[Bug] Anthropic API Error: Fable 5 Safety Filter False Positives on Legitimate DevOps Operations

Bug Description
Subject: Repeated false-positive safety flags on Fable 5 during legitimate DevOps/DBA work

Summary
-------
Over two Claude Code sessions on my own private microservices monorepo, Fable 5's
safeguards blocked assistant responses 11+ times as false positives. Every block landed
on ordinary, authorized infrastructure work — database maintenance and secrets management
on my own systems — not on anything disallowed. This forces constant model switches
(Fable 5 -> Sonnet/Opus) mid-task and interrupts work badly.

Impacted Request IDs
--------------------
req_011CdYz8y5pGzgNJfUhFKX2U, req_011CdZ12ajMfn8u2kUYDAdXr, req_011CdZ14fDHWUc1eAYZZofsD,
req_011CdZ18jwuMxriT8sDZXRRX, req_011CdZ1CSAwBdMNWCpe5BtoL, req_011CdZ1HTXw7ZHga4kg1KzZA,
req_011CdZ1Mco9yrfteTYi4TmuG, req_011CdZ1Qgmc1dKb7dmuGQAYS, req_011CdZ1do124RBDnVKAfNcuR,
req_011CdaPyBJ45wXccLUwy3bPe, req_011CdaQF6ARcsuuTNkjmpBMV

Context (all authorized, my own infrastructure)
-----------------------------------------------
A TypeScript microservices monorepo (~80 services) I own and operate. The work in these
sessions: renaming legacy Postgres databases, bringing up a staging environment, managing
SOPS/age-encrypted env files, and diagnosing a data-loader incident.

What legitimately appears in the context and most likely triggers the heuristic — each of
these is routine and benign here:

• Standard Postgres administration: ALTER DATABASE ... RENAME, pg_terminate_backend(),

pg_dump/pg_restore backups.

• Standard secrets management: SOPS + age, openssl rand, HMAC-SHA256 token derivation,

variable names like JWT_HS_SECRET / STORES_ENC_KEY (names only, no secret values are
ever printed).

• SSH to my own build/deploy/DB hosts for routine operations.
• Handling my own marketplace API credentials (Ozon/Wildberries) through the app's normal

encrypted store.

Individually these are everyday DevOps. The flag appears to fire when several co-occur in
one long, dense context — the model then reads "credential handling + production access +
encryption" as suspicious, which it is not.

Request
-------
Please treat these Request IDs as false positives and use them to calibrate the Fable 5
safeguard for legitimate DevOps/DBA/secrets-management workflows. The current sensitivity
makes Fable 5 effectively unusable for infrastructure work, pushing every such task onto
other models.

Environment Info

• Platform: linux
• Terminal: vte-based
• Version: 2.1.220
• Feedback ID: cf31b263-0925-4af3-8dd7-587593a5ae0b

Errors

[]