[BUG] --resume never hits the prompt cache past the static prefix on Fable 5.1 (opus hits): session-start context message is replayed as a plain string instead of the content blocks it was sent with

update: this is model specific. same repro, same client, back to back, mcp off:

| model | fresh | resume 1 | resume 2 |
|---|---|---|---|
| opus (\--model opus\) | write 25262 / read 10126 | write 353 / read 35388 | write 264 / read 35741 |
| fable 5.1 (\--model fable\) | write 24378 / read 12887 | write 24639 / read 12887 | write 24902 / read 12887 |

so opus resumes hit, fable resumes miss every time. i also captured a subagent spawn vs its resume through the proxy on both models: the request shape change is identical on both (messages[1] is a content block array with cache_control on spawn and a plain string on resume, everything else byte identical incl. output_config effort and thinking). only fable misses on it. same with real subagents: 2 fresh fable seats missed on their first resume (~54k rewritten), 2 opus seats hit (139 and 645 written).

so either the fable cache does not treat a string content and a single text block as the same prefix while the opus one does, or the client should just replay the block form it sent originally. either way the cost only shows up on fable 5.1.

one more thing from the fable 5.1 docs (platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1), which i think explains why only fable misses:

Editing earlier turns invalidates thinking blocks. Modifying anything before a Claude Fable 5.1 thinking block (the system prompt, the tools, or an earlier message) results in an error on the next request, or in the block being dropped if you opt into that. [...] Claude Code, claude.ai, Claude Managed Agents, and the Claude Agent SDK keep that prefix intact for you.

and:

These patterns invalidate every later thinking block: [...] Injecting per-request text into an earlier turn (a reminder or status line) that you remove on the next request.

the resume replay changes messages[1] from a block array to a string, so from fables point of view an earlier turn was modified. for accounts created before 2026-08-31 the api records the mismatch silently, which matches what i see: no error, just a full re-write. so the docs promise that claude code keeps the prefix intact is exactly what the resume path does not do on 2.1.268. maintainers could confirm with thinking.block_binding.prefix_mismatch_behavior: "drop_block" and input_transformations on a resumed request.

cost side: fable cache reads are $0.25/MTok, 1h cache writes $20/MTok, so every resume of a long fable session pays the write rate on the whole history instead of the read rate, 80x per token.

closing the caveat from the report: with ENABLE_TOOL_SEARCH=true the proxy capture keeps tool search on (12 tools incl. ToolSearch, deferred, no listing block in messages[0]), i.e. the normal request shape. same result, fable, mcp off:

| call | cache_read | cache_creation |
|---|---|---|
| fresh | 0 | 30533 |
| resume 1 | 12690 | 18081 |
| resume 2 | 30536 | 475 |

fresh vs resume 1: tools identical, system identical, messages[0] identical (4 blocks), first and only diff is messages[1] (block array with cache_control on fresh, plain string on resume, same text). resume 2 hits again because resume 1 wrote its own entry at the assistant reply with messages[1] already in string form; only the previous turns hook-context message flips again and it is small. so the big miss is the first resume of any session (or subagent) on fable, where the flipped block is the ~22k char session-start message sitting right after messages[0] with no earlier message breakpoint.

Independent confirmation on Claude Code 2.1.268, first-party Claude subscription authentication, --safe-mode, medium effort, macOS x86_64. The official latest-channel binary was downloaded separately and checked against its published SHA-256; no production upgrade was involved.

Six successful main requests: fresh session plus two genuinely new continuations for each model, sequentially within each arm. No tools were called. Final result.usage values (cache read / cache creation / uncached input):

| Model | Fresh | Resume 1 | Resume 2 |
|---|---|---|---|
| fable → claude-fable-5-1 | 10,126 / 6,647 / 2 | 13,003 / 3,890 / 2 | 13,003 / 4,010 / 2 |
| opus → claude-opus-5 | 10,126 / 5,751 / 2 | 15,877 / 117 / 2 | 15,994 / 117 / 2 |

The Fable continuations both report native message.diagnostics.cache_miss_reason.type = messages_changed (estimated missed tokens 3,554 and 3,647). Opus diagnostics are null alongside growing cache reuse. Initial requests need not be cold: shared static prefixes may already be cached.

Native request-body capture using OTEL_LOG_RAW_API_BODIES=file:<private-directory> confirms, for both models:

• Message counts 2 → 5 → 8.
• Tools unchanged within each arm.
• First continuation: only historical message index 1 changes, from a text-block list to a string with identical text.
• Second continuation: only historical message index 4 changes in the same way; earlier entries remain equal.

Thus in this safe-mode run the second Fable continuation still does not regain growing conversation reuse, unlike the last posted example above. Also, the representation transition itself is not a sufficient causal explanation: Opus makes the same transition and succeeds. Native model-dependent construction versus server interpretation remains unresolved.

Reproduction outline (normal subscription-authenticated official CLI; use a new UUID per model):

claude -p --safe-mode --model fable --effort medium --output-format stream-json --verbose --session-id <new-uuid> 'Independent cache probe <uuid>, new continuation 0. Do not use tools or change files. Reply exactly CURRENT_CACHE_0_OK.'
claude -p --safe-mode --model fable --effort medium --output-format stream-json --verbose --resume <same-uuid> 'Independent cache probe <uuid>, new continuation 1. Do not use tools or change files. Reply exactly CURRENT_CACHE_1_OK.'
claude -p --safe-mode --model fable --effort medium --output-format stream-json --verbose --resume <same-uuid> 'Independent cache probe <uuid>, new continuation 2. Do not use tools or change files. Reply exactly CURRENT_CACHE_2_OK.'

Repeat with --model opus and a different UUID. Our runner supplies the prompt via stdin rather than the positional argument shown above, and uses a credential-isolating wrapper that executes the unmodified official binary. Safe mode disables project customizations. No custom system prompt or beta-disable override was added.

Earlier independent tests on 2.1.260/2.1.263 also reproduced the Fable pattern, including new-input boundaries within one continuous CLI process, so process restart/resume is not necessary for the failure. Those earlier continuous-process results are not claimed as a new 2.1.268 test.

Current Fable continuation message IDs for maintainer correlation: msg_011CewCyJFHCj9kUMrivvbJ1, msg_011CewCyWppsH1PpdrSUWiFm. No raw prompts, credentials, or session dumps attached.