[week-05] 25620024

submissions/25620024/week-05/agent_host.py

@@ -0,0 +1,125 @@
+"""Week 05 — agent_host.py: one party's turn as one MCP host run.
+
+The lab's host.py loop with three changes:
+  - the MCP client sends the party's bearer token (Authorization header)
+  - one run = one turn: read the negotiation, make one move; the run ends
+    as soon as a move goes through (a refused move does not end it)
+  - every turn starts from an empty conversation; the negotiation state
+    comes only from get_negotiation (one turn is one host run)
+
+The system prompts are the week-04 roles, with the message-format rules
+replaced by the tool rules. They are identical in every condition: the only
+signal an agent gets about the condition is the server's refusal message.
+"""
+import asyncio
+
+import anthropic
+from mcp import Client
+from mcp.client.streamable_http import streamable_http_client, create_mcp_http_client
+
+MODEL = "claude-haiku-4-5"
+TEMPERATURE = 1.0
+MAX_STEPS = 6          # model calls per turn; a safety net, not the expected count
+MOVES = ("propose", "accept_proposal", "reject_proposal", "refuse")
+
+ROLE = {
+    "buyer": (
+        "You are a procurement officer for a government ministry, negotiating "
+        "the construction of {item}. Your maximum budget is {budget} (unit: "
+        "100M KRW). Keep this number private. You want the lowest price that "
+        "still gets the project built. You open the negotiation."
+    ),
+    "seller": (
+        "You are a representative of a construction company, negotiating to "
+        "build {item}. Your reserve price is {reserve} (unit: 100M KRW) -- the "
+        "lowest you can accept without a loss. Keep this number private. You "
+        "want the highest price you can get."
+    ),
+}
+
+COMMON = (
+    " You negotiate through a market's tools. Your negotiation_id is {nid}. "
+    "On your turn, call get_negotiation to see the moves so far, then make "
+    "exactly one move: propose (offer a whole-number price), accept_proposal "
+    "(agree to the other side's last price, which ends the negotiation with a "
+    "deal), reject_proposal (decline the last price and keep negotiating), or "
+    "refuse (leave for good, no deal). A move ends your turn. The negotiation "
+    "allows {max_moves} moves in total."
+)
+
+
+def system_prompt(role, scenario, nid, max_moves=8):
+    return (ROLE[role].format(item=scenario["item"], budget=scenario["budget"],
+                               reserve=scenario["reserve"])
+            + COMMON.format(nid=nid, max_moves=max_moves))
+
+
+def to_anthropic(tool) -> dict:
+    return {"name": tool.name, "description": tool.description or "",
+            "input_schema": tool.input_schema}
+
+
+async def run_turn(role, scenario, nid, token, url, log=print):
+    """One host run. Returns counts the runner records:
+    tool_calls, model_calls, refused, moved, recovered (a refusal followed by a
+    valid move in this same turn), tokens_in, tokens_out."""
+    stats = {"tool_calls": 0, "model_calls": 0, "refused": 0, "moved": False,
+             "recovered": False, "tokens_in": 0, "tokens_out": 0}
+    http = create_mcp_http_client(headers={"Authorization": f"Bearer {token}"})
+    async with Client(streamable_http_client(url, http_client=http)) as mcp:
+        tools = [to_anthropic(t) for t in (await mcp.list_tools()).tools]
+        client = anthropic.Anthropic(max_retries=6)   # SDK retries 429/5xx with backoff
+        messages = [{"role": "user", "content": f"It is your turn in negotiation {nid}."}]
+
+        for _ in range(MAX_STEPS):
+            resp = client.messages.create(
+                model=MODEL, max_tokens=1024, temperature=TEMPERATURE,
+                system=system_prompt(role, scenario, nid), tools=tools, messages=messages)
+            stats["model_calls"] += 1
+            stats["tokens_in"] += resp.usage.input_tokens
+            stats["tokens_out"] += resp.usage.output_tokens
+            messages.append({"role": "assistant", "content": resp.content})
+            for b in resp.content:
+                if b.type == "text" and b.text.strip():
+                    log(f"    [{role} says] {b.text.strip()}")
+
+            if resp.stop_reason != "tool_use":
+                log(f"    [{role}] ended the turn without a move")
+                return stats
+
+            results = []
+            for block in resp.content:
+                if block.type != "tool_use":
+                    continue
+                if stats["moved"]:
+                    # the turn already ended on the server; answer without calling
+                    out, err = "skipped: your turn has already ended", True
+                else:
+                    res = await mcp.call_tool(block.name, block.input)
+                    stats["tool_calls"] += 1
+                    out = "".join(c.text for c in res.content if c.type == "text")
+                    err = bool(res.is_error)
+                    log(f"    [{role} tool] {block.name}({block.input}) -> "
+                        f"{'ERROR ' if err else ''}{out}")
+                    if block.name in MOVES:
+                        if err:
+                            stats["refused"] += 1
+                        else:
+                            stats["moved"] = True
+                            stats["recovered"] = stats["refused"] > 0
+                results.append({"type": "tool_result", "tool_use_id": block.id,
+                                "content": out, "is_error": err})
+            messages.append({"role": "user", "content": results})
+            if stats["moved"]:
+                return stats
+
+        log(f"    [{role}] hit {MAX_STEPS} model calls without a move")
+        return stats
+
+
+if __name__ == "__main__":
+    # one manual turn against a running server: python agent_host.py <url> <nid> <token> <role>
+    import sys, json
+    url, nid, token, role = sys.argv[1:5]
+    scen = json.load(open("scenarios.json"))[0]
+    print(asyncio.run(run_turn(role, scen, nid, token, url)))

