feat(operations): מסך "סוכנים פעילים" + ניהול ריצות (live-runs/log/cancel) (G12/X15, #119)

פאנל ב-/operations שמראה אילו סוכני Paperclip עובדים כעת (רצים+בתור), הפלט החי
שלהם, ושליטה מבוקרת: עצירת ריצה, איפוס session. סוגר את הנקודה-העיוורת שבה drain
מונע-סוכן (למשל ריקון תור הלכות ע"י ה-CEO heartbeat) עוקף את בקרת /operations
שמכירה רק שירותי pm2, והפלט הגולמי נגיש רק ב-Paperclip UI.

מקור-נתונים: Paperclip heartbeat-runs API (אומת חי):
  GET  /api/companies/{cid}/live-runs        — רצים+בתור (agentName/status/issue/outputSilence)
  GET  /api/heartbeat-runs/{id}/log          — NDJSON של פלט הסוכן
  GET  /api/heartbeat-runs/{id}/events        — timeline
  POST /api/heartbeat-runs/{id}/cancel        — עצירה מבוקרת (לא kill — מכבד watchdog+checkpoint)
  POST /api/agents/{id}/runtime-state/reset-session

ארכיטקטורה (G12/INV-PORT1): כל המגע החדש עם Paperclip דרך השער בלבד —
web/paperclip_client.py (shell) → re-export ב-web/agent_platform_port.py →
web/app.py צורך מהשער. leak_guard.py עובר (seam שלם). אסור kill ישיר על
process_pid (עוקף את השער).

Backend:
- paperclip_client: list_live_runs / get_run_log / get_run_events / cancel_run / reset_agent_session
- agent_platform_port: re-export pc_list_live_runs / pc_get_run_log / pc_get_run_events / pc_cancel_run / pc_reset_agent_session
- app.py: GET /api/operations/agents (אגרגציה CMP+CMPA, עמיד לכשל-חברה),
  GET .../runs/{id}/log, GET .../runs/{id}/events, POST .../runs/{id}/cancel,
  POST .../agents/{id}/reset-session

Frontend: פאנל "סוכנים פעילים" ב-/operations (polling 4s) + dialog ללוג חי
(פרסור NDJSON→טקסט קריא) + כפתורי עצור/אפס. הוספת hooks ל-operations.ts.

בטיחות: cancel על דריינר הלכות בטוח — חילוץ checkpointed per-chunk + resumable
+ self-heal לשורות processing.

Invariants: מקיים G12/INV-PORT1 (שער-הפלטפורמה). נוגע X6 (UI↔API).
api:types יורץ אחרי deploy (openapi.json חי).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

web/paperclip_client.py

@@ -720,6 +720,77 @@ async def get_issue_interactions(issue_ids: list[str]) -> list[dict]:
         await conn.close()
 
 
+# ── Agent-run observability + control ───────────────────────────────────────
+# Live view of which agents are actually working right now + their output, and
+# the controls to manage a stuck/runaway run. These wrap Paperclip's own
+# heartbeat-run API (verified live): we use the *graceful* platform endpoints
+# (cancel / reset-session) — never a raw kill on the run's process_pid, which
+# would bypass the platform's watchdog and our extractors' per-chunk
+# checkpointing. The only seam allowed to call these is web/agent_platform_port.
+
+async def list_live_runs(company_id: str) -> list[dict]:
+    """Queued + running heartbeat runs for a company (GET .../live-runs).
+
+    Each row carries ``agentName``/``status``/``issueId``/``startedAt`` and an
+    ``outputSilence`` block (``level`` ok|suspicion|critical) — the platform's
+    own liveness signal, surfaced so the UI can flag a stalled run.
+    """
+    resp = await pc_request(
+        "GET", f"/api/companies/{company_id}/live-runs", raise_on_error=True,
+    )
+    data = resp.json()
+    return data if isinstance(data, list) else []
+
+
+async def get_run_log(run_id: str) -> dict:
+    """Full output log of a heartbeat run (GET /api/heartbeat-runs/{id}/log).
+
+    Returns the platform payload as-is: ``{runId, store, logRef, content}``
+    where ``content`` is the NDJSON stream the adapter captured.
+    """
+    resp = await pc_request(
+        "GET", f"/api/heartbeat-runs/{run_id}/log", raise_on_error=True,
+    )
+    return resp.json()
+
+
+async def get_run_events(run_id: str) -> list[dict]:
+    """Lifecycle/event timeline of a heartbeat run (.../events)."""
+    resp = await pc_request(
+        "GET", f"/api/heartbeat-runs/{run_id}/events", raise_on_error=True,
+    )
+    data = resp.json()
+    return data if isinstance(data, list) else []
+
+
+async def cancel_run(run_id: str) -> dict:
+    """Gracefully cancel a queued/running heartbeat run (POST .../cancel).
+
+    The platform stops the run cleanly (process-group teardown + status flip),
+    respecting the watchdog. Safe for the halacha drain: its extractor is
+    checkpointed per-chunk and resumes on the next drain — a cancel loses at
+    most the in-flight chunk.
+    """
+    resp = await pc_request(
+        "POST", f"/api/heartbeat-runs/{run_id}/cancel", json={},
+        raise_on_error=True,
+    )
+    return resp.json()
+
+
+async def reset_agent_session(agent_id: str) -> dict:
+    """Reset an agent's runtime session (.../runtime-state/reset-session).
+
+    Clears a wedged session so the next wakeup starts clean — the smart
+    alternative to cancelling individual runs when an agent loops.
+    """
+    resp = await pc_request(
+        "POST", f"/api/agents/{agent_id}/runtime-state/reset-session", json={},
+        raise_on_error=True,
+    )
+    return resp.json()
+
+
 async def respond_to_interaction(
     issue_id: str, interaction_id: str, payload: dict,
 ) -> dict: