fix(agents): stop instructing agents to ToolSearch built-in tools (#167)

The system-prompt directive layer and the briefing block both opened
with "FIRST ACTION REQUIRED: run ToolSearch to activate deferred
Edit/Write". That premise is false: per Claude Code 2.1.114, ToolSearch
gates only deferred MCP tools, never built-ins — and it is not even a
callable tool in the agent runtime. Built-ins are loaded at spawn via
the `--tools` flag and gated solely by the per-role permission rules
(the actual Edit/Write breakage was the global Write(*)/Edit(*) deny +
single-slash path, fixed in c0ba335). So weak models dutifully chased a
nonexistent ToolSearch, concluded Edit/Write were unavailable, and
rewrote whole files via destructive shell redirection.

Both touch points now affirm the role's built-in tools are loaded and
ready, tell the agent NOT to call ToolSearch, and (for authoring roles)
explicitly steer away from whole-file shell redirection — directly
countering the clobber behaviour. Role prompt files (developer,
cell_pm, main_pm, board) updated to match. Dead
_read_tool_load_from_role_prompt (no callers) removed. Directive tests
rewritten to lock the corrected behaviour.
This commit is contained in:
Renn F
2026-05-16 03:53:52 +02:00
parent c0ba335470
commit 38dba74837
8 changed files with 175 additions and 189 deletions
@@ -1,19 +1,19 @@
"""Smoke-8: briefing's _build_tool_load_block emits the ToolSearch directive
without depending on a "## Load on spawn" section in the role file.
"""#167: the briefing's _build_tool_load_block must not push ToolSearch.
The previous implementation scraped role prompts for that marker; since
no role file had it, the function returned "" and the briefing showed
agents no tool-load instruction. Combined with weak models skipping the
system-prompt directive, the agent went straight to Edit and hit "not
enabled in this context."
Earlier (smoke-8) the block instructed agents to run a ToolSearch call
to "activate deferred built-in tools". That premise was false —
ToolSearch is MCP-only, never gates built-ins, and is not callable in
the agent runtime — so weak models chased a nonexistent tool and fell
back to destructive shell file-writes. The real cause of "Edit not
enabled in this context" was a permission bug fixed separately.
Fix: per-role tool list lives in the orchestrator (mirrors the
factories layer). Pre-renders the directive directly — no file scrape.
The block now affirms the role's built-in tools are loaded and ready,
tells the agent NOT to call ToolSearch, and (for authoring roles)
steers away from whole-file shell redirection.
"""
from __future__ import annotations
import re
from unittest.mock import patch
from roboco.runtime.orchestrator import AgentOrchestrator
@@ -26,58 +26,54 @@ def _orch() -> AgentOrchestrator:
return orch
def test_developer_directive_includes_edit_and_write() -> None:
def _tool_names(block: str) -> list[str]:
"""Exact tool tokens (so 'TodoWrite' is not mistaken for 'Write')."""
line = next(ln for ln in block.splitlines() if "available now:" in ln)
seg = line.split("available now: ", 1)[1]
return seg.split(".", 1)[0].split(", ")
def test_developer_block_affirms_tools_no_toolsearch_call() -> None:
block = _orch()._build_tool_load_block("developer")
assert "First action required" in block
assert "ToolSearch" in block
assert "Edit" in block
assert "Write" in block
assert "Your tools are ready" in block
assert "ToolSearch(query=" not in block
assert "are deferred" not in block
names = _tool_names(block)
assert "Edit" in names and "Write" in names
def test_documenter_directive_includes_edit_and_write() -> None:
block = _orch()._build_tool_load_block("documenter")
assert "Edit" in block
assert "Write" in block
def test_developer_block_steers_away_from_shell_redirection() -> None:
block = _orch()._build_tool_load_block("developer")
assert "shell redirection" in block
assert "Edit/Write" in block
def test_qa_directive_excludes_edit_and_write() -> None:
def test_documenter_block_lists_edit_and_write() -> None:
names = _tool_names(_orch()._build_tool_load_block("documenter"))
assert "Edit" in names and "Write" in names
def test_qa_block_excludes_edit_and_write() -> None:
block = _orch()._build_tool_load_block("qa")
assert "First action required" in block
# The ToolSearch line itself: pull the comma list to be precise about
# which tools are listed (substring match would catch "TodoWrite").
match = re.search(r'select:([^"]+)"', block)
assert match is not None
tools = match.group(1).split(",")
assert "Edit" not in tools
assert "Write" not in tools
assert "Read" in tools
assert "Bash" in tools
assert "Your tools are ready" in block
names = _tool_names(block)
assert "Edit" not in names and "Write" not in names
assert "Read" in names and "Bash" in names
def test_pm_directives_exclude_edit_and_write() -> None:
def test_pm_blocks_exclude_edit_and_write() -> None:
for role in ("main_pm", "cell_pm", "product_owner", "head_marketing", "auditor"):
block = _orch()._build_tool_load_block(role)
match = re.search(r'select:([^"]+)"', block)
assert match is not None
tools = match.group(1).split(",")
assert "Edit" not in tools, f"{role} must not list Edit"
assert "Write" not in tools, f"{role} must not list Write"
names = _tool_names(_orch()._build_tool_load_block(role))
assert "Edit" not in names, f"{role} must not list Edit"
assert "Write" not in names, f"{role} must not list Write"
def test_unknown_role_returns_empty() -> None:
"""No directive for unknown roles (defensive)."""
block = _orch()._build_tool_load_block("nonexistent")
assert block == ""
def test_directive_warns_about_failure_mode() -> None:
block = _orch()._build_tool_load_block("developer")
assert "Edit exists but is not enabled" in block
assert _orch()._build_tool_load_block("nonexistent") == ""
def test_role_cache_works() -> None:
orch = _orch()
first = orch._build_tool_load_block("developer")
# Second call hits the cache (same return).
second = orch._build_tool_load_block("developer")
assert first is second # same string object
assert first is second