape (prompt strings omitted—see the test file for the full handoff text):

import requests

from llmwrap import openai_sdk_result_text, wrap_llm_line

WORKFLOW_GROUP = "Workflow Group 1"
WORKFLOW_A = "A"
WORKFLOW_B = "B"

# One Run_Id for the whole chain (one session id for all agent hops).
def allocate_run_id_from_api() -> str:
    base = CFG.wrap_arbis_base_url.rstrip("/")
    resp = requests.post(
        f"{base}/api/unique-run-id",
        json={
            "secret_key": CFG.key,
            "company_name": CFG.company,
        },
        timeout=30,
    )
    resp.raise_for_status()
    return str(resp.json()["run_id"])


run_id = allocate_run_id_from_api()

METADATA = {"user_id": "user-42", "role": "analyst"}  # optional; same dict every hop


def run_agent(*, client, model, agent_name, workflow_name, agent_parent, workflow_parent_name, prompt):
    return wrap_llm_line(
        llm_call=lambda p: client.chat.completions.create(
            model=model,
            messages=[{"role": "user", "content": p}],
            temperature=0,
        ),
        prompt=prompt,
        company_name=CFG.company,
        project_name=CFG.project,
        agent_name=agent_name,
        Run_Id=run_id,
        Workflow_name=workflow_name,
        workflow_group_name=WORKFLOW_GROUP,
        agent_parent=agent_parent,
        Workflow_parent_name=workflow_parent_name,
        metadata=METADATA,
        secret_key=CFG.key,
        max_tries=3,
    )


# Agent-1: root of the agent graph, workflow A, no workflow parent.
out1 = run_agent(
    client=openai_client,
    model=CFG.model,
    agent_name="Agent-1",
    workflow_name=WORKFLOW_A,
    agent_parent=None,
    workflow_parent_name=None,
    prompt='...',  # e.g. ask for a single marked line — see test_linear_workflow_group_chain.py
)
text1 = openai_sdk_result_text(out1).strip()

# Agent-2: still workflow A; parent agent is Agent-1.
out2 = run_agent(
    client=openai_client,
    model=CFG.model,
    agent_name="Agent-2",
    workflow_name=WORKFLOW_A,
    agent_parent=["Agent-1"],
    workflow_parent_name=None,
    prompt=f"The previous agent output is:\n{text1}\n...",
)
text2 = openai_sdk_result_text(out2).strip()

# Agent-3: first step in workflow B; consumes Agent-2; workflow B follows A.
out3 = run_agent(
    client=openai_client,
    model=CFG.model,
    agent_name="Agent-3",
    workflow_name=WORKFLOW_B,
    agent_parent=["Agent-2"],
    workflow_parent_name=[WORKFLOW_A],
    prompt=f"Previous workflow A ended with:\n{text2}\n...",
)
text3 = openai_sdk_result_text(out3).strip()

# Agent-4: still workflow B; parent agent Agent-3 only.
out4 = run_agent(
    client=openai_client,
    model=CFG.model,
    agent_name="Agent-4",
    workflow_name=WORKFLOW_B,
    agent_parent=["Agent-3"],
    workflow_parent_name=None,
    prompt=f"Previous agent in workflow B produced:\n{text3}\n...",
)

Distinct example pairs

Convention. Snippets assume CFG, openai_client, RUN_ID, WORKFLOW_NAME, WORKFLOW_GROUP_NAME, AGENT_PARENT, WORKFLOW_PARENT_NAME, METADATA, and (where needed) TOOLS / MESSAGES exist in your app. The same tracking kwargs appear in every call so ingest can correlate steps.

How to read this section. Each numbered scenario is one behavioral class for the wrapper. For each class we show the same behavior twice: wrap_llm_call binds the prompt via a decorated function argument; wrap_llm_line passes llm_call and prompt at the call site. Ingest and response handling are the same; only binding style differs.

Quick contrast

#	Scenario	What makes it different from the others
1	Chat Completions	Normal chat completion object (or dict with the same choices / message / content layout). OpenRouter uses this same shape via an OpenAI-compatible client.
2	Responses API	Uses client.responses.create and string input, not chat.completions.
3	Multipart assistant content	Assistant content is a list of parts (e.g. multiple {"type":"text",...}), not a single string.
4	Tool calls + passthrough	passthrough_when returns the raw SDK object on tool-call turns so you can run another round trip before a final wrapped answer.
5	Nested wrapped calls	A parent function calls other wrapped functions; each invocation is a separate wrapped LLM call with its own agent_name.
6	Custom response_extractor	Answer lives in a custom field (here raw_text); without return_merger / answer pointer you often get a plain string back.
7	Non-reconstructable return	Object cannot be deep-copied or rebuilt; the wrapper uses model_dump()-style serialization and returns a dict tree instead of the original type.
8	Tool agent + wrapped LLM tools	Top chat turn uses tools= and passthrough_when (same idea as row 4), but each tool name is implemented by a separate wrapped LLM function, so every tool invocation is its own ingest (agent_name per tool).

1. Chat Completions

What this demonstrates. The common Chat Completions path: messages in, completion object out. The library can detect the answer slot and preserve the SDK return shape when possible.

How it differs. This is the baseline. OpenRouter (or any OpenAI-compatible base URL) is not a separate pattern: use the same chat.completions.create call with a different client and model id.

wrap_llm_call — prompt is the messages argument of your function.

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex1_chat_call",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="messages",
    max_tries=1,
)
def one_turn(messages):
    return openai_client.chat.completions.create(
        model=CFG.model, messages=messages, temperature=0
    )

wrap_llm_line — prompt lives inside a dict; prompt_json_pointer selects the user text to wrap.

payload = {
    "model": CFG.model,
    "messages": [{"role": "user", "content": "one sentence"}],
}
out = wrap_llm_line(
    llm_call=lambda p: openai_client.chat.completions.create(
        model=p["model"], messages=p["messages"], temperature=0
    ),
    prompt=payload,
    prompt_json_pointer="/messages/0/content",
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex1_chat_line",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    max_tries=1,
)

2. Responses API

What this demonstrates. The OpenAI Responses API: a string (or structured) input and responses.create, which returns a different object shape than chat completions.

How it differs. From example 1: different method, different prompt parameter name (input vs messages), different default extraction rules inside the wrapper.

wrap_llm_call

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex2_responses_call",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="question",
    max_tries=1,
)
def ask(question: str):
    return openai_client.responses.create(model=CFG.model, input=question)

wrap_llm_line

out = wrap_llm_line(
    llm_call=lambda prompt: openai_client.responses.create(
        model=CFG.model, input=prompt
    ),
    prompt="Return one sentence.",
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex2_responses_line",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    max_tries=1,
)

3. Multipart assistant content

What this demonstrates. The assistant message content is a list of segments (multipart), not a single string. The wrapper still extracts or merges answer text from that structure.

How it differs. From examples 1–2: tests list-shaped content under choices[0].message, which is a different merge path than a scalar string.

wrap_llm_call

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex3_multipart_call",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="messages",
    max_tries=1,
)
def one_turn(messages):
    return {
        "choices": [{
            "message": {
                "role": "assistant",
                "content": [
                    {"type": "text", "text": "Part A"},
                    {"type": "text", "text": "Part B"},
                ],
            }
        }]
    }

wrap_llm_line

payload = {"messages": [{"role": "user", "content": "multipart"}]}
out = wrap_llm_line(
    llm_call=lambda _p: {
        "choices": [{
            "message": {
                "role": "assistant",
                "content": [
                    {"type": "text", "text": "Part A"},
                    {"type": "text", "text": "Part B"},
                ],
            }
        }]
    },
    prompt=payload,
    prompt_json_pointer="/messages/0/content",
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex3_multipart_line",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    max_tries=1,
)

4. Tool calls with passthrough

What this demonstrates. When the model returns tool calls, you usually want the raw completion object back so your app can execute tools and call the model again. passthrough_when does that: no parse/ingest on those turns; later turns without tool calls follow the normal wrapped path.

How it differs. From examples 1–3: introduces branching behavior based on the raw return value and requires tools in the request.

