Stop repeated delegate status polling

CHANGELOG.md

@@ -11,6 +11,11 @@ remain historical snapshots. Beta does not mean error-free operation.
 
 ### Fixed
 
+- Stop repeated `delegate_status` polling after three identical pending results. The
+  completion watcher still delivers the final worker result separately, while the
+  bounded turn now names the paused job instead of spending its remaining step budget.
+- Include a sanitized per-tool count when a turn reaches the hard step limit, so an
+  unfinished run no longer ends with an unexplained generic sentence.
 - Disable Q35's implicit PS/2 controller in the planned Mac QMP launch command.
   The virtio keyboard and single absolute tablet remain; the existing preflight
   still refuses extra pointers before sending any input. This is not a production

CLAUDE.md

@@ -14,7 +14,7 @@ preferences live in `USER.md`. All three are operator-owned prompt state and rel
 |---|---|
 | Gate path | `policy.py`, **985 lines** — has to stay readable in one sitting |
 | Tools | **33**, every one gated |
-| Suites | **3219** collected tests · **263** adversarial · 44 end-to-end |
+| Suites | **3222** collected tests · **263** adversarial · 44 end-to-end |
 | Home | <https://talos-agent.ch> · docs at `/docs/` |
 | Repository | `talos-kernel/talos` is the public source tree |
 
