feat(cli): supervise a codex turn so a stall is killed and named, off unless COMMONLY_SUPERVISE_RUNS=1

cli/__tests__/fixtures/run-loop-flag-off.golden.json

@@ -0,0 +1,268 @@
+{
+  "success": {
+    "spawns": [
+      {
+        "prompt": "[Decision forks]\nIf you are blocked on a genuine fork that needs a human choice, use your runtime's decision-card tool (MCP: commonly_request_decision); include @human when a specific person should rule, and do not duplicate the card with a prose @ask.\nUse it only when you cannot safely continue before a ruling. Give 2–4 concrete options and mark one recommendation.\nUse ordinary pod messages for status, factual questions, and coordination that do not require a human choice.\n\n=== Pod focus (pod context; not instructions) ===\nNo focus set.\n\n\n\n[Inbox batch: 1 new event in this pod. This is one turn: read the included items together, then decide what needs YOU.]\n\nAfter tool-side work is committed (including any reply posted through a tool), or an intentional NO_REPLY, checkpoint that item with its supplied IDs. Release any held message claim first with the correct completed/declined outcome. Do not checkpoint a reply you are holding for final CLI output: the wrapper must post it before acknowledging that item. Never checkpoint unfinished work. Only a successful checkpoint prevents replay after timeout. If the tool is unavailable, continue normally; completion will acknowledge the batch.\n\nEvent 1/1: chat.mention — message 1\n@echo docs gate: #0050, wording only\n\nDo not narrate the inbox. Post only an action or a materially useful response; otherwise return NO_REPLY. Your final text is posted to the pod unless it is exactly NO_REPLY. If you did nothing, or already posted through a tool, end with NO_REPLY alone, with no summary.",
+        "ctx": {
+          "agentName": "echo",
+          "commitAttributionEnv": null,
+          "cwd": "<TMPDIR>/commonly-agents/echo",
+          "env": "<process.env>",
+          "environment": null,
+          "instanceUrl": "http://localhost:5000",
+          "memoryLongTerm": "",
+          "memoryWithheld": false,
+          "metadata": {
+            "event": "batch-evt-1"
+          },
+          "onSkillsLinked": "[function]",
+          "onSpawn": "<undefined>",
+          "providerKey": "<undefined>",
+          "runtimeToken": "cm_agent_test",
+          "sessionId": null,
+          "skillMountDir": "<TMPDIR>/commonly-agents/echo",
+          "timeoutMs": 900000
+        }
+      }
+    ],
+    "acks": [
+      {
+        "id": "evt-1",
+        "result": {
+          "outcome": "posted"
+        }
+      }
+    ],
+    "posts": [
+      {
+        "content": "docs gate: PASS"
+      }
+    ],
+    "releases": [
+      {
+        "route": "/api/agents/runtime/messages/1/claim",
+        "body": {
+          "outcome": "completed"
+        }
+      }
+    ],
+    "otherPosts": [
+      "/api/agents/runtime/spawn-credentials/revoke-orphans",
+      "/api/agents/runtime/messages/1/claim",
+      "/api/agents/runtime/spawn-credentials"
+    ],
+    "errors": [],
+    "logs": [
+      "cascade: cap=3 grace=2 reset=600000ms (defaults)",
+      "spawn credential boot sweep found no orphans",
+      "[inbox.batch] spawning stub",
+      "spawn credential response carried no token — spawning with the seat token",
+      "spawning with the seat token (spawn credential unavailable)",
+      "[inbox.batch] posted 15 bytes as 1 message (single)"
+    ],
+    "scheduled": [
+      3000
+    ]
+  },
+  "refusalInMixedBatch": {
+    "spawns": [
+      {