Related pattern. Example 8 uses the same passthrough mechanism when each named tool is backed by its own wrapped LLM (see test_top_level_tool_agent_with_two_real_tools_and_nested_wrapped_llm_tools in tests/testing_different_interfaces.py).

wrap_llm_call

def has_tool_calls(raw):
    try:
        return bool(raw.choices[0].message.tool_calls)
    except Exception:
        return False

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex4_tools_call",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="messages",
    passthrough_when=has_tool_calls,
    max_tries=1,
)
def run_turn(messages):
    return openai_client.chat.completions.create(
        model=CFG.model, messages=messages, tools=TOOLS, tool_choice="auto"
    )

wrap_llm_line

out = wrap_llm_line(
    llm_call=lambda p: openai_client.chat.completions.create(
        model=p["model"],
        messages=p["messages"],
        tools=p["tools"],
        tool_choice="auto",
    ),
    prompt={"model": CFG.model, "messages": MESSAGES, "tools": TOOLS},
    prompt_json_pointer="/messages/1/content",
    passthrough_when=has_tool_calls,
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex4_tools_line",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    max_tries=1,
)

5. Nested wrapped calls

What this demonstrates. Composition: a top-level wrapped function calls other wrapped functions. Each inner call is a full wrapper invocation (its own agent_name, same or different Run_Id depending on how you thread context). This replaces the older README variants that repeated the same idea under different names (separate tools vs “manager” vs “hierarchy”).

How it differs. From examples 1–4: not about return shape; about call graph and multiple ingest events per user request.

wrap_llm_call

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex5_subagent_a",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="question",
    max_tries=1,
)
def subagent_a(question: str):
    return openai_client.responses.create(model=CFG.model, input=question)

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex5_subagent_b",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="question",
    max_tries=1,
)
def subagent_b(question: str):
    return openai_client.responses.create(model=CFG.model, input=question)

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex5_top_call",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="question",
    max_tries=1,
)
def top_agent(question: str):
    return subagent_a(question) + "\n" + subagent_b(question)

wrap_llm_line

def top_agent(question: str):
    a = wrap_llm_line(
        llm_call=lambda q: openai_client.responses.create(model=CFG.model, input=q),
        prompt=question,
        company_name=CFG.company,
        project_name=CFG.project,
        agent_name="ex5_subagent_a",
        Run_Id=RUN_ID,
        Workflow_name=WORKFLOW_NAME,
        workflow_group_name=WORKFLOW_GROUP_NAME,
        agent_parent=AGENT_PARENT,
        Workflow_parent_name=WORKFLOW_PARENT_NAME,
        metadata=METADATA,
        secret_key=CFG.key,
        max_tries=1,
    )
    b = wrap_llm_line(
        llm_call=lambda q: openai_client.responses.create(model=CFG.model, input=q),
        prompt=question,
        company_name=CFG.company,
        project_name=CFG.project,
        agent_name="ex5_subagent_b",
        Run_Id=RUN_ID,
        Workflow_name=WORKFLOW_NAME,
        workflow_group_name=WORKFLOW_GROUP_NAME,
        agent_parent=AGENT_PARENT,
        Workflow_parent_name=WORKFLOW_PARENT_NAME,
        metadata=METADATA,
        secret_key=CFG.key,
        max_tries=1,
    )
    return f"{a}\n{b}"

6. Custom response extractor

What this demonstrates. Your model function returns a custom object shape. You supply response_extractor so the wrapper knows where the answer text lives for ingest.

How it differs. From examples 1–5: answer is not in the default chat/response slots; if you do not also supply return_merger or response_answer_json_pointer, the wrapper may give you a plain string instead of preserving the original object.

wrap_llm_call

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex6_extractor_call",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="messages",
    response_extractor=lambda obj: obj["raw_text"],
    max_tries=1,
)
def one_turn(messages):
    return {"raw_text": "Model content here", "meta": {"id": "abc"}}

wrap_llm_line

out = wrap_llm_line(
    llm_call=lambda _prompt: {"raw_text": "model answer", "meta": {"id": "abc"}},
    prompt="hello",
    response_extractor=lambda obj: obj["raw_text"],
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex6_extractor_line",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    max_tries=1,
)

7. Non-reconstructable model output

What this demonstrates. Some return values cannot be cloned or rebuilt into the same Python type. The wrapper still needs a structured view for merge/ingest, so it falls back to a dict produced from model_dump()-style data.

