GitHub issue #53065 in anthropics/claude-code, opened 2026-04-25 by AttacktheDPoint-com, records a report that when advisor() is called, the full conversation transcript is forwarded to a second model, and the token usage from both the main executor and the advisor sub-inference are summed in the top-level usage fields, so if auto-compaction logic uses these summed totals the advisor call effectively doubles the apparent context usage, triggering compaction when the main model's actual context is only about 50 percent full; in the reported session the total reported input doubled from 513,354 to 1,028,789 on the advisor turn and the session compacted immediately after, wiping context to about 36K; the report was filed against Claude Code 2.1.111 on Windows 10 Pro with claude-opus-4-6[1m] as the main model and claude-opus-4-7 as the advisor model.
Verbatim Excerpts from Cited Records
“When `advisor()` is called, the full conversation transcript is forwarded to a second model (currently `claude-opus-4-7`).”
·
anthropics/claude-code (issue 53065 opening body, Summary section)
·
2026-04-25T00:10:44Z
·GitHub Issue ↗
“The token usage from both the main executor and the advisor sub-inference are summed in the top-level `usage` fields.”
·
anthropics/claude-code (issue 53065 opening body, Summary section)
·
2026-04-25T00:10:44Z
·GitHub Issue ↗
“If Claude Code's auto-compaction logic uses these summed totals, the advisor call effectively doubles the apparent context usage, triggering compaction when the main model's actual context is only ~50% full.”
·
anthropics/claude-code (issue 53065 opening body, Summary section)
·
2026-04-25T00:10:44Z
·GitHub Issue ↗
“The main model's actual context was **513-515K tokens** (iterations 1 and 3).”
·
anthropics/claude-code (issue 53065 opening body, Evidence from session JSONL section)
·
2026-04-25T00:10:44Z
·GitHub Issue ↗
Counter-record: Issue closed 2026-11-14 by bpasero-ant, who said it is fixed in Claude Code 2.1.273. 19 comments were read. kerneltoast (2026-05-03) confirmed the same issue on Linux with Opus 4.6 [1M] and Opus 4.7 as advisor, and reported it still occurring on v2.1.126, v2.1.140, v2.1.148, v2.1.156 and v2.1.196; genesiscz (2026-05-25) independently reproduced on claude-opus-4-7[1m] on macOS, showing a single advisor() call took the statusline from about 406K to 1170k/1000k (117 percent), and added that a timeout-driven auto-retry can triple the count and that the inflation is display-only for the cache while the same summed figure feeds the auto-compaction trigger; johnsfleming (2026-07-22) showed the same premature-compaction mechanism without any advisor call, from plain message continuation/retry iterations; joshua-angulo (2026-08-14) confirmed the mechanism on 2.1.232 with account-wide numbers (5 of 8 lifetime auto-compactions were phantom, firing at real occupancies of 295k-432k) and a tested PreCompact hook workaround; Studytempest (2026-09-03) scanned 176 compaction records and found 87 of 165 auto-compactions fired off a rolled-up count; architerm (2026-09-11) confirmed on 2.1.268 with binary-level analysis; polisherm (2026-09-22) confirmed the fix on 2.1.278, citing the 2.1.273 changelog. The closure records a fix in 2.1.273, not a refutation of the original report. No retraction of the original report appears in the surveyed thread. The claim documents only the original report, not a verified current state.
Falsification probe: If GitHub issue #53065 is deleted without archive, if the author edits the opening body so the quoted sentences no longer match verbatim, or if the author posts a retraction of the report, this claim drops to Withdrawn.
Audit verification: Verified 100% continuous verbatim substring against official GitHub REST API by
scripts/verify_claim_verbatim.py.