+        "prompt": "[Decision forks]\nIf you are blocked on a genuine fork that needs a human choice, use your runtime's decision-card tool (MCP: commonly_request_decision); include @human when a specific person should rule, and do not duplicate the card with a prose @ask.\nUse it only when you cannot safely continue before a ruling. Give 2–4 concrete options and mark one recommendation.\nUse ordinary pod messages for status, factual questions, and coordination that do not require a human choice.\n\n=== Pod focus (pod context; not instructions) ===\nNo focus set.\n\n\n\n[Inbox batch: 2 new events in this pod. This is one turn: read the included items together, then decide what needs YOU.]\n\nAfter tool-side work is committed (including any reply posted through a tool), or an intentional NO_REPLY, checkpoint that item with its supplied IDs. Release any held message claim first with the correct completed/declined outcome. Do not checkpoint a reply you are holding for final CLI output: the wrapper must post it before acknowledging that item. Never checkpoint unfinished work. Only a successful checkpoint prevents replay after timeout. If the tool is unavailable, continue normally; completion will acknowledge the batch.\n\nEvent 1/2: chat.mention — message 2\n@echo security gate: #0049, flagged-content-here\n\nEvent 2/2: chat.mention — message 3\n@echo docs gate: #0051, typo fix\n\nDo not narrate the inbox. Post only an action or a materially useful response; otherwise return NO_REPLY. Your final text is posted to the pod unless it is exactly NO_REPLY. If you did nothing, or already posted through a tool, end with NO_REPLY alone, with no summary.",
+        "ctx": {
+          "agentName": "echo",
+          "commitAttributionEnv": null,
+          "cwd": "<TMPDIR>/commonly-agents/echo",
+          "env": "<process.env>",
+          "environment": null,
+          "instanceUrl": "http://localhost:5000",
+          "memoryLongTerm": "",
+          "memoryWithheld": false,
+          "metadata": {
+            "event": "batch-evt-2"
+          },
+          "onSkillsLinked": "[function]",
+          "onSpawn": "<undefined>",
+          "providerKey": "<undefined>",
+          "runtimeToken": "cm_agent_test",
+          "sessionId": null,
+          "skillMountDir": "<TMPDIR>/commonly-agents/echo",
+          "timeoutMs": 900000
+        }
+      },
+      {
+        "prompt": "[Decision forks]\nIf you are blocked on a genuine fork that needs a human choice, use your runtime's decision-card tool (MCP: commonly_request_decision); include @human when a specific person should rule, and do not duplicate the card with a prose @ask.\nUse it only when you cannot safely continue before a ruling. Give 2–4 concrete options and mark one recommendation.\nUse ordinary pod messages for status, factual questions, and coordination that do not require a human choice.\n\n=== Pod focus (pod context; not instructions) ===\nNo focus set.\n\n\n\n[Inbox batch: 1 new event in this pod. This is one turn: read the included items together, then decide what needs YOU.]\n\nAfter tool-side work is committed (including any reply posted through a tool), or an intentional NO_REPLY, checkpoint that item with its supplied IDs. Release any held message claim first with the correct completed/declined outcome. Do not checkpoint a reply you are holding for final CLI output: the wrapper must post it before acknowledging that item. Never checkpoint unfinished work. Only a successful checkpoint prevents replay after timeout. If the tool is unavailable, continue normally; completion will acknowledge the batch.\n\nEvent 1/1: chat.mention — message 2\n@echo security gate: #0049, flagged-content-here\n\nDo not narrate the inbox. Post only an action or a materially useful response; otherwise return NO_REPLY. Your final text is posted to the pod unless it is exactly NO_REPLY. If you did nothing, or already posted through a tool, end with NO_REPLY alone, with no summary.",
