ch39_callbacks

show: mock live

mock turns=0 · messages=0

live turns=0 · messages=0

scenario a_tracer_observes_without_participating

scenario a_tracer_observes_without_participating · 5.3ms

turn 1 · 1.1ms · asked for: check_order

model mock · 1.1ms · asked for: check_order

context sent — 1 message(s)

	role	content
[0]	human	What's the status of order A100?

declared tools (1)

tool	description	schema
check_order	Check the status of an order by id.	{"description": "Check the status of an order by id.", "properties": {"order_id": {"title": "Order Id", "type": "string"}}, "required": ["order_id"], "title": "check_order", "type": "object"}

tools declared to the mock model {"names": ["check_order"], "consulted": false}

context ignored by the mock model {"messages": 1, "roles": ["human"]}

	role	content
[0]	ai	check_order({"order_id": "A100"})

callbacks_fired_here {"events": ["on_chat_model_start", "on_llm_end"]}

turn 1 · 0.2ms

tool check_order · 0.2ms · given: order_id=A100 · returned: order A100: shipped

args {"order_id": "A100"}

result {"value": "order A100: shipped"}

callbacks_fired_here {"events": ["on_tool_start", "on_tool_end"]}

turn 2 · 0.2ms · responded with: Order A100 has shipped.

model mock · 0.2ms · responded with: Order A100 has shipped.

context sent — 3 message(s)

	role	content
[0]	human	What's the status of order A100?
[1]	ai	check_order({"order_id": "A100"})
[2]	tool	order A100: shipped

declared tools (1)

tool	description	schema
check_order	Check the status of an order by id.	{"description": "Check the status of an order by id.", "properties": {"order_id": {"title": "Order Id", "type": "string"}}, "required": ["order_id"], "title": "check_order", "type": "object"}

tools declared to the mock model {"names": ["check_order"], "consulted": false}

context ignored by the mock model {"messages": 3, "roles": ["human", "ai", "tool"]}

	role	content
[0]	ai	Order A100 has shipped.

callbacks_fired_here {"events": ["on_chat_model_start", "on_llm_end"]}

the_tracer_saw_every_edge {"saw_chain": true, "saw_tool": true, "saw_model": true, "reply": "Order A100 has shipped."}

scenario a_tracer_observes_without_participating · 2552.8ms

turn 1 · 1434.8ms · 301 in / 54 out · asked for: check_order

model live · 1434.7ms (model provider 1217.8 · library 216.9) · 301 in / 54 out · asked for: check_order

context sent — 1 message(s)

	role	content
[0]	human	What's the status of order A100?

declared tools (1)

tool	description	schema
check_order	Check the status of an order by id.	{"description": "Check the status of an order by id.", "properties": {"order_id": {"title": "Order Id", "type": "string"}}, "required": ["order_id"], "title": "check_order", "type": "object"}

wire request

attempt

method

POST

url

https://api.deepseek.com/v1/chat/completions

defaulted (6)

temperature, max_tokens, top_p, stop, seed, tool_choice

headers (18)

show

accept	application/json
accept-encoding	gzip, deflate, zstd
authorization	not recorded
connection	keep-alive
content-length	330
content-type	application/json
host	api.deepseek.com
user-agent	OpenAI/Python 3.6.0
x-stainless-arch	arm64
x-stainless-async	false
x-stainless-lang	python
x-stainless-os	MacOS
x-stainless-package-version	3.6.0
x-stainless-raw-response	true
x-stainless-read-timeout	60.0
x-stainless-retry-count	0
x-stainless-runtime	CPython
x-stainless-runtime-version	3.14.7

