fix(prompt): restore two dropped qualifiers from the reference Doing-tasks section

src/context_system/prompt_assembly.py

@@ -877,12 +877,33 @@ def _emit_group(
     "- In general, do not propose changes to code you haven't read. "
     "If a user asks about or wants you to modify a file, read it first. "
     "Understand existing code before suggesting modifications.\n"
-    "- Do not create files unless they're absolutely necessary. "
-    "Generally prefer editing an existing file to creating a new one.\n"
+    # Both qualifiers below are load-bearing and were dropped in the port
+    # (reference: typescript/src/constants/prompts.ts "Doing tasks"). Without
+    # "for achieving your goal" the bullet reads as a flat prohibition on
+    # creating files, and without the rationale it loses the scope that makes
+    # it a bloat-avoidance rule rather than a ban. Cost of the truncation,
+    # observed on terminal-bench 2.1 (torch-tensor-parallelism, 2026-07-25):
+    # the task said "Create the file /app/parallel_linear.py", /app was empty,
+    # the container had no Python to verify against — and the agent spent all
+    # 8 of its steps hunting for an interpreter and never wrote the file.
+    # Claude Code wrote it on its 5th action and scored 1.0.
+    "- Do not create files unless they're absolutely necessary for achieving "
+    "your goal. Generally prefer editing an existing file to creating a new "
+    "one, as this prevents file bloat and builds on existing work more "
+    "effectively.\n"
+    # Final clause restored from the reference, which ends this bullet with
+    # "Escalate to the user with <AskUserQuestion> only when you're genuinely
+    # stuck after investigation, not as a first response to friction."
+    # Deliberately phrased without naming the tool: it is unregistered on the
+    # headless surface (nothing there can answer), and a prompt that names an
+    # unavailable tool is its own failure mode \u2014 fix-git, 2026-07-25, asked
+    # twice and then handed the task back to a user who did not exist.
     "- If an approach fails, diagnose why before switching tactics\u2014read "
     "the error, check your assumptions, try a focused fix. Don't retry "
     "the identical action blindly, but don't abandon a viable approach "
-    "after a single failure either.\n"
+    "after a single failure either. Escalate to the user only when you're "
+    "genuinely stuck after investigation, not as a first response to "
+    "friction.\n"
     "- For ambiguous bugs, failures, or unfamiliar data, form a small set "
     "of plausible hypotheses from the available evidence and rank them by "
     "likelihood and the cost of checking them. Start with the cheapest "

tests/test_system_prompt_full.py

@@ -87,6 +87,42 @@ def test_task_prompt_groups_related_mechanical_edits(self):
         assert "long sequence of tiny edits" in prompt
         assert "Keep unrelated or semantically distinct changes separate" in prompt
 
+    def test_task_prompt_keeps_the_reference_file_creation_qualifiers(self):
+        """The file-creation bullet must carry BOTH reference qualifiers.
+
+        Reference (typescript/src/constants/prompts.ts, "Doing tasks"):
+        "Do not create files unless they're absolutely necessary for
+        achieving your goal. Generally prefer editing an existing file to
+        creating a new one, as this prevents file bloat and builds on
+        existing work more effectively."
+
+        The port dropped "for achieving your goal" and the rationale,
+        leaving what reads as a flat ban on new files. That cost a
+        terminal-bench 2.1 task outright (torch-tensor-parallelism,
+        2026-07-25): the instruction was "Create the file
+        /app/parallel_linear.py", /app was empty, no interpreter existed to
+        verify against, and the agent spent every step hunting for one
+        instead of writing the deliverable it had been asked for.
+        """
+        prompt = build_full_system_prompt(use_cache=False)
+        assert "absolutely necessary for achieving your goal" in prompt, (
+            "dropping 'for achieving your goal' turns a bloat-avoidance rule "
+            "into a prohibition on producing a requested file"
+        )
+        assert "prevents file bloat" in prompt
+
+    def test_task_prompt_scopes_escalation_to_genuine_blockage(self):
+        """Restored from the reference's "If an approach fails" bullet.
+
+        The named tool is omitted on purpose: AskUserQuestion is
+        unregistered on the headless surface, and a prompt that names an
+        unavailable tool is its own failure mode (fix-git, 2026-07-25 —
+        asked twice, then handed the task back to a user who did not exist).
+        """
+        prompt = build_full_system_prompt(use_cache=False)
+        assert "genuinely stuck after investigation" in prompt
+        assert "not as a first response to friction" in prompt
+
     def test_identity_prompt_backward_compat(self):
         """_IDENTITY_PROMPT is an alias for _INTRO_SECTION."""
         assert _IDENTITY_PROMPT is _INTRO_SECTION