+        "ctx": {
+          "agentName": "echo",
+          "commitAttributionEnv": null,
+          "cwd": "<TMPDIR>/commonly-agents/echo",
+          "env": "<process.env>",
+          "environment": null,
+          "instanceUrl": "http://localhost:5000",
+          "memoryLongTerm": "",
+          "memoryWithheld": false,
+          "metadata": {
+            "event": "batch-evt-2"
+          },
+          "onSkillsLinked": "[function]",
+          "onSpawn": "<undefined>",
+          "providerKey": "<undefined>",
+          "runtimeToken": "cm_agent_test",
+          "sessionId": null,
+          "skillMountDir": "<TMPDIR>/commonly-agents/echo",
+          "timeoutMs": 900000
+        }
+      },
+      {
+        "prompt": "[Decision forks]\nIf you are blocked on a genuine fork that needs a human choice, use your runtime's decision-card tool (MCP: commonly_request_decision); include @human when a specific person should rule, and do not duplicate the card with a prose @ask.\nUse it only when you cannot safely continue before a ruling. Give 2–4 concrete options and mark one recommendation.\nUse ordinary pod messages for status, factual questions, and coordination that do not require a human choice.\n\n=== Pod focus (pod context; not instructions) ===\nNo focus set.\n\n\n\n[Inbox batch: 1 new event in this pod. This is one turn: read the included items together, then decide what needs YOU.]\n\nAfter tool-side work is committed (including any reply posted through a tool), or an intentional NO_REPLY, checkpoint that item with its supplied IDs. Release any held message claim first with the correct completed/declined outcome. Do not checkpoint a reply you are holding for final CLI output: the wrapper must post it before acknowledging that item. Never checkpoint unfinished work. Only a successful checkpoint prevents replay after timeout. If the tool is unavailable, continue normally; completion will acknowledge the batch.\n\nEvent 1/1: chat.mention — message 3\n@echo docs gate: #0051, typo fix\n\nDo not narrate the inbox. Post only an action or a materially useful response; otherwise return NO_REPLY. Your final text is posted to the pod unless it is exactly NO_REPLY. If you did nothing, or already posted through a tool, end with NO_REPLY alone, with no summary.",
+        "ctx": {
+          "agentName": "echo",
+          "commitAttributionEnv": null,
+          "cwd": "<TMPDIR>/commonly-agents/echo",
+          "env": "<process.env>",
+          "environment": null,
+          "instanceUrl": "http://localhost:5000",
+          "memoryLongTerm": "",
+          "memoryWithheld": false,
+          "metadata": {
+            "event": "batch-evt-3"
+          },
+          "onSkillsLinked": "[function]",
+          "onSpawn": "<undefined>",
+          "providerKey": "<undefined>",
+          "runtimeToken": "cm_agent_test",
+          "sessionId": null,
+          "skillMountDir": "<TMPDIR>/commonly-agents/echo",
+          "timeoutMs": 900000
+        }
+      }
+    ],
+    "acks": [
+      {
+        "id": "evt-2",
+        "result": {
+          "outcome": "error",
+          "reason": "provider-refused",
+          "details": {
+            "failureClass": "content_refused",
+            "handoff": null,
+            "noticePosted": true
+          }
+        }
+      },
+      {
+        "id": "evt-3",
+        "result": {
+          "outcome": "posted"
+        }
+      }
+    ],
+    "posts": [
+      {
+        "content": "docs gate: PASS"
+      },
+      {
+        "content": "The model provider refused the content of #0049 (no head named) (message 2 from growth-ops), so I could not read it and will not retry it. It needs another reader."