How it differs. Focuses on serialization / copy failure, not on chat tools or custom extractors.

wrap_llm_call

class NonCopyable:
    def model_dump(self):
        return {
            "choices": [{"message": {"role": "assistant", "content": "hello"}}]
        }

    def __deepcopy__(self, memo):
        raise RuntimeError("cannot deepcopy")

@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex7_noncopy_call",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="messages",
    max_tries=1,
)
def one_turn(messages):
    return NonCopyable()

wrap_llm_line

class NonCopyable:
    def model_dump(self):
        return {
            "choices": [{"message": {"role": "assistant", "content": "hello"}}]
        }

    def __deepcopy__(self, memo):
        raise RuntimeError("cannot deepcopy")

out = wrap_llm_line(
    llm_call=lambda _prompt: NonCopyable(),
    prompt="one sentence",
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex7_noncopy_line",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    max_tries=1,
)

8. Tool agent with wrapped LLM tools

Plain summary. The coordinator model gets a tool menu via chat.completions.create(..., tools=...). When it returns tool_calls, your code runs real Python for each name. If a “tool” is actually another LLM call, wrap that call too: then ingest sees one event per tool LLM (its own agent_name) and separate events for the coordinator.

What goes in tools?
tools is a list of JSON descriptions for the API only. Each entry is usually {"type": "function", "function": {...}} with:

• name: string the model will put in tool_calls[].function.name
• description: free text shown to the model
• parameters: a small JSON Schema object (type, properties, required, …) describing the arguments JSON the model should output

It is not a list of Python callables. The provider uses this list so the model knows what to ask for; you map name → your own functions (or wrapped LLM helpers) after the response comes back.

Do I need something like a run_coordinator function?
llmwrap does not require it. Chat Completions tool use is multi-step: one response is either “here are tool_calls” or “here is final text”. After tool calls you append the assistant turn and one {"role": "tool", ...} message per call, then call the coordinator again. Any name (run_coordinator, your agent framework, inline code) is fine—the README uses one small driver so the flow is obvious.

Runnable reference. test_top_level_tool_agent_with_two_real_tools_and_nested_wrapped_llm_tools and test_top_level_tool_agent_with_two_real_tools_and_nested_wrapped_llm_tools_line_wrapper in tests/testing_different_interfaces.py.

How it differs from example 4. Same passthrough_when on the coordinator; example 8 adds wrapped LLM functions as the bodies behind specific tool names.

wrap_llm_call

import json

from llmwrap import openai_sdk_result_text, wrap_llm_call

# Declares two callable tools the MODEL may request (API schema, not Python functions).
TOOL_DEFINITIONS_FOR_API = [
    {
        "type": "function",
        "function": {
            "name": "get_operational_fact",
            "description": "Return one short operational fact about a topic.",
            "parameters": {
                "type": "object",
                "properties": {"topic": {"type": "string", "description": "Subject to summarize"}},
                "required": ["topic"],
            },
        },
    },
    {
        "type": "function",
        "function": {
            "name": "get_risk_fact",
            "description": "Return one short risk-oriented fact about a topic.",
            "parameters": {
                "type": "object",
                "properties": {"topic": {"type": "string"}},
                "required": ["topic"],
            },
        },
    },
]


def has_tool_calls(raw):
    try:
        return bool(raw.choices[0].message.tool_calls)
    except Exception:
        return False


@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex8_weather_tool",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="question",
    max_tries=1,
)
def weather_tool_llm(question: str):
    return openai_client.chat.completions.create(
        model=CFG.model,
        messages=[{"role": "user", "content": question}],
        temperature=0,
    )


@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex8_risk_tool",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="question",
    max_tries=1,
)
def risk_tool_llm(question: str):
    return openai_client.chat.completions.create(
        model=CFG.model,
        messages=[{"role": "user", "content": question}],
        temperature=0,
    )


@wrap_llm_call(
    company_name=CFG.company,
    project_name=CFG.project,
    agent_name="ex8_tool_coordinator",
    Run_Id=RUN_ID,
    Workflow_name=WORKFLOW_NAME,
    workflow_group_name=WORKFLOW_GROUP_NAME,
    agent_parent=AGENT_PARENT,
    Workflow_parent_name=WORKFLOW_PARENT_NAME,
    metadata=METADATA,
    secret_key=CFG.key,
    prompt_arg="messages",
    passthrough_when=has_tool_calls,
    max_tries=1,
)
def coordinator_turn(messages: list, tool_definitions: list):
    return openai_client.chat.completions.create(
        model=CFG.model,
        messages=messages,
        tools=tool_definitions,
        tool_choice="auto",
        temperature=0,
    )