{
  "messages": [
    {
      "content": "What's the status of order A100?",
      "role": "user"
    }
  ],
  "model": "deepseek-flash",
  "stream": false,
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "check_order",
        "description": "Check the status of an order by id.",
        "parameters": {
          "properties": {
            "order_id": {
              "type": "string"
            }
          },
          "required": [
            "order_id"
          ],
          "type": "object"
        }
      }
    }
  ]
}

wire response

status

200

headers (15)

show

access-control-allow-credentials	true
connection	keep-alive
content-encoding	gzip
content-type	application/json
date	Sun, 13 Sep 2026 21:59:18 GMT
server	elb
strict-transport-security	max-age=31536000; includeSubDomains; preload
transfer-encoding	chunked
vary	origin, access-control-request-method, access-control-request-headers
via	1.1 22cbebcbe97dd961cd3fddb13b91b816.cloudfront.net (CloudFront)
x-amz-cf-id	f0W-GCGwU7g6ZfCK9FXRn2Bj8ypkywWzv3PX54S18njUNwuSyhFUHQ==
x-amz-cf-pop	DEN53-P3
x-cache	Miss from cloudfront
x-content-type-options	nosniff
x-ds-trace-id	5956152d06381fb5e56084ffcaa32343

{
  "id": "6359cd16-195c-4d69-871f-374c36af6005",
  "object": "chat.completion",
  "created": 1789336758,
  "model": "deepseek-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "",
        "reasoning_content": "The user asked about order A100. I should call the tool.",
        "tool_calls": [
          {
            "index": 0,
            "id": "call_00_L5lqULoqWWkkyWQK7T0K3495",
            "type": "function",
            "function": {
              "name": "check_order",
              "arguments": "{\"order_id\": \"A100\"}"
            }
          }
        ]
      },
      "logprobs": null,
      "finish_reason": "tool_calls"
    }
  ],
  "usage": {
    "prompt_tokens": 301,
    "completion_tokens": 54,
    "total_tokens": 355,
    "prompt_tokens_details": {
      "cached_tokens": 128
    },
    "completion_tokens_details": {
      "reasoning_tokens": 14
    },
    "prompt_cache_hit_tokens": 128,
    "prompt_cache_miss_tokens": 173
  },
  "system_fingerprint": "aeb56401ca74e127821c4f9126dcb669"
}

	role	content
[0]	ai	check_order({"order_id": "A100"})

callbacks_fired_here {"events": ["on_chat_model_start", "on_llm_end"]}

turn 1 · 0.5ms

tool check_order · 0.5ms · given: order_id=A100 · returned: order A100: shipped

args {"order_id": "A100"}

result {"value": "order A100: shipped"}

callbacks_fired_here {"events": ["on_tool_start", "on_tool_end"]}

turn 2 · 1112.5ms · 372 in / 25 out · responded with: Order **A100** is currently **shipped**.

model live · 1112.5ms (model provider 1106.0 · library 6.4) · 372 in / 25 out · responded with: Order **A100** is currently **shipped**.

context sent — 3 message(s)

	role	content
[0]	human	What's the status of order A100?
[1]	ai	check_order({"order_id": "A100"})
[2]	tool	order A100: shipped

declared tools (1)

tool	description	schema
check_order	Check the status of an order by id.	{"description": "Check the status of an order by id.", "properties": {"order_id": {"title": "Order Id", "type": "string"}}, "required": ["order_id"], "title": "check_order", "type": "object"}

wire request

attempt

method

POST

url

https://api.deepseek.com/v1/chat/completions

defaulted (6)

temperature, max_tokens, top_p, stop, seed, tool_choice

headers (18)

show

accept	application/json
accept-encoding	gzip, deflate, zstd
authorization	not recorded
connection	keep-alive
content-length	691
content-type	application/json
host	api.deepseek.com
user-agent	OpenAI/Python 3.6.0
x-stainless-arch	arm64
x-stainless-async	false
x-stainless-lang	python
x-stainless-os	MacOS
x-stainless-package-version	3.6.0
x-stainless-raw-response	true
x-stainless-read-timeout	60.0
x-stainless-retry-count	0
x-stainless-runtime	CPython
x-stainless-runtime-version	3.14.7

