[BUG] v2.1.290: --resume on opus-5-5/sonnet-5-5 re-writes the whole history to the prompt cache every time (haiku unaffected)

Summary

On Claude Code v2.1.290, every --resume of a headless (-p) session on claude-opus-5-5 and claude-sonnet-5-5 reads only the system+tools prefix from the prompt cache and re-writes the entire conversation history as cache_creation_input_tokens. This happens on every resume, with or without a gap, and with both the 5-minute and the 1-hour TTL. On claude-haiku-4-5, the same resumes read the whole prior context from cache as expected.

This looks like the system-block scatter behaviour reported in #43657 and refiled as #67497. Both were closed as not-planned by inactivity sweeps. It is still present, so here are fresh numbers.

Measurements (per-call usage, not total_cost_usd, which is session-cumulative on resume)

| Model | Resume | cache_read | cache_creation |
|---|---|---|---|
| haiku-4-5 | turn 30, no gap | 153,194 | 4,855 (only the new turn) |
| opus-5-5 | turn 6, no gap, 5m TTL | 988 | 39,126 (whole history) |
| opus-5-5 | turn 6, no gap, 1h TTL | 988 | 39,188 |
| sonnet-5-5 | each resume | ~950 | whole history |

Positive control: an identical fresh dispatch of the same first prompt reads it all from cache on the same models (opus: 21,972 of 21,974; sonnet: 21,621 of 21,623). Inside one process, a tool-use loop also reads the earlier call's prompt. So caching works on these models, and the miss is specific to the resumed history.

Ruled out

• --agent / --agents: also reproduces on a bare claude -p.
• MCP: reproduces with --strict-mcp-config and no servers.
• Extended thinking: reproduces with MAX_THINKING_TOKENS=0.
• Effort mismatch (#66005): passing --effort high on the first turn and on every resume still gives the same miss (opus, each resume [read 12,196, create ~22k]). Dropping user settings via --setting-sources project,local also does not help.

Repro

claude -p --model claude-opus-5-5 --output-format stream-json --verbose "<~20k-token prompt>"   # note session_id
claude -p --model claude-opus-5-5 --output-format stream-json --verbose --resume <session_id> "next message"
# compare the resume's result usage: cache_read_input_tokens ≈ system+tools only

The same two commands with --model claude-haiku-4-5 read the full prior context on resume.

Impact

A process-per-turn orchestrator that resumes an opus session 30 times, with the context growing to about 158k, pays a full history write on every resume. We estimate about $12.50–$20, against about $1.50 if resume caching worked (derived from the measured per-resume rows). Recall across resumes was perfect (60 checks up to 158k), so correctness is fine; only cost is affected.

Linux, Claude Code 2.1.290.