# App-side driver (name arbitrary): repeat coordinator → execute tools → append results.
def run_until_final_answer(user_text: str) -> str:
    messages = [
        {
            "role": "system",
            "content": "Use the tools when needed, then answer in plain language.",
        },
        {"role": "user", "content": user_text},
    ]
    for _ in range(8):
        out = coordinator_turn(messages, TOOL_DEFINITIONS_FOR_API)
        msg = out.choices[0].message
        if not msg.tool_calls:
            return openai_sdk_result_text(out)
        messages.append(msg.model_dump(exclude_none=True))
        for tc in msg.tool_calls:
            args = json.loads(tc.function.arguments or "{}")
            topic = str(args.get("topic", ""))
            if tc.function.name == "get_operational_fact":
                fact = openai_sdk_result_text(
                    weather_tool_llm(f"One operational fact about {topic}")
                )
            elif tc.function.name == "get_risk_fact":
                fact = openai_sdk_result_text(
                    risk_tool_llm(f"One risk fact about {topic}")
                )
            else:
                fact = "unknown tool"
            messages.append(
                {
                    "role": "tool",
                    "tool_call_id": tc.id,
                    "content": json.dumps({"fact": fact}),
                }
            )
    raise RuntimeError("Tool loop limit exceeded")

wrap_llm_line

Same TOOL_DEFINITIONS_FOR_API and the same driver idea: coordinator = one wrap_llm_line with passthrough_when and prompt_json_pointer on the user message; each tool = wrap_llm_line on a small payload (here {"model", "question"}).

import json

from llmwrap import openai_sdk_result_text, wrap_llm_line

# TOOL_DEFINITIONS_FOR_API: identical list as in the wrap_llm_call example above.


def weather_tool_llm(question: str):
    payload = {"model": CFG.model, "question": question}
    return wrap_llm_line(
        llm_call=lambda p: openai_client.chat.completions.create(
            model=p["model"],
            messages=[{"role": "user", "content": p["question"]}],
            temperature=0,
        ),
        prompt=payload,
        prompt_json_pointer="/question",
        company_name=CFG.company,
        project_name=CFG.project,
        agent_name="ex8_weather_tool",
        Run_Id=RUN_ID,
        Workflow_name=WORKFLOW_NAME,
        workflow_group_name=WORKFLOW_GROUP_NAME,
        agent_parent=AGENT_PARENT,
        Workflow_parent_name=WORKFLOW_PARENT_NAME,
        metadata=METADATA,
        secret_key=CFG.key,
        max_tries=1,
    )


def risk_tool_llm(question: str):
    payload = {"model": CFG.model, "question": question}
    return wrap_llm_line(
        llm_call=lambda p: openai_client.chat.completions.create(
            model=p["model"],
            messages=[{"role": "user", "content": p["question"]}],
            temperature=0,
        ),
        prompt=payload,
        prompt_json_pointer="/question",
        company_name=CFG.company,
        project_name=CFG.project,
        agent_name="ex8_risk_tool",
        Run_Id=RUN_ID,
        Workflow_name=WORKFLOW_NAME,
        workflow_group_name=WORKFLOW_GROUP_NAME,
        agent_parent=AGENT_PARENT,
        Workflow_parent_name=WORKFLOW_PARENT_NAME,
        metadata=METADATA,
        secret_key=CFG.key,
        max_tries=1,
    )


def has_tool_calls(raw):
    try:
        return bool(raw.choices[0].message.tool_calls)
    except Exception:
        return False


def coordinator_turn(messages: list, tool_definitions: list):
    return wrap_llm_line(
        llm_call=lambda p: openai_client.chat.completions.create(
            model=p["model"],
            messages=p["messages"],
            tools=p["tools"],
            tool_choice="auto",
            temperature=0,
        ),
        prompt={"model": CFG.model, "messages": messages, "tools": tool_definitions},
        prompt_json_pointer="/messages/1/content",
        passthrough_when=has_tool_calls,
        company_name=CFG.company,
        project_name=CFG.project,
        agent_name="ex8_tool_coordinator",
        Run_Id=RUN_ID,
        Workflow_name=WORKFLOW_NAME,
        workflow_group_name=WORKFLOW_GROUP_NAME,
        agent_parent=AGENT_PARENT,
        Workflow_parent_name=WORKFLOW_PARENT_NAME,
        metadata=METADATA,
        secret_key=CFG.key,
        max_tries=1,
    )