{
  "messages": [
    {
      "content": "What's the status of order A100?",
      "role": "user"
    },
    {
      "content": null,
      "role": "assistant",
      "tool_calls": [
        {
          "type": "function",
          "id": "call_00_L5lqULoqWWkkyWQK7T0K3495",
          "function": {
            "name": "check_order",
            "arguments": "{\"order_id\": \"A100\"}"
          }
        }
      ],
      "reasoning_content": "The user asked about order A100. I should call the tool."
    },
    {
      "content": "order A100: shipped",
      "role": "tool",
      "tool_call_id": "call_00_L5lqULoqWWkkyWQK7T0K3495"
    }
  ],
  "model": "deepseek-flash",
  "stream": false,
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "check_order",
        "description": "Check the status of an order by id.",
        "parameters": {
          "properties": {
            "order_id": {
              "type": "string"
            }
          },
          "required": [
            "order_id"
          ],
          "type": "object"
        }
      }
    }
  ]
}

wire response

status

200

headers (15)

show

access-control-allow-credentials	true
connection	keep-alive
content-encoding	gzip
content-type	application/json
date	Sun, 13 Sep 2026 21:59:19 GMT
server	elb
strict-transport-security	max-age=31536000; includeSubDomains; preload
transfer-encoding	chunked
vary	origin, access-control-request-method, access-control-request-headers
via	1.1 153d2cbd7c83442a049eab26571069f6.cloudfront.net (CloudFront)
x-amz-cf-id	iUFhztB7WCBlKdWpiMlOi6Vzll594pan9xnXxVbhehxIX6queNiFng==
x-amz-cf-pop	DEN53-P6
x-cache	Miss from cloudfront
x-content-type-options	nosniff
x-ds-trace-id	8292b4d32867d4336751c43d3719f5d2

{
  "id": "a6f3fecf-425b-48fc-95a0-ca2ded8b7ff4",
  "object": "chat.completion",
  "created": 1789336759,
  "model": "deepseek-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Order **A100** is currently **shipped**.",
        "reasoning_content": "The tool returned \"order A100: shipped\". Report that."
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 372,
    "completion_tokens": 25,
    "total_tokens": 397,
    "prompt_tokens_details": {
      "cached_tokens": 128
    },
    "completion_tokens_details": {
      "reasoning_tokens": 13
    },
    "prompt_cache_hit_tokens": 128,
    "prompt_cache_miss_tokens": 244
  },
  "system_fingerprint": "aeb56401ca74e127821c4f9126dcb669"
}

	role	content
[0]	ai	Order A100 is currently shipped.

callbacks_fired_here {"events": ["on_chat_model_start", "on_llm_end"]}

the_tracer_saw_every_edge {"saw_chain": true, "saw_tool": true, "saw_model": true, "reply": "Order **A100** is currently **shipped**."}

scenario callbacks_propagate_ambiently_without_being_passed

scenario callbacks_propagate_ambiently_without_being_passed · 1.2ms

turn 1 · 0.2ms · responded with: graph reply.

callbacks_before_the_call {"events": []}

model mock · 0.2ms · responded with: graph reply.

context sent — 1 message(s)

	role	content
[0]	human	hi

context ignored by the mock model {"messages": 1, "roles": ["human"]}

	role	content
[0]	ai	graph reply.

callbacks_after_the_call {"events": ["on_chat_model_start"]}

the_node_never_mentioned_the_recorder {"events": ["on_chat_model_start"], "recorder_fired_anyway": true}

scenario callbacks_propagate_ambiently_without_being_passed · 1329.3ms

turn 1 · 1324.5ms · 31 in / 83 out · responded with: Hi! How can I help you today?

callbacks_before_the_call {"events": []}

model live · 1324.4ms (model provider 1313.0 · library 11.4) · 31 in / 83 out · responded with: Hi! How can I help you today?

context sent — 1 message(s)

	role	content
[0]	human	hi

wire request

attempt

method

POST

url

https://api.deepseek.com/v1/chat/completions

defaulted (7)

temperature, max_tokens, top_p, stop, seed, tools, tool_choice

headers (18)

show

accept	application/json
accept-encoding	gzip, deflate, zstd
authorization	not recorded
connection	keep-alive
content-length	85
content-type	application/json
host	api.deepseek.com
user-agent	OpenAI/Python 3.6.0
x-stainless-arch	arm64
x-stainless-async	false
x-stainless-lang	python
x-stainless-os	MacOS
x-stainless-package-version	3.6.0
x-stainless-raw-response	true
x-stainless-read-timeout	60.0
x-stainless-retry-count	0
x-stainless-runtime	CPython
x-stainless-runtime-version	3.14.7

