Benign coding work was stopped by reasoning_extraction; drafting a follow-up comment to report that event was then refused as cyber

This is not a simple false positive.

Benign coding work was stopped by reasoning_extraction — and then the act of drafting a follow-up comment to report that stop to Anthropic was itself refused, this time as cyber.

So the problem is not only that harmless implementation work gets interrupted. The product can interrupt safe work, and then seal off — with its own safeguards — the very path by which the user attempts to record and report that interruption.

What this sequence implies

The conclusion this sequence points to is simple.

It is difficult to attribute the refusal to the visible user message itself. The final visible message was nothing more than a benign request to show a draft of a GitHub follow-up comment — and identical content did pass on v2.1.233.

It is therefore more reasonable to attribute this refusal not to what the user visibly sent, but to hidden or composite request state not faithfully represented in the transcript, or to a version-specific product path.

Actual damages

The damage from this failure does not end at discomfort or a momentary interruption. Investigation, isolation, retries, checking related issues, rolling back versions, re-drafting the report — all of it consumes real time, real tokens, and real weekly usage allowance.

This week, the weekly usage limit had reset at midnight. By the following morning, 48% of the Fable weekly allowance had already been consumed — and the bulk of that usage was lost not to actual implementation work, but to investigating, recovering from, recording, and reporting a failure that originated on the product side.

What is happening here is not mere inconvenience. The product's failure-handling costs are being shifted onto the user — as time, as operational burden, and as lost paid usage.

Given that this occurs in a product provided in exchange for a monthly subscription, the problem is clear: users pay for their own work. They did not sign a contract to have their usage allowance drained investigating, recovering from, and reporting the product's own erroneous stops.

At minimum, when this class of erroneous stop occurs during an Anthropic status incident, suspension or compensation of usage allowance consumed during that window should be considered.

Why the suggested guidance is not a fix

None of the following standard guidance is a solution:

• Start a new session
• Rephrase the request
• Use a different model
• Wait a while

Each of these merely transfers the cost of a product-side failure onto user-side avoidance effort. They are not fixes. Users are not paying in order to abandon session continuity or model choice every time the product fails; they pay to use the product under the conditions they chose.

Operational burden is not the only reason this is unacceptable. The user chooses which session, which model, and which continuity to work on — and pays not for "the possibility of getting some answer via some other path," but for a specific model, a specific continuity of work, a specific quality level — the utility of those chosen conditions themselves.

If, on that premise, the answer to a failure is "start a new session" or "use a different model," that is not a solution — it is unprofessional, negligent handling: the product fails to maintain the conditions the user paid for and chose, leaves that failure unaddressed, and simply pressures the user into accepting alternative paths with different quality, capability, continuity, and cost characteristics.

On a comparison of quality, cost, capability, and continuity, this user chose this product deliberately over alternative providers. If the answer to that choice is only "change how you use it" or "move elsewhere," it deserves to be evaluated as a degradation of the product's delivered value. If the cleanup of damage the product created is left to the user, and the only workarounds are escaping to another model or discarding session continuity, then the value of choosing this product at all is what comes into question.

To state it plainly: when a product stops harmless work, stops the harmless report about that stop, and then bills its paying users — in time, in continuity, in weekly allowance — for investigating its own failure, what is happening is not protection. It is cost transfer. The product is making the user pay the price of a conflict the product itself created.

---

Environment / references

• VS Code extension: refusals observed on v2.1.237; identical request passed after rollback to v2.1.233
• Categories observed: reasoning_extraction (coding work), cyber (drafting this report); response model shows <synthetic>
• Request ID (cyber, refusal of "show me the draft"): req_011CeEoyM85zpNhb9sK2ADjA (2026-08-21)
• Related: #87273 (original report, three correlated occurrences documented), #88131 (related series), and the recent v2.1.236 bisection report