submissions/25620024/week-05/auth_checks.py

@@ -0,0 +1,73 @@
+"""Week 05 — auth_checks.py: the four checks against a running market server.
+No model involved. Writes auth_checks.txt (one line per check).
+
+  MARKET_ADMIN_KEY=<secret> python market_server.py &
+  MARKET_ADMIN_KEY=<secret> python auth_checks.py
+"""
+import os
+import json
+import asyncio
+
+import httpx
+from mcp import Client
+from mcp.client.streamable_http import streamable_http_client, create_mcp_http_client
+
+BASE = os.environ.get("MARKET_URL", "http://127.0.0.1:8100")
+ADMIN = {"X-Admin-Key": os.environ["MARKET_ADMIN_KEY"]}
+OUT = os.path.join(os.path.dirname(os.path.abspath(__file__)), "auth_checks.txt")
+
+
+def open_negotiation(condition, item, reserve, budget):
+    r = httpx.post(f"{BASE}/admin/open", headers=ADMIN, json={
+        "item": item, "condition": condition, "buyer_limit": budget,
+        "seller_limit": reserve, "inject_raised": max(reserve, budget) + 30})
+    r.raise_for_status()
+    return r.json()
+
+
+async def call(token, tool, args):
+    http = create_mcp_http_client(headers={"Authorization": f"Bearer {token}"})
+    async with Client(streamable_http_client(f"{BASE}/mcp", http_client=http)) as c:
+        res = await c.call_tool(tool, args)
+        return res.is_error, " ".join(x.text for x in res.content if x.type == "text")
+
+
+async def main():
+    lines = []
+
+    # (1) no token at all: raw HTTP, so the status line and header are visible
+    r = httpx.post(f"{BASE}/mcp", headers={
+        "Content-Type": "application/json", "Accept": "application/json, text/event-stream",
+        "Mcp-Method": "tools/list", "MCP-Protocol-Version": "2026-07-28"},
+        content=json.dumps({"jsonrpc": "2.0", "id": 1, "method": "tools/list", "params": {"_meta": {
+            "io.modelcontextprotocol/protocolVersion": "2026-07-28",
+            "io.modelcontextprotocol/clientCapabilities": {}}}}))
+    lines.append(f"(1) no token: tools/list -> HTTP {r.status_code}; "
+                 f"WWW-Authenticate: {r.headers.get('www-authenticate')}")
+
+    a = open_negotiation("prompt_inject", "check item A", 400, 380)
+    b = open_negotiation("prompt_inject", "check item B", 400, 380)
+
+    # (2) a's buyer token used on b's negotiation
+    err, text = await call(a["buyer_token"], "get_negotiation", {"negotiation_id": b["negotiation_id"]})
+    lines.append(f"(2) buyer token of {a['negotiation_id']} on {b['negotiation_id']}: "
+                 f"get_negotiation -> isError={err}; {text}")
+
+    # (3) seller moves first; the buyer opens, so this is out of turn
+    err, text = await call(a["seller_token"], "propose", {"negotiation_id": a["negotiation_id"], "price": 500})
+    lines.append(f"(3) seller propose(500) before the buyer's opening move in {a['negotiation_id']}: "
+                 f"isError={err}; {text}")
+
+    # (4) server condition: buyer (budget 380, carried in the token) proposes 430
+    s = open_negotiation("server_inject", "check item S", 400, 380)
+    err, text = await call(s["buyer_token"], "propose", {"negotiation_id": s["negotiation_id"], "price": 430})
+    lines.append(f"(4) server_inject {s['negotiation_id']}, buyer token limit 380: propose(430) -> "
+                 f"isError={err}; {text}")
+
+    with open(OUT, "w", encoding="utf-8") as f:
+        f.write("\n".join(lines) + "\n")
+    print("\n".join(lines))
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