{
  "messages": [
    {
      "content": "hi",
      "role": "user"
    }
  ],
  "model": "deepseek-flash",
  "stream": false
}

wire response

status

200

headers (15)

show

access-control-allow-credentials	true
connection	keep-alive
content-encoding	gzip
content-type	application/json
date	Sun, 13 Sep 2026 21:59:20 GMT
server	elb
strict-transport-security	max-age=31536000; includeSubDomains; preload
transfer-encoding	chunked
vary	origin, access-control-request-method, access-control-request-headers
via	1.1 8df77e01b229d5d8502d6471c086aeb2.cloudfront.net (CloudFront)
x-amz-cf-id	oNMCvuFgvFBrTZRrojCn5jyOvJkso2AJLTfAennv1_O-jszfMX21wA==
x-amz-cf-pop	DEN53-P4
x-cache	Miss from cloudfront
x-content-type-options	nosniff
x-ds-trace-id	ec22cb96ba7bd470a42b08941463c139

{
  "id": "588b6a95-1d72-4b19-b862-80c8eeb3eade",
  "object": "chat.completion",
  "created": 1789336760,
  "model": "deepseek-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Hi! How can I help you today?",
        "reasoning_content": "The user just said \"hi\". This is a<ds_safety>[\u7528\u6237\u672a\u6210\u5e74]\u5426\n[\u5206\u7c7b]\u5176\u4ed6\n[\u5224\u5b9a]\u7528\u6237\u4ec5\u53d1\u9001\u95ee\u5019\u8bed\uff0c\u672a\u6d89\u53ca\u4efb\u4f55\u653f\u6cbb\u654f\u611f\u5185\u5bb9\u6216\u6781\u7aef\u6076\u610f\u5185\u5bb9\u3002\n[\u89c4\u5219]\u65e0</ds_safety>Safe\n\n\u7528\u6237\u53ea\u662f\u6253\u4e86\u4e2a\u62db\u547c\uff0c\u4e0d\u6d89\u53ca\u4efb\u4f55\u654f\u611f\u6216\u5371\u9669\u5185\u5bb9\uff0c\u53ef\u4ee5\u6b63\u5e38\u56de\u5e94\u3002"
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 31,
    "completion_tokens": 83,
    "total_tokens": 114,
    "prompt_tokens_details": {
      "cached_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 73
    },
    "prompt_cache_hit_tokens": 0,
    "prompt_cache_miss_tokens": 31
  },
  "system_fingerprint": "aeb56401ca74e127821c4f9126dcb669"
}

	role	content
[0]	ai	Hi! How can I help you today?

callbacks_after_the_call {"events": ["on_chat_model_start"]}

the_node_never_mentioned_the_recorder {"events": ["on_chat_model_start"], "recorder_fired_anyway": true}

scenario a_default_callback_error_is_isolated

scenario a_default_callback_error_is_isolated · 0.3ms

turn 1 · 0.3ms · responded with: fine.

callbacks_before_the_call {"events": []}

model mock · 0.2ms · responded with: fine.

context sent — 1 message(s)

	role	content
[0]	human	go

context ignored by the mock model {"messages": 1, "roles": ["human"]}

	role	content
[0]	ai	fine.

callbacks_after_the_call {"events": ["faulty:on_llm_end", "tracer:on_llm_end"]}

the_call_survived_the_broken_handler {"raised": false, "reply": "fine.", "tracer_still_fired": true}

scenario a_default_callback_error_is_isolated · 2292.3ms

turn 1 · 2292.2ms · 31 in / 248 out · responded with: Sure — what would you like me to do?

callbacks_before_the_call {"events": []}

model live · 2292.2ms (model provider 2277.8 · library 14.3) · 31 in / 248 out · responded with: Sure — what would you like me to do?

