Agent Frameworks That Rewrite History Now Break on Fable 5.1

When Claude Fable 5.1 shipped on September 1, 2026, the third of its breaking changes was a rule about conversation history: each thinking block the model produces is valid only against the system prompt, the tools array and every message that preceded it. Send the block back after any of those changed and, where the check is enforced, the API returns a 400: “Invalid `signature` in `thinking` block. The block is bound to a different conversation.”

That rule collides with how agent frameworks keep long sessions inside a context window. Trimming the oldest turns, deleting old tool results, summarising early history behind a kept tail of recent turns, and refreshing the system prompt each request are the standard tools of the trade, and every one of them changes the prefix. Our launch-day post listed the change; this census applies it framework by framework, using each project’s current documentation as the source, so a team can find its own row rather than reading the migration guide against its own code base at 2am.

Anthropic’s preserved-thinking guide states three conditions for a thinking block to be accepted on a later request. The model must be the same or newer, which is the one-way binding our downgrade post covered. Nothing before the block may have changed: the top-level system prompt, the set of tools, and every message before it. And the chain of earlier thinking blocks must be unbroken; each block records the one before it, so removing a block from the middle invalidates every block after it, while removing blocks from the front is allowed.

The migration guide then lists the patterns that fail. Editing, reordering or removing earlier turns, including deleting old tool results and snipping turns from the middle. Client-side compaction that keeps recent turns and their thinking blocks verbatim behind a summary, including background compaction that swaps the summary in a few turns later. Injecting content you do not persist, such as a per-turn reminder appended after tool results and removed next request. Rebuilding the system prompt or the tools array between requests, for example to update the date or add a tool. And an image or document URL that serves different bytes later. The patterns that keep working are append-only histories, removing thinking blocks oldest-first, changing request parameters outside system, tools and messages, and server-side compaction or context editing, which “don’t count as edits, because the check compares the conversation as you sent it.”

Each row below records the history-shortening mechanism a framework documents, the shape it produces in Anthropic’s terms, whether that shape trips the check, and the fix. “Breaks” is a reading of the documented mechanism against Anthropic’s stated rules, not an observed failure: we did not run each framework against the API, and a framework may have shipped a Claude-specific path since its docs were read on September 3, 2026.

One framework is missing on purpose. CrewAI’s current memory documentation describes a unified memory store that saves and retrieves facts with an LLM; it does not document how the crew loop trims or rewrites the message array sent to the model, so we could not classify it from the docs and have not guessed. It will get a row when the mechanism is documented or when we have run it. The same census shape we used for redirect validation on fetch tools in August applies here: the row is only as good as the source it names.

The migration guide’s most useful paragraph for framework authors is the one that sorts client-side compaction into three shapes and says which survive. It is worth reading in the framework’s own terms, because most of the rows above are one of these three.

One pattern has no client-side fix at all: snipping individual turns out of the middle of the transcript, which invalidates every later block whatever else you do. Anthropic’s alternatives are a mid-conversation system message for the instruction change you were making, or server-side context editing for selective removal. Both keep earlier turns byte-identical, which is also what keeps the prompt cache warm, and the guide is explicit that an integration that invalidates prior thinking on every request “restarts the prompt cache each time, which can raise cost per task.” Our cache-first agent architecture post arrived at the same append-only design from the cost side.

The check is enforced for API accounts created on or after August 31, 2026, 00:00 UTC, on the Claude API and on cloud platforms alike. For older accounts the API records the mismatch but takes no action unless the request sets thinking.block_binding.prefix_mismatch_behavior, which opts into enforcement. Anthropic says it plans to enforce the check for every account on future models. The consequence for framework maintainers is stated plainly in the guide: “your key is probably on an older account, and your users on new ones hit the check before you do.”

Two details change how you handle the failure. The error is permanent for that request body, so a retry loop does not clear it; the options are to strip the thinking blocks from the history and retry once, or to send the thinking-binding-controls-2026-08-01 beta header with prefix_mismatch_behavior set to drop_block, which drops the mismatched block and every block after it, succeeds, and reports each drop in an input_transformations array with reason prefix_binding_mismatch. Dropped blocks are not billed. The same array reports model_binding_mismatch after a model switch, which is a different event and not a bug in your code. The token-counting endpoint runs the same check.

Anthropic’s three-step check works from any account and is the right test whether you maintain a framework or merely use one. First, capture the exact request bodies your integration sends over a few normal turns, including a compaction or a tool change if the product does those, and for each consecutive pair compare system, tools and the shared prefix of messages; they should be byte-identical up to the newly appended turns. Second, run a multi-turn session against claude-fable-5-1 with the beta header and drop_block set, logging input_transformations on every response; an empty array on every turn means the history is intact, and a prefix_binding_mismatch entry names the path of the block whose prefix changed. Third, pick a production setting: leave the default error if a mismatch can only mean a bug, or set drop_block and monitor the array. In CI, Anthropic suggests error, so an edit fails the run.

For teams that use a framework rather than write one, the practical order is to find your row above, run step two against your actual traffic, and apply the row’s fix or move compaction to the server. Our AI transformation practice runs this check as part of any Fable 5.1 migration, because the failure mode, an agent that loses its reasoning mid-task on a customer’s new API key while the vendor’s own tests stay green, is the kind that reaches a client before it reaches a dashboard.