claude-code - 💡(How to fix) Fix opus-4-8 appears to re-fetch identical tool results after normal (non-error) returns — 2-3x vs opus-4-7

Summary: On Claude Code, opus-4-8 shows a markedly higher rate of compulsively re-running the same tool call (same Read file_path / same Bash command) within a short window, compared to opus-4-7 — even when the prior result was a normal success (not an error, timeout, or empty result). This burns cache_read tokens (each repeat re-bills the growing context) and can feed sibling-tool-call cascades.

Method: Scanned local Claude Code transcripts (~/.claude/projects/<proj>/*.jsonl, ~175 sessions). Defined a "burst" = same (tool, key) appearing >=3x within any 12-call window, where key = Read.file_path or Bash.command. Scripts available on request.

Data — burst rate by model, controlled for task type (to rule out "the 4.8 window happened to be heavy meta-debug work"):

Same task type, opus-4-8 is 2-3x opus-4-7. Task type adds a further ~1.5-2x, but model is the dominant axis.

Key finding: of 130 burst onsets, 88% were immediately preceded by a normal-ok tool result (not error/empty/cascade). So this is not a defensive reaction to a flaky channel — the model re-fetches results that already returned fine.

Spiral: in heavy sessions, repeats concentrate in the latter half (median position 0.64), i.e. once it starts it compounds.

Caveats (honest): single user, one ~2-week window; opus-4-8 sample is smaller (21 sessions / 2,581 calls vs 4.7's 99 / 7,416); "burst" is a self-defined heuristic, not an official metric. I cannot separate "model weights" from "4.8-era harness changes" (system prompt / deferred tools / CC version) from the client side — that needs Anthropic's cross-user telemetry to confirm whether other 4.8 users see the same lift.

Ask: please check cross-user telemetry for an opus-4-8 vs 4.7 difference in same-tool-call repetition rate after non-error results.