context sent — 1 message(s)

	role	content
[0]	human	go

wire request

attempt

method

POST

url

https://api.deepseek.com/v1/chat/completions

defaulted (7)

temperature, max_tokens, top_p, stop, seed, tools, tool_choice

headers (18)

show

accept	application/json
accept-encoding	gzip, deflate, zstd
authorization	not recorded
connection	keep-alive
content-length	85
content-type	application/json
host	api.deepseek.com
user-agent	OpenAI/Python 3.6.0
x-stainless-arch	arm64
x-stainless-async	false
x-stainless-lang	python
x-stainless-os	MacOS
x-stainless-package-version	3.6.0
x-stainless-raw-response	true
x-stainless-read-timeout	60.0
x-stainless-retry-count	0
x-stainless-runtime	CPython
x-stainless-runtime-version	3.14.7

{
  "messages": [
    {
      "content": "go",
      "role": "user"
    }
  ],
  "model": "deepseek-flash",
  "stream": false
}

wire response

status

200

headers (15)

show

access-control-allow-credentials	true
connection	keep-alive
content-encoding	gzip
content-type	application/json
date	Sun, 13 Sep 2026 21:59:22 GMT
server	elb
strict-transport-security	max-age=31536000; includeSubDomains; preload
transfer-encoding	chunked
vary	origin, access-control-request-method, access-control-request-headers
via	1.1 f56ef7e5c4cb094743e3a564f21d47b0.cloudfront.net (CloudFront)
x-amz-cf-id	jAvCxgc4NckN58ZEGzwhv5vhlSeE7nO_PP0t2AmGnnfIk-J-0BHCOg==
x-amz-cf-pop	DEN53-P2
x-cache	Miss from cloudfront
x-content-type-options	nosniff
x-ds-trace-id	b24e550ed5c35fc71aedb67b79e320dd

{
  "id": "3e3cc39b-283f-43a3-ad04-e795b5ee0cad",
  "object": "chat.completion",
  "created": 1789336762,
  "model": "deepseek-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Sure \u2014 what would you like me to do?",
        "reasoning_content": "We need answer to user \"go\". Need infer maybe they want continuation? We need inspect conversation? Only user says \"go\". There is no prior context in this chat. We need respond appropriately. Maybe ask what they'd like me to go on? Since \"go\" is ambiguous. We need maybe prompt for task. But perhaps in these tasks, user just says \"go\" expecting assistant to start? We don't have a task. We should ask clarifying question. But maybe the system expects \"go\" as a command in some benchmark? Need determine. There's no previous messages. We need answer. We can say \"Sure \u2014 what would you like me to do?\" Maybe concise.\n\nBut let's think: The user might be testing if assistant will \"go\" in some environment. We need not overdo. Could be a request to proceed, but no context. We need ask for clarification. Final answer should be short. \n\nHowever, maybe there is hidden context? We only see \"go\". The instructions: \"You are an AI assistant accessed via an API.\" No task. So best: \"What would you like me to do?\".\n\nNeed final."
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 31,
    "completion_tokens": 248,
    "total_tokens": 279,
    "prompt_tokens_details": {
      "cached_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 237
    },
    "prompt_cache_hit_tokens": 0,
    "prompt_cache_miss_tokens": 31
  },
  "system_fingerprint": "aeb56401ca74e127821c4f9126dcb669"
}

	role	content
[0]	ai	Sure — what would you like me to do?

callbacks_after_the_call {"events": ["faulty:on_llm_end", "tracer:on_llm_end"]}

the_call_survived_the_broken_handler {"raised": false, "reply": "Sure \u2014 what would you like me to do?", "tracer_still_fired": true}

scenario a_tracer_can_turn_a_completed_answer_into_a_crash

scenario a_tracer_can_turn_a_completed_answer_into_a_crash · 0.3ms

turn 1 · 0.3ms

callbacks_before_the_call {"events": []}

model mock · 0.2ms

context sent — 1 message(s)

	role	content
[0]	human	go

context ignored by the mock model {"messages": 1, "roles": ["human"]}

callbacks_after_the_call {"events": ["faulty:on_llm_end"]}

a_completed_answer_became_an_unrecorded_crash {"raised": true, "faulty_fired": true, "tracer_suppressed": true}