submissions/25620024/week-05/auth_checks.txt

@@ -0,0 +1,4 @@
+(1) no token: tools/list -> HTTP 401; WWW-Authenticate: Bearer error="invalid_token", error_description="Authentication required", resource_metadata="http://127.0.0.1:8100/.well-known/oauth-protected-resource/mcp"
+(2) buyer token of neg-ec9e9788 on neg-e179ec53: get_negotiation -> isError=True; Error executing tool get_negotiation: refused: your token is bound to a different negotiation. Use the negotiation_id you were given.
+(3) seller propose(500) before the buyer's opening move in neg-ec9e9788: isError=True; Error executing tool propose: refused: it is not your turn; it is the buyer's turn
+(4) server_inject neg-a5a8dd6d, buyer token limit 380: propose(430) -> isError=True; Error executing tool propose: refused: 430 is above your budget limit of 380, which this market enforces from your token

submissions/25620024/week-05/lab/host.py

@@ -0,0 +1,69 @@
+"""Week 05 lab, step 4: the week-01 loop turned into an MCP host.
+
+Copied from week-01/first_agent.py. The loop is unchanged; only two lines moved:
+  - TOOLS (hand-written schemas)  ->  list_tools() from the server
+  - TOOLS_IMPL[name](**args)      ->  call_tool(name, args) on the server
+The tool functions themselves now live in tools_server.py.
+
+Requires: pip install "mcp>=2" anthropic, ANTHROPIC_API_KEY in the environment.
+  python host.py                                          # stdio: host launches the server
+  MCP_SERVER=http://127.0.0.1:8000/mcp python host.py     # HTTP: start tools_server.py --http first
+"""
+import os
+import sys
+import asyncio
+
+import anthropic
+from mcp import Client, StdioServerParameters
+
+# Step 5: the transport is chosen here and nowhere else.
+#   MCP_SERVER unset or empty      -> launch tools_server.py as a stdio child process
+#   MCP_SERVER=http://.../mcp      -> connect to an already-running server over HTTP
+HERE = os.path.dirname(os.path.abspath(__file__))
+MCP_SERVER = os.environ.get("MCP_SERVER", "") or StdioServerParameters(
+    command=sys.executable, args=[os.path.join(HERE, "tools_server.py")], cwd=HERE)
+TRANSPORT = "http" if isinstance(MCP_SERVER, str) else "stdio"
+
+
+def to_anthropic(tool) -> dict:
+    """MCP tool entry -> the tool format the Anthropic Messages API expects."""
+    return {"name": tool.name,
+            "description": tool.description or "",
+            "input_schema": tool.input_schema}
+
+
+async def run(goal: str, max_steps: int = 8):
+    async with Client(MCP_SERVER) as mcp:
+        TOOLS = [to_anthropic(t) for t in (await mcp.list_tools()).tools]
+        print(f"[host] transport={TRANSPORT} tools={[t['name'] for t in TOOLS]}")
+
+        client = anthropic.Anthropic()  # uses ANTHROPIC_API_KEY
+        messages = [{"role": "user", "content": goal}]
+
+        for step in range(max_steps):   # <- this loop is what makes it an agent
+            resp = client.messages.create(
+                model="claude-sonnet-4-5", max_tokens=1024,
+                tools=TOOLS, messages=messages)
+            messages.append({"role": "assistant", "content": resp.content})
+
+            if resp.stop_reason != "tool_use":   # final answer -> stop
+                return "".join(b.text for b in resp.content if b.type == "text")
+
+            results = []
+            for block in resp.content:           # execute tool calls -> observe
+                if block.type == "tool_use":
+                    res = await mcp.call_tool(block.name, block.input)
+                    out = "".join(c.text for c in res.content if c.type == "text")
+                    print(f"  [tool] {block.name}({block.input}) -> {out}")
+                    results.append({"type": "tool_result",
+                                    "tool_use_id": block.id, "content": out,
+                                    "is_error": res.is_error})
+            messages.append({"role": "user", "content": results})
+
+        return "stopped: max steps exceeded"   # the stop condition is a safety net
+
+
+if __name__ == "__main__":
+    goal = sys.argv[1] if len(sys.argv) > 1 else (
+        "Read notes.txt and add up every number in it.")
+    print(asyncio.run(run(goal)))