def run_until_final_answer(user_text: str) -> str:
    messages = [
        {"role": "system", "content": "Use the tools when needed, then answer."},
        {"role": "user", "content": user_text},
    ]
    for _ in range(8):
        out = coordinator_turn(messages, TOOL_DEFINITIONS_FOR_API)
        msg = out.choices[0].message
        if not msg.tool_calls:
            return openai_sdk_result_text(out)
        messages.append(msg.model_dump(exclude_none=True))
        for tc in msg.tool_calls:
            args = json.loads(tc.function.arguments or "{}")
            topic = str(args.get("topic", ""))
            if tc.function.name == "get_operational_fact":
                fact = openai_sdk_result_text(
                    weather_tool_llm(f"One operational fact about {topic}")
                )
            elif tc.function.name == "get_risk_fact":
                fact = openai_sdk_result_text(
                    risk_tool_llm(f"One risk fact about {topic}")
                )
            else:
                fact = "unknown tool"
            messages.append(
                {
                    "role": "tool",
                    "tool_call_id": tc.id,
                    "content": json.dumps({"fact": fact}),
                }
            )
    raise RuntimeError("Tool loop limit exceeded")

Console output

Each wrapped call may emit one line to stdout when ingest succeeds or when ingest / logs reporting fails. The line starts with [Arbis-Wrapper], followed by a single JSON object (no pretty-printing). If stdout is a TTY, that line is wrapped in green (success) or red (failure) ANSI sequences; if stdout is not a TTY (pipes, CI, log capture), the same JSON is printed without color codes.

Successful POSTs to /api/logs do not print anything to the console.

Success (after ingest HTTP 2xx)

Emitted when encrypted POST …/api/ingest returns a success status code (after the model output was parsed successfully).

Expected shape (values are examples):

[Arbis-Wrapper] {"message": "Query processed successfully; ingest accepted.", "api_status": 200, "path": "/api/ingest", "query": "<original logical prompt passed to the wrapper>"}

• message: Fixed success text.
• api_status: HTTP status from the ingest response (typically 200).
• path: "/api/ingest" for the normal ingest path.
• query: The wrapper’s original prompt string used for correlation (the same logical input ingest uses as prompt, not the augmented “wrapped” block sent to the model).

If the pipeline used the fallback ingest instead (POST …/ingest/fallback after parse retries were exhausted), a success line looks the same except path is "/ingest/fallback" and query is still the logical prompt:

[Arbis-Wrapper] {"message": "Query processed successfully; ingest accepted.", "api_status": 200, "path": "/ingest/fallback", "query": "<original logical prompt>"}

Failure

Emitted when ingest or fallback ingest fails, or when POST …/api/logs fails after all retries. Same [Arbis-Wrapper] prefix and single JSON object; red when stdout is a TTY.

Main ingest failed (model output parsed, but /api/ingest did not return 2xx, or the request raised before a response):

[Arbis-Wrapper] {"message": "Ingest request failed.", "http_status": 401, "error": "<API error body or short reason>"}

Fallback ingest failed (/ingest/fallback):

[Arbis-Wrapper] {"message": "Ingest fallback request failed.", "http_status": 503, "error": "<API error body or short reason>"}

Wrapper logs failed (all attempts to /api/logs failed):

[Arbis-Wrapper] {"message": "Failed to record wrapper logs to the server.", "http_status": 500, "error": "<API error body or exception text>"}

When there is no HTTP response (timeouts, connection errors, etc.), http_status is JSON null:

[Arbis-Wrapper] {"message": "Ingest request failed.", "http_status": null, "error": "<exception message>"}

If the library has no usable error text, error may be "(no detailed error available)".

Notes

• Keep credentials in environment variables (.env) and never hardcode production keys.
• Use distinct agent_name values per workflow for clean tracking.
• Pass Run_Id, workflow fields, and metadata when you need ingest to group steps or attach user context (user_id, role, and so on); omit any argument you do not use.
• For custom return shapes, pair response_extractor with response_answer_json_pointer or return_merger when needed.
• Exhaustive runnable variants (including redundant naming patterns) are in tests/testing_different_interfaces.py. For a linear workflow-group chain demo, see tests/test_linear_workflow_group_chain.py.

License

MIT. See LICENSE.