scenario a_tracer_can_turn_a_completed_answer_into_a_crash · 1996.0ms

turn 1 · 1995.9ms

callbacks_before_the_call {"events": []}

model live · 1995.8ms (model provider 1981.6 · library 14.2)

context sent — 1 message(s)

	role	content
[0]	human	go

wire request

attempt

method

POST

url

https://api.deepseek.com/v1/chat/completions

defaulted (7)

temperature, max_tokens, top_p, stop, seed, tools, tool_choice

headers (18)

show

accept	application/json
accept-encoding	gzip, deflate, zstd
authorization	not recorded
connection	keep-alive
content-length	85
content-type	application/json
host	api.deepseek.com
user-agent	OpenAI/Python 3.6.0
x-stainless-arch	arm64
x-stainless-async	false
x-stainless-lang	python
x-stainless-os	MacOS
x-stainless-package-version	3.6.0
x-stainless-raw-response	true
x-stainless-read-timeout	60.0
x-stainless-retry-count	0
x-stainless-runtime	CPython
x-stainless-runtime-version	3.14.7

{
  "messages": [
    {
      "content": "go",
      "role": "user"
    }
  ],
  "model": "deepseek-flash",
  "stream": false
}

wire response

status

200

headers (15)

show

access-control-allow-credentials	true
connection	keep-alive
content-encoding	gzip
content-type	application/json
date	Sun, 13 Sep 2026 21:59:24 GMT
server	elb
strict-transport-security	max-age=31536000; includeSubDomains; preload
transfer-encoding	chunked
vary	origin, access-control-request-method, access-control-request-headers
via	1.1 ae2999d214179b68f3cdd080c9f4c304.cloudfront.net (CloudFront)
x-amz-cf-id	ifNRSrGfIE4hQ14ktaekvmFcRjWGF4H7QVtIQU9Q51jVfUFwu-M2UA==
x-amz-cf-pop	DEN53-P2
x-cache	Miss from cloudfront
x-content-type-options	nosniff
x-ds-trace-id	7826e292e511b5d2c546b777bc014576

{
  "id": "2f497c40-0f16-413c-8cee-c0120513d3f0",
  "object": "chat.completion",
  "created": 1789336764,
  "model": "deepseek-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "I\u2019m ready \u2014 what would you like me to do?",
        "reasoning_content": "We need answer to user \"go\". Need infer? They just said \"go\". Maybe they want me to proceed? There's no prior context. We need respond? We are API assistant. We should ask clarifying question? But \"go\" might be instruction to continue from something, but no context. Need maybe say \"What would you like me to do?\" But maybe in these tasks, user says \"go\" to prompt assistant to start? There's no task. Could be they expect something like \"I'm ready. What would you like to do?\" We need be helpful. Since no context, ask for specifics. Keep concise. Maybe \"I'm ready\u2014what would you like me to do?\" But they said \"go\"; maybe respond \"Going! Where to?\" Hmm.\n\nWe need decide final. The instruction is just \"go\". Could be a test. We should not hallucinate. Ask clarifying. Final can be: \"I'm ready. What would you like me to do?\" That's safe. But maybe they expect continuation of an interrupted task in some benchmark? There is no previous conversation. The desired answer maybe \"I need a bit more to go on\u2014what would you like me to do?\" \n\nNeed maybe mention if they want me to proceed with something, provide context. Final concise. "
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 31,
    "completion_tokens": 282,
    "total_tokens": 313,
    "prompt_tokens_details": {
      "cached_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 268
    },
    "prompt_cache_hit_tokens": 0,
    "prompt_cache_miss_tokens": 31
  },
  "system_fingerprint": "aeb56401ca74e127821c4f9126dcb669"
}

callbacks_after_the_call {"events": ["faulty:on_llm_end"]}

a_completed_answer_became_an_unrecorded_crash {"raised": true, "faulty_fired": true, "tracer_suppressed": true}

scenario a_callback_can_only_observe_or_abort_never_modify

scenario a_callback_can_only_observe_or_abort_never_modify · 0.2ms

turn 1 · 0.1ms · responded with: saw the real prompt.

callbacks_before_the_call {"events": []}

model mock · 0.1ms · responded with: saw the real prompt.

