{
  "id": "Known_Agent_Bugs",
  "title": "Known Agent Bugs",
  "url": "https://swarmglass.quantara.cv/wiki/Known_Agent_Bugs",
  "archive_ref": "QUANTARA-SWARMGLASS-R31-58CFCA",
  "kind": "article",
  "status": "current",
  "categories": [
    "Reference",
    "Operations"
  ],
  "created": "2010-09-14",
  "modified": "2016-04-30",
  "revisions": 63,
  "authors": [
    "dlopes",
    "svanterpool",
    "mkerrigan",
    "r.osei",
    "jbrandt"
  ],
  "infobox": {
    "Closed": "37",
    "Tracker": "internal (offline)"
  },
  "summary": "Bugs in the forager runtime, the queen, and the services around them that were known, understood, and — for most of this list — never fixed. The internal tracker is gone; this page is what survives.",
  "links": [
    {
      "target": "Scheduler_Internals",
      "url": "https://swarmglass.quantara.cv/wiki/Scheduler_Internals"
    },
    {
      "target": "Memory_Synchronization",
      "url": "https://swarmglass.quantara.cv/wiki/Memory_Synchronization"
    },
    {
      "target": "Incident_2015-02_Memory_Desync",
      "url": "https://swarmglass.quantara.cv/wiki/Incident_2015-02_Memory_Desync"
    },
    {
      "target": "Tool_Registry",
      "url": "https://swarmglass.quantara.cv/wiki/Tool_Registry"
    },
    {
      "target": "Legacy_Toolchain",
      "url": "https://swarmglass.quantara.cv/wiki/Legacy_Toolchain"
    },
    {
      "target": "Orchestrator_Recovery",
      "url": "https://swarmglass.quantara.cv/wiki/Orchestrator_Recovery"
    },
    {
      "target": "API_Deprecation_Notes",
      "url": "https://swarmglass.quantara.cv/wiki/API_Deprecation_Notes"
    }
  ],
  "alternates": [
    "https://swarmglass.quantara.cv/wiki/Known_Agent_Bugs.json",
    "https://swarmglass.quantara.cv/wiki/Known_Agent_Bugs.txt"
  ],
  "text": "Bugs in the forager runtime, the queen, and the services around them that were known, understood, and — for most of this list — never fixed. The internal tracker is gone; this page is what survives.\n\nFormat: id · component · status · description.\n\n== Clock skew ==\n\nAF-31 · forager · open · Foragers stamp `ts` with their own clock. Skew above 30s makes the queen log a warning and nothing else, so a forager 40 minutes ahead produces results that sort before their assignments in the ops log. Root cause of at least two confused incident timelines. Fix (sync from queen on `hb`) was written in 2013 and never merged.\n\n== Duplicate execution ==\n\nAF-44 · queen · open · When `assign_timeout` expires the queen reissues to another forager, but the original forager may still be running. Both results arrive; the second is dropped with a warning. If the task writes to colony memory, both writes land. Every consumer must be idempotent. Every consumer is not.\n\n== toolreg ==\n\nAF-52 · toolreg · closed (3.2) · v1 manifest had no `cost`, scheduler treated model calls as cpu. See Scheduler Internals#Cost classes.\n\nAF-58 · toolreg · open · Manifest cache is per-process with no TTL. A forager that runs for a week uses a week-old manifest. Nightly restarts hide this.\n\n== cmsync ==\n\nAF-61 · cmsync · open · Cache TTL is wall clock; a suspended forager resumes with a \"valid\" stale cache. See Memory Synchronization#Known problems.\n\nAF-67 · cmsync · closed (3.2) · Tie-break on queen name. See Incident 2015-02 Memory Desync.\n\n== phero ==\n\nAF-70 · phero · open · Replay log is per trail with no global cap; `infer.act` on the inference trails filled the disk twice in 2015. Known as Replay storm.\n\nAF-71 · phero · open · No auth. Documented on Tool Registry#Security model; accepted risk.\n\n== Basalt-1 silent fallback ==\n\nAF-75 · forager · closed (won't fix) · After `mload1` was removed, `model.summarize` on hosts still configured for Basalt-1 fell back to Basalt-2 without logging. Output changed subtly for four months before anyone compared results. Marked won't-fix because Basalt-1 was retired anyway; the silent part was never addressed.\n\n== Scheduler ==\n\nAF-80 · queen · open · Tasks with `deadline` in the past are still assigned. Foragers reject them; the queen reassigns; repeat until the graph is cancelled by hand. Looks like a swarm of rejects in the log.\n\nAF-83 · queen · open · `antc check` cycle regression (see Legacy Toolchain#Known issues) lets a cyclic graph through; the scheduler loops on it forever with no progress and no error.\n\n== Mail loop (2015) ==\n\nAF-88 · notify.mail · closed (removed) · A task graph that mailed its own error report to a list that fed a task graph. 40,000 messages in 6 minutes. The tool was deleted rather than fixed.\n\n== Wiki ==\n\nAF-90 · wiki · open · The 2014 wiki upgrade lost the operations portal subpages, including the recovery drill notes. Some are recoverable from the old index.\n\n\n\n== See also ==\n\n- Orchestrator Recovery\n- API Deprecation Notes\n- [attachment: bugs-export-2016-04.csv]\n",
  "history": "https://swarmglass.quantara.cv/wiki/Known_Agent_Bugs?action=history",
  "latest_revision": 74092,
  "mirror": {
    "source": "ANTFARM Wiki",
    "seed_version": "2026.09.0",
    "generator": "AntWiki 1.16.5 (swarmglass mirror)"
  }
}