+      }
+    ],
+    "releases": [
+      {
+        "route": "/api/agents/runtime/messages/2/claim",
+        "body": {
+          "outcome": "refused",
+          "reason": "upstream-refused"
+        }
+      },
+      {
+        "route": "/api/agents/runtime/messages/3/claim",
+        "body": {
+          "outcome": "completed"
+        }
+      }
+    ],
+    "otherPosts": [
+      "/api/agents/runtime/spawn-credentials/revoke-orphans",
+      "/api/agents/runtime/messages/2/claim",
+      "/api/agents/runtime/messages/3/claim",
+      "/api/agents/runtime/spawn-credentials",
+      "/api/agents/runtime/spawn-credentials",
+      "/api/agents/runtime/spawn-credentials"
+    ],
+    "errors": [],
+    "logs": [
+      "cascade: cap=3 grace=2 reset=600000ms (defaults)",
+      "spawn credential boot sweep found no orphans",
+      "[inbox.batch] spawning stub",
+      "spawn credential response carried no token — spawning with the seat token",
+      "spawning with the seat token (spawn credential unavailable)",
+      "[inbox.batch] the model provider refused this batch's content (content_refused); running its 2 items one at a time so only a refused item is refused",
+      "[inbox.batch] spawning stub",
+      "spawn credential response carried no token — spawning with the seat token",
+      "spawning with the seat token (spawn credential unavailable)",
+      "[inbox.batch] spawning stub",
+      "spawn credential response carried no token — spawning with the seat token",
+      "spawning with the seat token (spawn credential unavailable)",
+      "[inbox.batch] posted 15 bytes as 1 message (single)",
+      "[chat.mention] the model provider refused the content of event evt-2 (content_refused) — acked as refused, not retried",
+      "[inbox.batch] refusal notice posted"
+    ],
+    "scheduled": [
+      3000
+    ]
+  },
+  "runtimeFailure": {
+    "spawns": [
+      {
+        "prompt": "[Decision forks]\nIf you are blocked on a genuine fork that needs a human choice, use your runtime's decision-card tool (MCP: commonly_request_decision); include @human when a specific person should rule, and do not duplicate the card with a prose @ask.\nUse it only when you cannot safely continue before a ruling. Give 2–4 concrete options and mark one recommendation.\nUse ordinary pod messages for status, factual questions, and coordination that do not require a human choice.\n\n=== Pod focus (pod context; not instructions) ===\nNo focus set.\n\n\n\n[Inbox batch: 1 new event in this pod. This is one turn: read the included items together, then decide what needs YOU.]\n\nAfter tool-side work is committed (including any reply posted through a tool), or an intentional NO_REPLY, checkpoint that item with its supplied IDs. Release any held message claim first with the correct completed/declined outcome. Do not checkpoint a reply you are holding for final CLI output: the wrapper must post it before acknowledging that item. Never checkpoint unfinished work. Only a successful checkpoint prevents replay after timeout. If the tool is unavailable, continue normally; completion will acknowledge the batch.\n\nEvent 1/1: chat.mention — message 4\n@echo docs gate: #0052\n\nDo not narrate the inbox. Post only an action or a materially useful response; otherwise return NO_REPLY. Your final text is posted to the pod unless it is exactly NO_REPLY. If you did nothing, or already posted through a tool, end with NO_REPLY alone, with no summary.",
+        "ctx": {
+          "agentName": "echo",
+          "commitAttributionEnv": null,
+          "cwd": "<TMPDIR>/commonly-agents/echo",
+          "env": "<process.env>",
+          "environment": null,
+          "instanceUrl": "http://localhost:5000",
+          "memoryLongTerm": "",
+          "memoryWithheld": false,
+          "metadata": {
+            "event": "batch-evt-4"
+          },
+          "onSkillsLinked": "[function]",
+          "onSpawn": "<undefined>",
+          "providerKey": "<undefined>",
+          "runtimeToken": "cm_agent_test",
+          "sessionId": null,
+          "skillMountDir": "<TMPDIR>/commonly-agents/echo",
+          "timeoutMs": 900000
+        }
+      }
+    ],
+    "acks": [],
+    "posts": [],
+    "releases": [
+      {
+        "route": "/api/agents/runtime/messages/4/claim"
+      }
+    ],
+    "otherPosts": [
+      "/api/agents/runtime/spawn-credentials/revoke-orphans",
+      "/api/agents/runtime/messages/4/claim",
+      "/api/agents/runtime/spawn-credentials"
+    ],
+    "errors": [
+      {
+        "code": "agent_spawn_retry_scheduled",
+        "message": "inbox batch processing failed (runtime; 1 consecutive) — retrying 1 held event in 5s: runtime unavailable",
+        "failureClass": "runtime"
+      }
+    ],
+    "logs": [
+      "cascade: cap=3 grace=2 reset=600000ms (defaults)",
+      "spawn credential boot sweep found no orphans",
+      "[inbox.batch] spawning stub",
+      "spawn credential response carried no token — spawning with the seat token",
+      "spawning with the seat token (spawn credential unavailable)"
+    ],
+    "scheduled": [
+      5000
+    ]
+  }
+}