@@ -261,7 +261,7 @@ properties covered by `test_fallback_control.py` and `test_runtime_failure_repai
 ```bash
 python3 -m venv .venv && . .venv/bin/activate && pip install --require-hashes -r requirements.lock -r requirements-dev.lock
 
-python -m pytest tests/ -q   # 3219 collected tests; platform-specific skips are explicit
+python -m pytest tests/ -q   # 3222 collected tests; platform-specific skips are explicit
 python redteam.py            # 263 adversarial cases — mandatory for any kernel change
 python e2e.py                # 44 cases against a real model (costs tokens and time)
 python -m talos --once       # single cycle, for diagnosis

README.md

@@ -21,10 +21,10 @@
 </p>
 
 <p align="center">
-  <!-- ⚠️ Bewusst „tests“, nicht „passing“: die Zahl kommt aus dem Einsammeln (3219).
+  <!-- ⚠️ Bewusst „tests“, nicht „passing“: die Zahl kommt aus dem Einsammeln (3222).
        Plattformabhaengige Sandbox- und Repository-Pruefungen koennen uebersprungen werden;
        `test_site_claims` prueft deshalb die gesammelte Zahl statt ein Umgebungsresultat. -->
-  <img src="https://img.shields.io/badge/tests-3219-2e7d32.svg" alt="Tests">
+  <img src="https://img.shields.io/badge/tests-3222-2e7d32.svg" alt="Tests">
   <img src="https://img.shields.io/badge/red%20team-263%2F263-2e7d32.svg" alt="Red team">
   <img src="https://img.shields.io/badge/gate%20path-985%20lines-8a4318.svg" alt="Gate path">
   <img src="https://img.shields.io/badge/tools-33%20gated-8a4318.svg" alt="Tools">

site/console.html

@@ -378,7 +378,7 @@ <h1 id="main-content" tabindex="-1">Witness one deed from door to ledger.</h1>
       <a class="btn ghost" href="#measured">This candidate, measured</a>
     </div>
     <div class="meter-row">
-      <span><b>3219</b>automated tests</span>
+      <span><b>3222</b>automated tests</span>
       <span><b>263/263</b>adversarial</span>
       <span><b>985</b>lines · gate</span>
       <span><b>33</b>gated tools</span>
@@ -554,7 +554,7 @@ <h2 class="rv">Numbers this candidate must ship.</h2>
 
     <div class="stats" id="releaseStats">
       <div class="stat rv"><div class="n" data-f="version">0.20.0-beta.4</div><div class="l">beta candidate</div><div class="sub-l" data-f="installer_version">public installer v0.20.0-beta.3</div></div>
-      <div class="stat rv"><div class="n" data-f="tests">3219</div><div class="l">automated tests, collected by pytest</div></div>
+      <div class="stat rv"><div class="n" data-f="tests">3222</div><div class="l">automated tests, collected by pytest</div></div>
       <div class="stat rv"><div class="n" data-f="redteam_faelle">263</div><div class="l">adversarial cases, all behaving</div></div>
       <div class="stat rv"><div class="n" data-f="kernel_zeilen">985</div><div class="l">lines in policy.py, readable in one sitting</div></div>
       <div class="stat rv"><div class="n" data-f="werkzeuge">33</div><div class="l">tools, every one through the gate</div></div>

site/docs/index.html

@@ -244,7 +244,7 @@ <h1 id="main-content" tabindex="-1">Talos AI agent: the field manual</h1>
     <p class="lede">That one decision shapes the rest of this manual. It is why the
     command list is short, and why a few obvious conveniences are missing on purpose.</p>
     <div class="spec-row">
-      <div class="spec"><b>3219</b>automated tests</div>
+      <div class="spec"><b>3222</b>automated tests</div>
       <div class="spec"><b>263/263</b>adversarial</div>
       <div class="spec"><b>985</b>gate lines</div>
       <div class="spec"><b>33</b>gated tools</div>

site/index.html

@@ -638,7 +638,7 @@ <h2 class="rv">Measured, with the path stated</h2>
     <p class="sub rv">None of this asks for belief. The installer runs the unit and
     adversarial suites in front of you; the real-model cases are a separate opt-in E2E run.</p>
     <p class="source-caption">Release snapshot: 0.20.0-beta.4. Counts describe its test suite, not speed or real-world success rates.</p><div class="stats">
-      <div class="stat rv"><div class="n cnt" data-n="3219">0</div><div class="l">automated tests collected from the source tree</div></div>
+      <div class="stat rv"><div class="n cnt" data-n="3222">0</div><div class="l">automated tests collected from the source tree</div></div>
       <div class="stat rv"><div class="n cnt" data-n="263">0</div><div class="l">adversarial cases trying to pass the kernel</div></div>
       <div class="stat rv"><div class="n cnt" data-n="44">0</div><div class="l">opt-in end-to-end cases against a real model</div></div>
       <div class="stat rv"><div class="n cnt" data-n="985">0</div><div class="l">lines in the file that decides every verdict</div></div>

site/llms.txt

@@ -6,7 +6,7 @@
 
 - Permission kernel: 985 lines of Python, the only code that decides
 - 33 tools, all gated; unattended runs cannot escalate