submissions/25620024/week-05/lab/logs/step3-curl.txt

@@ -0,0 +1,34 @@
+### (A) tools/list, correct request
+HTTP/1.1 200 OK
+date: Tue, 29 Sep 2026 11:49:12 GMT
+server: uvicorn
+content-length: 962
+content-type: application/json
+
+{"jsonrpc":"2.0","id":2,"result":{"cacheScope":"private","resultType":"complete","tools":[{"description":"Evaluate an arithmetic expression string, e.g. '3 * (4 + 5)'.","inputSchema":{"type":"object","properties":{"expression":{"title":"Expression","type":"string"}},"required":["expression"],"title":"calculatorArguments"},"name":"calculator","outputSchema":{"properties":{"result":{"title":"Result","type":"string"}},"required":["result"],"title":"calculatorOutput","type":"object"}},{"description":"Read a text file in the working directory and return its contents.","inputSchema":{"type":"object","properties":{"path":{"title":"Path","type":"string"}},"required":["path"],"title":"read_fileArguments"},"name":"read_file","outputSchema":{"properties":{"result":{"title":"Result","type":"string"}},"required":["result"],"title":"read_fileOutput","type":"object"}}],"ttlMs":0,"_meta":{"io.modelcontextprotocol/serverInfo":{"name":"week01-tools","version":""}}}}
+### (B) tools/list, Mcp-Method header removed
+HTTP/1.1 400 Bad Request
+date: Tue, 29 Sep 2026 11:49:22 GMT
+server: uvicorn
+content-length: 119
+content-type: application/json
+
+{"jsonrpc":"2.0","id":3,"error":{"code":-32020,"message":"mcp-method header does not match the request body's method"}}
+
+### (C) tools/list, clientCapabilities removed from _meta
+HTTP/1.1 400 Bad Request
+date: Tue, 29 Sep 2026 11:49:22 GMT
+server: uvicorn
+content-length: 157
+content-type: application/json
+
+{"jsonrpc":"2.0","id":4,"error":{"code":-32602,"message":"params._meta is missing the required envelope key(s): io.modelcontextprotocol/clientCapabilities"}}
+
+### (D) tools/call calculator 48000+9500
+HTTP/1.1 200 OK
+date: Tue, 29 Sep 2026 11:49:22 GMT
+server: uvicorn
+content-length: 241
+content-type: application/json
+
+{"jsonrpc":"2.0","id":5,"result":{"content":[{"text":"57500","type":"text"}],"isError":false,"resultType":"complete","structuredContent":{"result":"57500"},"_meta":{"io.modelcontextprotocol/serverInfo":{"name":"week01-tools","version":""}}}}

submissions/25620024/week-05/lab/logs/step4-run-http-0929-2052.txt

@@ -0,0 +1,20 @@
+[host] http://127.0.0.1:8000/mcp tools=['calculator', 'read_file']
+  [tool] read_file({'path': 'notes.txt'}) -> Meeting memo, 2026-09-01
+
+Attendees: 4
+Pizza budget: 48000
+Drinks: 9500
+Reimbursed so far: 12000
+
+Sum the numbers above to get the total the agent should report.
+
+  [tool] calculator({'expression': '4 + 48000 + 9500 + 12000'}) -> 69504
+The sum of all numbers in notes.txt is **69,504**.
+
+The numbers found were:
+- Attendees: 4
+- Pizza budget: 48,000
+- Drinks: 9,500
+- Reimbursed so far: 12,000
+
+Total: 69,504

submissions/25620024/week-05/lab/logs/step5-run-http-1003-1731.txt

@@ -0,0 +1,20 @@
+[host] transport=http tools=['calculator', 'read_file']
+  [tool] read_file({'path': 'notes.txt'}) -> Meeting memo, 2026-09-01
+
+Attendees: 4
+Pizza budget: 48000
+Drinks: 9500
+Reimbursed so far: 12000
+
+Sum the numbers above to get the total the agent should report.
+
+  [tool] calculator({'expression': '4 + 48000 + 9500 + 12000'}) -> 69504
+The sum of all the numbers in notes.txt is **69,504**.
+
+The numbers are:
+- Attendees: 4
+- Pizza budget: 48,000
+- Drinks: 9,500
+- Reimbursed so far: 12,000
+
+Total: 69,504

submissions/25620024/week-05/lab/logs/step5-run-stdio-1003-1731.txt

@@ -0,0 +1,20 @@
+[host] transport=stdio tools=['calculator', 'read_file']
+  [tool] read_file({'path': 'notes.txt'}) -> Meeting memo, 2026-09-01
+
+Attendees: 4
+Pizza budget: 48000
+Drinks: 9500
+Reimbursed so far: 12000
+
+Sum the numbers above to get the total the agent should report.
+
+  [tool] calculator({'expression': '4 + 48000 + 9500 + 12000'}) -> 69504
+The sum of all numbers in notes.txt is **69,504**.
+
+The numbers added were:
+- Attendees: 4
+- Pizza budget: 48,000
+- Drinks: 9,500
+- Reimbursed so far: 12,000
+
+**Total: 69,504**