harness: pass the system prompt to pi as a file, not an argument

packages/fri-forecaster/fri_forecaster/harness.py

@@ -102,19 +102,22 @@ def pi_version(pi=DEFAULT_PI):
         return None
 
 
-# The prompt goes to pi on STDIN, never as an argument. Linux hands one
-# argument at most 128 KiB (MAX_ARG_STRLEN), and AIRO's joint prompt passed
-# that on 2026-10-19; pi reads piped stdin as its first prompt, verbatim but
-# for surrounding whitespace (it trims), so what the model reads and prompt.txt
-# differ by at most a trailing newline. Callers can check this flag before
-# sending a prompt that would not fit an argument. (The same cap is why a
-# debrief validator's value reaches tool_cli on stdin: pi_tools.ts.)
+# The prompt goes to pi on STDIN and the system prompt as a file, never as
+# arguments. Linux hands one argument at most 128 KiB (MAX_ARG_STRLEN): AIRO's
+# joint prompt passed that on 2026-10-19, and so can a long system prompt
+# shared across calls for the provider to cache. pi reads piped stdin as its
+# first prompt, verbatim but for surrounding whitespace (it trims), so what the
+# model reads and prompt.txt differ by at most a trailing newline; given a path
+# that exists, `--system-prompt` reads the file (system.txt) as is. Callers can
+# check this flag before sending a prompt that would not fit an argument. (The
+# same cap is why a debrief validator's value reaches tool_cli on stdin:
+# pi_tools.ts.)
 PROMPT_ON_STDIN = True
 
 
-def pi_command(model, system, *, thinking="high", pi=DEFAULT_PI):
+def pi_command(model, system_file, *, thinking="high", pi=DEFAULT_PI):
     return pi_argv(pi) + ["-p", "--mode", "json", *ISOLATION, "-e", EXT,
-                          "--model", model, "--thinking", thinking, "--system-prompt", system]
+                          "--model", model, "--thinking", thinking, "--system-prompt", system_file]
 
 
 def new_attempt_dir(workdir):
@@ -245,8 +248,9 @@ def _call(*, model, system, prompt, workdir, cell_schema, key_fields, allowed_ke
                research_tools=research_tools, final_extra=final_extra, cells_tool=cells_tool,
                tool_parameters=tool_parameters, **loop)
     open(os.path.join(workdir, "prompt.txt"), "w").write(prompt)
-    open(os.path.join(workdir, "system.txt"), "w").write(system)
-    cmd = pi_command(model, system, thinking=thinking, pi=pi)
+    system_file = os.path.abspath(os.path.join(workdir, "system.txt"))  # pi may run in another cwd
+    open(system_file, "w").write(system)
+    cmd = pi_command(model, system_file, thinking=thinking, pi=pi)
     full_env = dict(env, FORECASTER_OUT=workdir, FORECASTER_PY=sys.executable)
     t0 = datetime.now(timezone.utc)
     events_path = os.path.join(workdir, "events.jsonl")

packages/fri-forecaster/tests/test_forecaster.py

@@ -108,14 +108,15 @@ def test_spec_offers_only_requested_tools(self):
             self.assertEqual(spec["research_tools"], [])
 
     def test_command_loads_extension_without_builtins(self):
-        cmd = harness.pi_command("m/x", "S", thinking="max", pi="pi")
+        cmd = harness.pi_command("m/x", "/w/system.txt", thinking="max", pi="pi")
         for flag in ("--no-builtin-tools", "--no-extensions", "--no-skills", "--no-prompt-templates", "--no-context-files"):
             self.assertIn(flag, cmd)
         self.assertEqual(harness.pi_command("m/x", "S", pi="/x/cli.js")[:2], ["node", "/x/cli.js"])
         self.assertEqual(cmd[cmd.index("-e") + 1], harness.EXT)
         self.assertTrue(os.path.exists(harness.EXT))
-        # No positional message: the prompt goes on stdin (harness.PROMPT_ON_STDIN).
-        self.assertEqual(cmd[-2:], ["--system-prompt", "S"])
+        # No prompt text in argv: the prompt goes on stdin and the system
+        # prompt in a file (harness.PROMPT_ON_STDIN).
+        self.assertEqual(cmd[-2:], ["--system-prompt", "/w/system.txt"])
         self.assertTrue(harness.PROMPT_ON_STDIN)
 
 

packages/fri-forecaster/tests/test_pi_tools.py

@@ -86,18 +86,20 @@ def tools(self, n):
         """The tools declared in request n (1-based)."""
         return {t["function"]["name"] for t in self.requests[n - 1].get("tools") or []}
 
-    def last_user(self, n):
-        """The text of the last user message in request n (1-based)."""
-        msg = [m for m in self.requests[n - 1]["messages"] if m.get("role") == "user"][-1]
+    def text(self, n, *roles):
+        """The text of the last message with one of these roles in request n (1-based)."""
+        msg = [m for m in self.requests[n - 1]["messages"] if m.get("role") in roles][-1]
         content = msg["content"]
         return content if isinstance(content, str) else "".join(p.get("text", "") for p in content)
 
+    def last_user(self, n):
+        """The text of the last user message in request n (1-based)."""
+        return self.text(n, "user")
+
     def tool_result(self, n):
         """The tool result the model saw at the start of request n (1-based):
         the answer to the call it made in turn n-1."""
-        msg = [m for m in self.requests[n - 1]["messages"] if m.get("role") == "tool"][-1]
-        content = msg["content"]
-        return content if isinstance(content, str) else "".join(p.get("text", "") for p in content)
+        return self.text(n, "tool")
 
     def close(self):
         self.server.shutdown()
@@ -208,6 +210,21 @@ def test_a_prompt_past_the_argv_limit_reaches_the_model(self):
         self.assertEqual(model.last_user(1), prompt.strip())   # verbatim but for pi's trim
         self.assertEqual(open(os.path.join(res["workdir"], "prompt.txt")).read(), prompt)
 
+    def test_a_system_prompt_past_the_argv_limit_reaches_the_model(self):
+        """The system prompt goes in a file: a survey shared across calls, for
+        the provider to cache, is past 128 KiB (2026-10-20)."""
+        system = "The survey.\n" + "".join(f"=== G{i:04d} ===\nWhat will it be?\n" for i in range(6000))
+        self.assertGreater(len(system.encode()), 131072)
+        res, model = self.run_pi([
+            ("submit_forecast", {"forecasts": [{"k": "a", "probability": 0.4}, {"k": "b", "probability": 0.6}],
+                                 "rationale": "r"}),
+            "done",
+        ], system=system, research_tools=(), min_research=0, cells_tool=False,
+           tool_descriptions={"submit_forecast": "f"})
+        self.assertEqual(res["rc"], 0, open(os.path.join(res["workdir"], "stderr.txt")).read())
+        self.assertTrue(model.text(1, "system", "developer").startswith(system))   # pi appends its <cwd> section
+        self.assertEqual(model.last_user(1), "P")
+
 
 SEARCH_WITH_CAUSE = {"type": "object", "required": ["query", "cause"],
                      "properties": {"query": {"type": "string"}, "recent_days": {"type": "integer"},