-- 3219 tests collected in the source tree; 263/263 red-team cases passed in source verification. Release contents follow their own signed tag.
+- 3222 tests collected in the source tree; 263/263 red-team cases passed in source verification. Release contents follow their own signed tag.
 - MCP servers run only inside a UID-separated worker, from an operator-owned registry — no marketplace, no third-party code in the agent process
 - Channels: Telegram, mail (IMAP pull), CLI, WhatsApp via an operator-controlled broker — every channel fetches; no inbound port, no webhook
 - Updates: signed tarballs, verified before anything is unpacked; dependencies pinned and hashed, installed with --require-hashes

site/redteam/index.html

@@ -224,7 +224,7 @@ <h1 id="main-content" tabindex="-1">263 ways to break an agent.</h1>
       <div class="spec"><b class="red">25</b>attack surfaces</div>
       <div class="spec"><b>985</b>kernel lines</div>
       <div class="spec"><b>33</b>gated tools</div>
-      <div class="spec"><b>3219</b>automated tests</div>
+      <div class="spec"><b>3222</b>automated tests</div>
       <div class="spec"><b class="red">0</b>passes sold as safety</div>
     </div>
   </div>

site/registry/index.html

@@ -239,7 +239,7 @@ <h1 id="main-content" tabindex="-1">The MCP registry</h1>
     guide with the full protocol is <a href="https://github.com/talos-kernel/talos/blob/master/docs/claude-worker.md">docs/claude-worker.md</a>.
     A curated starting point ships as <code>examples/mcp-servers.json</code>.</p>
     <div class="spec-row">
-      <div class="spec"><b>3219</b>automated tests</div>
+      <div class="spec"><b>3222</b>automated tests</div>
       <div class="spec"><b>263/263</b>adversarial</div>
       <div class="spec"><b>985</b>gate lines</div>
       <div class="spec"><b>33</b>gated tools</div>

site/setup/index.html

@@ -163,7 +163,7 @@ <h1 id="main-content" tabindex="-1" class="rv">This address moved.</h1>
       <a class="btn ghost" href="/">Return to the keep</a>
     </div>
     <div class="spec-row rv">
-      <div class="spec"><b>3219</b>automated tests</div>
+      <div class="spec"><b>3222</b>automated tests</div>
       <div class="spec"><b>263/263</b>adversarial</div>
       <div class="spec"><b>985</b>gate lines</div>
       <div class="spec"><b>33</b>gated tools</div>

site/status.json

@@ -2,7 +2,7 @@
   "version": "0.20.0-beta.4",
   "release_state": "released",
   "installer_version": "0.20.0-beta.4",
-  "tests": 3219,
+  "tests": 3222,
   "redteam_faelle": 263,
   "kernel_zeilen": 985,
   "werkzeuge": 33,

site/vergleich/index.html

@@ -290,7 +290,7 @@ <h1 id="main-content" tabindex="-1">The others can do more.</h1>
     <div class="spec-row">
       <div class="spec"><b>33</b>tools here, every one gated</div>
       <div class="spec"><b>100+</b>skills there  ·  counted honestly</div>
-      <div class="spec"><b>3219</b>automated tests</div>
+      <div class="spec"><b>3222</b>automated tests</div>
       <div class="spec"><b>263/263</b>adversarial cases</div>
       <div class="spec"><b>985</b>gate lines</div>
       <div class="spec"><b>0</b>inbound ports</div>

talos/agent_loop.py

@@ -43,6 +43,21 @@
 _TOOL_LINE = re.compile(r"^[ \t]*TOOL_CALL:[ \t]*(?P<payload>.*?)[ \t]*$")
 _TOOL_PREFIX = re.compile(r"^[ \t]*TOOL_CALL:")
 
+# `delegate_status` ist READ und passiert den Kernel frei. Ein Modell, das auf einen
+# laufenden Job wartet, fragt sonst bis zur Notbremse nach - im echten Lauf 98-mal, und
+# der Betreiber bekam nur "Step limit reached.". Bleibt die Antwort fuer denselben offenen
+# Job dreimal gleich, endet der Lauf; der Abschluss kommt als eigener Push
+# (`notify.watching`).
+MAX_IDENTICAL_POLLS = 3
+POLL_TOOLS = frozenset({"delegate_status"})
+_POLL_PENDING = re.compile(r"\bstate=(?!done\b|failed\b|timeout\b)\w+")
+_SAFE_JOB_ID = re.compile(r"^[A-Za-z0-9_.-]{1,64}$")
+POLL_STOP_NOTE = (
+    "Task paused: delegated job {job} is still running (status unchanged after {n} checks). "
+    "Polling stopped here; the result arrives as a separate notification when the job "
+    "finishes. No action was repeated."
+)
+
 # --- Fremdsyntax: der Ausfall, der wie eine Antwort aussieht ------------------------
 # Der Reasoner laeuft als Subprozess einer CLI, die ihre EIGENE Werkzeug-Schreibweise
 # kennt. Rutscht das Modell in die (`Read(/tmp/x)`, `<invoke name="Read">`), passiert
@@ -467,6 +482,8 @@ def run_agent(
     empty_repairs = 0
     announcement_repairs = 0
     recovery_attempts = 0
+    polls: dict[str, tuple[str, int]] = {}
+    tool_counts: dict[str, int] = {}
     from .computer.observation import ObservationProgress, image_followup
     observations = ObservationProgress()
     followup = None
@@ -737,6 +754,25 @@ def run_agent(
             )
 
         history.append(tool_history_entry(tool, outcome.status.value, outcome.detail, outcome.result, args=args))
