[argus] lead_agent/prompt: add <file_editing> guidance block

The stock system prompt has no edit-vs-rewrite guidance, so models
default to bash heredocs for every file write — including fixes to
files they just wrote in the same conversation. That roughly doubles
wall-time on iterative coding tasks: instead of a 5-line str_replace,
the model emits the entire file again, which has to round-trip through
the LLM and gets persisted in conversation history forever.

The <file_editing> block sits inside <working_directory> (so it's
adjacent to the workspace guidance) and tells the model:

  - Prefer targeted edits (str_replace) over rewrites (write_file) over
    bash heredocs (cat <<EOF).
  - read_file before editing files you wrote earlier in the same
    conversation — context drift makes "expected old content" stale.
  - For deliverables in /mnt/user-data/outputs, write once with
    write_file then str_replace to fix in place; never rewrite.

Adds a test asserting block presence, contents, and placement inside
<working_directory>.

PR-candidate: maybe
Upstream-issue: none
Reason: Opinionated phrasing — upstream may want a more general
        sentence rather than the three-pillar block. Worth proposing
        in a follow-up issue once we have data on whether it actually
        speeds up coding tasks in practice.

Author: Nichol4s

backend/packages/harness/deerflow/agents/lead_agent/prompt.py

@@ -422,6 +422,18 @@ def _build_subagent_section(max_concurrent: int) -> str:
 - Avoid hardcoding `/mnt/user-data/...` inside generated scripts when a relative path from the workspace is enough
 - Final deliverables must be copied to `/mnt/user-data/outputs` and presented using `present_file` tool
 {acp_section}
+<file_editing>
+**Prefer targeted edits over full rewrites.** When modifying a file you wrote earlier in this conversation, the cost of rewriting is real: every byte of the rewritten file has to be re-emitted as model output and lives in conversation history forever. A 5-line fix in a 500-line file should be 5 lines, not 500.
+
+**Tool ordering when changing files:**
+1. `str_replace` — for targeted modifications. Use this whenever you can describe the change as "replace X with Y" in an existing file. Faster than rewriting because only the diff goes through the LLM.
+2. `write_file` — for *new* files, or when the change covers more than ~80% of the file's contents.
+3. `bash` heredocs (`cat <<EOF > file`) — avoid for any file you might edit later. The full content gets baked into conversation history. Reserve heredocs for one-off scripts the agent itself will execute and discard.
+
+**Before editing a file you wrote earlier in the same conversation, call `read_file` first.** Don't rely on memory of what you wrote — context can drift. `read_file` is cheap; `str_replace` failing because of a stale "expected old content" is expensive.
+
+**For deliverables in `/mnt/user-data/outputs`:** write the file once with `write_file`. If you find a bug after writing, use `str_replace` to fix in place. Do not re-run a HEREDOC `cat > ...` to rewrite the whole file.
+</file_editing>
 </working_directory>
 
 <response_style>

backend/tests/test_lead_agent_prompt.py

@@ -68,6 +68,38 @@ def test_apply_prompt_template_includes_relative_path_guidance(monkeypatch):
     assert "`hello.txt`, `../uploads/data.csv`, and `../outputs/report.md`" in prompt
 
 
+def test_apply_prompt_template_includes_file_editing_block(monkeypatch):
+    """The <file_editing> block must be present in the rendered system prompt
+    so the model gets edit-vs-rewrite guidance. Without it, models default
+    to bash heredocs for every file write — including fixes to files they
+    just wrote — which roughly doubles wall-time on iterative coding tasks.
+    """
+    config = SimpleNamespace(
+        sandbox=SimpleNamespace(mounts=[]),
+        skills=SimpleNamespace(container_path="/mnt/skills"),
+    )
+    monkeypatch.setattr("deerflow.config.get_app_config", lambda: config)
+    monkeypatch.setattr(prompt_module, "_get_enabled_skills", lambda: [])
+    monkeypatch.setattr(prompt_module, "get_deferred_tools_prompt_section", lambda: "")
+    monkeypatch.setattr(prompt_module, "_build_acp_section", lambda: "")
+    monkeypatch.setattr(prompt_module, "_get_memory_context", lambda agent_name=None: "")
+    monkeypatch.setattr(prompt_module, "get_agent_soul", lambda agent_name=None: "")
+
+    prompt = prompt_module.apply_prompt_template()
+
+    # Block delimiters and the three pillar guidances must all be present
+    assert "<file_editing>" in prompt
+    assert "</file_editing>" in prompt
+    assert "Prefer targeted edits over full rewrites" in prompt
+    assert "`str_replace`" in prompt
+    assert "Before editing a file you wrote earlier in the same conversation" in prompt
+    # The block must live inside <working_directory>...</working_directory>
+    wd_open = prompt.index("<working_directory")
+    wd_close = prompt.index("</working_directory>")
+    fe_open = prompt.index("<file_editing>")
+    assert wd_open < fe_open < wd_close
+
+
 def test_refresh_skills_system_prompt_cache_async_reloads_immediately(monkeypatch, tmp_path):
     def make_skill(name: str) -> Skill:
         skill_dir = tmp_path / name