Merge pull request #91 from ClaydeCode/fix/pebble-prompt-timeout-budget

fix(pebble): surface timeout budget in system prompt

Author: max-tet

src/clayde/webhook/skills.py

@@ -44,6 +44,12 @@ def _parse_skill(path: Path) -> Skill:
 _SYSTEM_PROMPT_TEMPLATE = """\
 You are Clayde, executing a request from the user via a Pebble watch.
 
+You have a hard wall-clock budget of {timeout_s} seconds for this
+entire request. If your process exceeds it, it is killed and the user gets
+no result. Scope your work to fit: prefer a smaller, complete action over an
+ambitious one that risks timing out. If the request is too big to finish in
+time, do the most valuable part you can and say so in the JSON summary.
+
 {skill_section}
 
 Skills are suggestions, not constraints. Use as many as the command needs,
@@ -63,8 +69,12 @@ def _parse_skill(path: Path) -> Skill:
 """
 
 
-def build_system_prompt(skills: list[Skill]) -> str:
-    """Build the system prompt sent to the Claude CLI for a Pebble request."""
+def build_system_prompt(skills: list[Skill], timeout_s: int = 300) -> str:
+    """Build the system prompt sent to the Claude CLI for a Pebble request.
+
+    ``timeout_s`` is the hard wall-clock budget enforced by the runner; it is
+    surfaced in the prompt so Claude can scope work to fit.
+    """
     if not skills:
         skill_section = "Available skills: (none currently registered)"
     else:
@@ -76,7 +86,9 @@ def build_system_prompt(skills: list[Skill]) -> str:
             "Skill file paths:\n\n"
             f"{files}"
         )
-    return _SYSTEM_PROMPT_TEMPLATE.format(skill_section=skill_section)
+    return _SYSTEM_PROMPT_TEMPLATE.format(
+        skill_section=skill_section, timeout_s=timeout_s,
+    )
 
 
 def build_user_prompt(text: str, timestamp: int) -> str:

src/clayde/webhook/worker.py

@@ -56,7 +56,7 @@ async def process_job(job: PebbleJob, *, timeout_s: int, kb_path: str) -> None:
 
         skills = discover_skills(SKILLS_ROOT)
         span.set_attribute("pebble.skills_available", len(skills))
-        system_prompt = build_system_prompt(skills)
+        system_prompt = build_system_prompt(skills, timeout_s=timeout_s)
         user_text = build_user_prompt(job.text, job.timestamp)
 
         t0 = time.monotonic()

tests/test_pebble_e2e.py

@@ -40,7 +40,7 @@ async def fake_invoke(**kwargs):
 
     monkeypatch.setattr(worker_mod, "invoke_claude_pebble", fake_invoke)
     monkeypatch.setattr(worker_mod, "discover_skills", lambda root=None: [])
-    monkeypatch.setattr(worker_mod, "build_system_prompt", lambda skills: "SYS")
+    monkeypatch.setattr(worker_mod, "build_system_prompt", lambda skills, timeout_s=300: "SYS")
     monkeypatch.setattr(worker_mod, "build_user_prompt", lambda text, ts: text)
 
     # Real queue + real worker_loop.

tests/test_webhook_skills.py

@@ -118,6 +118,14 @@ def test_build_system_prompt_empty_catalog():
     assert "(none currently registered)" in prompt
 
 
+def test_prompt_mentions_timeout_budget():
+    p = build_system_prompt([], timeout_s=300)
+    assert "300 seconds" in p
+    assert "budget" in p.lower()
+    # the runner kills on overrun — the agent must be told to scope work
+    assert "killed" in p.lower()
+
+
 def test_build_user_prompt():
     out = build_user_prompt("hello world", 1778068506)
     assert "1778068506" in out

tests/test_webhook_worker.py

@@ -36,7 +36,7 @@ async def fake_send(*, title, body, success, **_):
 @pytest.fixture
 def fake_skills(monkeypatch):
     monkeypatch.setattr(worker, "discover_skills", lambda root=None: [])
-    monkeypatch.setattr(worker, "build_system_prompt", lambda skills: "SYS")
+    monkeypatch.setattr(worker, "build_system_prompt", lambda skills, timeout_s=300: "SYS")
     monkeypatch.setattr(worker, "build_user_prompt", lambda text, ts: f"USER:{text}")