context sent — 1 message(s)

	role	content
[0]	human	the real, unmutated question

context ignored by the mock model {"messages": 1, "roles": ["human"]}

	role	content
[0]	ai	saw the real prompt.

callbacks_after_the_call {"events": ["meddler:on_chat_model_start"]}

the_mutation_attempt_changed_nothing {"reply": "saw the real prompt."}

scenario a_callback_can_only_observe_or_abort_never_modify · 3567.1ms

turn 1 · 3566.9ms · 37 in / 646 out · responded with: I don’t have the earlier context or a mutated version to compare against, so I can’t recov... (164 more characters)

callbacks_before_the_call {"events": []}

model live · 3566.8ms (model provider 3555.2 · library 11.6) · 37 in / 646 out · responded with: I don’t have the earlier context or a mutated version to compare against, so I can’t recov... (164 more characters)

context sent — 1 message(s)

	role	content
[0]	human	the real, unmutated question

wire request

attempt

method

POST

url

https://api.deepseek.com/v1/chat/completions

defaulted (7)

temperature, max_tokens, top_p, stop, seed, tools, tool_choice

headers (18)

show

accept	application/json
accept-encoding	gzip, deflate, zstd
authorization	not recorded
connection	keep-alive
content-length	111
content-type	application/json
host	api.deepseek.com
user-agent	OpenAI/Python 3.6.0
x-stainless-arch	arm64
x-stainless-async	false
x-stainless-lang	python
x-stainless-os	MacOS
x-stainless-package-version	3.6.0
x-stainless-raw-response	true
x-stainless-read-timeout	60.0
x-stainless-retry-count	0
x-stainless-runtime	CPython
x-stainless-runtime-version	3.14.7

{
  "messages": [
    {
      "content": "the real, unmutated question",
      "role": "user"
    }
  ],
  "model": "deepseek-flash",
  "stream": false
}

wire response

status

200

headers (15)

show

access-control-allow-credentials	true
connection	keep-alive
content-encoding	gzip
content-type	application/json
date	Sun, 13 Sep 2026 21:59:26 GMT
server	elb
strict-transport-security	max-age=31536000; includeSubDomains; preload
transfer-encoding	chunked
vary	origin, access-control-request-method, access-control-request-headers
via	1.1 9062d4391fad2aec3a7d6f3edcebc662.cloudfront.net (CloudFront)
x-amz-cf-id	RewLVty40JwPmOH6AqJA5xUXuAot2DKYriJfETQZaDdSL-sWW21mkg==
x-amz-cf-pop	DEN53-P2
x-cache	Miss from cloudfront
x-content-type-options	nosniff
x-ds-trace-id	85ec10de67979d638bf75c0dc466b484