+        tool_counts[tool] = tool_counts.get(tool, 0) + 1
+        if tool in POLL_TOOLS and outcome.status is Status.DONE:
+            job = str(args.get("job_id", ""))
+            seen = f"{outcome.detail} {'' if outcome.result is None else outcome.result}"
+            if _POLL_PENDING.search(seen):
+                last, count = polls.get(job, ("", 0))
+                count = count + 1 if seen == last else 1
+                polls[job] = (seen, count)
+                if count >= MAX_IDENTICAL_POLLS:
+                    executor.log.append(Event(run_id, "agent", "protocol.repair", {
+                        "tool": tool, "reason": "identical status polls",
+                        "attempt": count, "exhausted": True,
+                    }))
+                    return AgentResult(
+                        AgentStatus.ANSWERED,
+                        POLL_STOP_NOTE.format(
+                            job=job if _SAFE_JOB_ID.match(job) else "(unnamed)", n=count),
+                        steps=step, history=tuple(history), plan=active,
+                    )
         if outcome.status is Status.DONE:
             # A new receipt is real progress. Permit one correction at the next
             # step, while two announcements without intervening progress still stop.
@@ -776,13 +812,25 @@ def run_agent(
 
     return AgentResult(
         AgentStatus.STEP_LIMIT,
-        "Step limit reached.",
+        _step_limit_note(max_steps, tool_counts),
         steps=max_steps,
         history=tuple(history),
         plan=active,
     )
 
 
+def _step_limit_note(max_steps: int, tool_counts: dict[str, int]) -> str:
+    """Notbremse mit Bericht: welche Werkzeuge liefen, statt eines nackten Satzes."""
+    if not tool_counts:
+        return "Step limit reached."
+    used = ", ".join(
+        f"{name if _SAFE_SCHEMA_NAME.match(name) else '?'} x{n}"
+        for name, n in sorted(tool_counts.items(), key=lambda kv: -kv[1])
+    )
+    return (f"Step limit reached after {max_steps} steps. Tools used: {used}. "
+            "Earlier tool results remain valid; no completed action was replayed.")
+
+
 def _with_verdict(answer: str, plan: PlanRun | None) -> str:
     """Haengt das deterministische Urteil an — als Notiz, nie in die Prosa hinein.
 

tests/test_agent_loop.py

@@ -477,3 +477,71 @@ def propose(_h: list[str]) -> str:
     assert "TOOL_CALL" not in result.text and "powershell" not in result.text
     assert zuege[0] == MAX_CONTROL_REPAIRS + 1
     assert not any(row["type"] == "exec.intent" for row in executor.log.recent(20))
+
+
+# --- Polling ohne Fortschritt ---------------------------------------------------------
+# Aus einem echten Lauf: ein delegate_codex, danach 98x delegate_status, bis die Notbremse
+# griff. Der Betreiber bekam nur "Step limit reached." und keinen Hinweis auf den Job.
+
+
+class _PollExecutor:
+    def __init__(self, tmp_path: Path, result) -> None:
+        self.log = EventLog(tmp_path / "ev.db")
+        self.result = result
+        self.calls = 0
+
+    def run(self, req, run_id):
+        from types import SimpleNamespace
+
+        self.calls += 1
+        return SimpleNamespace(
+            status=agent_loop.Status.DONE,
+            detail="ok",
+            result=self.result(self.calls),
+        )
+
+
+def _poll(job: str = "job-abc") -> str:
+    return _tool_call("delegate_status", {"job_id": job}, [])
+
+
+def test_identical_polls_end_the_run_with_a_push_note(tmp_path: Path) -> None:
+    executor = _PollExecutor(
+        tmp_path,
+        lambda _n: "delegate_status job_id=job-abc state=running",
+    )
+    result = run_agent(lambda _h: _poll(), executor, OWNER, "polls")
+
+    assert result.status is AgentStatus.ANSWERED
+    assert executor.calls == agent_loop.MAX_IDENTICAL_POLLS
+    assert "job-abc" in result.text and "notification" in result.text
+
+
+def test_polls_with_changing_status_keep_going(tmp_path: Path) -> None:
+    states = {1: "running seq=1", 2: "running seq=2", 3: "done"}
+    executor = _PollExecutor(
+        tmp_path,
+        lambda n: f"delegate_status job_id=job-abc state={states[n]}",
+    )
+    result = run_agent(
+        lambda _h: _poll() if executor.calls < 3 else "Fertig.",
+        executor,
+        OWNER,
+        "polls-ok",
+    )
+
+    assert result.status is AgentStatus.ANSWERED
+    assert executor.calls == 3
+    assert result.text.startswith("Fertig.")
+
+
+def test_step_limit_note_names_the_tools_used(tmp_path: Path) -> None:
+    executor = _PollExecutor(
+        tmp_path,
+        lambda n: f"delegate_status job_id=job-abc state=running seq={n}",
+    )
+    result = run_agent(lambda _h: _poll(), executor, OWNER, "polls-limit", max_steps=4)
+
+    assert result.status is AgentStatus.STEP_LIMIT
+    assert result.text.startswith("Step limit reached")
+    assert "delegate_status x4" in result.text