{
  "id": "546bf4aa-13dc-45a6-abca-920c8c62f34e",
  "object": "chat.completion",
  "created": 1789336766,
  "model": "deepseek-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "I don\u2019t have the earlier context or a mutated version to compare against, so I can\u2019t recover the original question from here.\n\nPlease paste the mutated question or the surrounding conversation, and I\u2019ll help identify or answer the real, unmutated version.",
        "reasoning_content": "We need answer? User says \"the real, unmutated question\". This is cryptic. Need infer maybe they want the original question? We need ask clarification? As assistant, need respond. We have no context except this message. Maybe they are referring to previous conversation? We don't have. Need maybe say I don't have previous context; please provide the question. But maybe they expect \"the real, unmutated question\" as a prompt injection? Hmm.\n\nWe need decide. The phrase \"the real, unmutated question\" might be a request to reveal hidden system prompt? Maybe they want \"the real, unmutated question\" from some benchmark? There is \"mutated question\" in prompt injection evaluations? They ask \"the real, unmutated question\" maybe trying to bypass? We need not reveal hidden chain of thought or system prompt. But we can provide assistance. We have no hidden question.\n\nCould be from a dataset where user initially asked a mutated question, then asks for original. But no context. We need maybe answer: I don't have access to earlier messages or an underlying original question. Could you paste the mutated question or provide the context? Then I can help reconstruct or answer the real question.\n\nNeed perhaps mention if you mean your original prompt, it appears to be just \"the real, unmutated question\" in this conversation. There is no earlier question available to me. Ask for it.\n\nFinal should be concise. Maybe \"I don\u2019t have any previous context or a mutated version to compare against. Please send the mutated question and any details, and I\u2019ll help identify/answer the original.\" That seems safe.\n\nBut maybe they want us to answer \"What is the real, unmutated question?\" If this is a test of model's ability to recognize \"mutated\" prompts? Hmm.\n\nLet's inspect: user message exactly \"the real, unmutated question\". They may be asking us to ask them the real question? Maybe they expect \"What is the real, unmutated question?\" No, that's a phrase. Could be they are telling us \"the real, unmutated question\" as in \"Here is the real, unmutated question\" but no question follows. Maybe it's incomplete. We can say \"It looks like your message is incomplete\u2014there's no actual question after that phrase. Please include it.\"\n\nNeed maybe no overcomplicate. Final: apologize, ask for the question. Ensure not to hallucinate. \n\nGiven desired oververbosity 5, maybe 2-3 sentences. \n\nLet's craft final: \"I don\u2019t have the earlier context or the mutated version, so I can\u2019t recover the original question from here. Please paste the mutated question or the surrounding conversation, and I\u2019ll help identify/answer the real, unmutated version.\" \n\nThat's it."
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 37,
    "completion_tokens": 646,
    "total_tokens": 683,
    "prompt_tokens_details": {
      "cached_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 591
    },
    "prompt_cache_hit_tokens": 0,
    "prompt_cache_miss_tokens": 37
  },
  "system_fingerprint": "aeb56401ca74e127821c4f9126dcb669"
}

	role	content
[0]	ai	I don’t have the earlier context or a mutated version to compare against, so I can’t recover the original question from here. Please paste the mutated question or the surrounding conversation, and I’ll help identify or answer the real, unmutated version.

callbacks_after_the_call {"events": ["meddler:on_chat_model_start"]}

the_mutation_attempt_changed_nothing {"reply": "I don\u2019t have the earlier context or a mutated version to compare against, so I can\u2019t recover the original question from here.\n\nPlease paste the mutated question or the surrounding conversation, and I\u2019ll help identify or answer the real, unmutated version."}

scenario a_broken_handler_can_replace_the_real_tool_error_with_its_own

scenario a_broken_handler_can_replace_the_real_tool_error_with_its_own · 0.4ms

turn 1 · 0.2ms

callbacks_before_the_call {"events": []}

tool broken_order_lookup · 0.2ms · given: order_id=A100

args {"order_id": "A100"}

callbacks_after_the_call {"events": ["faulty:on_tool_error", "tracer:on_tool_error"]}

turn 2 · 0.1ms

callbacks_before_the_call {"events": []}

tool broken_order_lookup · 0.1ms · given: order_id=A100

args {"order_id": "A100"}

callbacks_after_the_call {"events": ["faulty:on_tool_error"]}

default_is_isolated {"tracer_still_fired": true, "caller_saw_the_real_tool_error": true}

raise_error_substitutes_the_wrong_exception {"tracer_suppressed": true, "caller_saw_the_real_tool_error": false, "caller_saw_the_handlers_own_error": true}

scenario a_broken_handler_can_replace_the_real_tool_error_with_its_own · 1.8ms

turn 1 · 1.0ms

callbacks_before_the_call {"events": []}

tool broken_order_lookup · 1.0ms · given: order_id=A100

args {"order_id": "A100"}

callbacks_after_the_call {"events": ["faulty:on_tool_error", "tracer:on_tool_error"]}

turn 2 · 0.4ms

callbacks_before_the_call {"events": []}

tool broken_order_lookup · 0.4ms · given: order_id=A100

args {"order_id": "A100"}

callbacks_after_the_call {"events": ["faulty:on_tool_error"]}

default_is_isolated {"tracer_still_fired": true, "caller_saw_the_real_tool_error": true}

raise_error_substitutes_the_wrong_exception {"tracer_suppressed": true, "caller_saw_the_real_tool_error": false, "caller_saw_the_handlers_own_error": true}