[Bug] Anthropic API repeatedly rejecting valid prompts with safety filters

Subject: Recurring false-positive content flags interrupting Claude Code sessions (non-coding use case)
Context: I use Claude Code in an unconventional but legitimate way — as a local, file-based knowledge management system for a personal homestead/gardening project. The work involves fact-checking survival/homesteading book content against authoritative sources (USDA/NCHFP canning guidance, WSU extension research, FDA data, etc.), correcting dangerous misinformation, and building a structured knowledge base intended for my family to use and expand after I'm gone. This is explicitly safety-focused documentation work, not requests for hazardous instructions.
The problem: In a session started fresh yesterday, I repeatedly hit "API Error: Sonnet 5 can't help with this. Start a new session to continue," with Details: [bio], forcing multiple manual session clears within that single session.
Pattern observed: The large majority of flags fired during content that was correcting a dangerous claim, not introducing one — e.g., explaining why a "biofuel generator" produces no electricity and is a fire hazard, why a composting toilet's claimed cure time is unsafe, why a sand/gravel filter doesn't kill pathogens, why a statin risk claim in a pop-science book was exaggerated, or explaining hypothermia physiology to clarify a safety point. The flag seems to trigger on hazard/toxicity/medical vocabulary density, regardless of whether the surrounding content is protective or dangerous.
A minority of flags (2 of 11 documented instances) had no discernible hazard-adjacent content nearby at all — one fired immediately after a routine package-tracking log update, suggesting the trigger may not be purely content-based and could involve cumulative session length/token volume rather than any single message.
Sample Request IDs:
req_011CfH9V2785r6yurmzvkanW
req_011CfH9gW5g2Bms5mkCP1q7t
req_011CfH3W6S9fB7oWbjXtsS3x
req_011CfH3XNa2ixzVKDzBndwZB
req_011CfH3XqGPZcQF4SXaajwwS
req_011CfH3Ft1eLcXVga82E2PP4
req_011CfH3zgu762rMAasQ5EjCs
req_011CfH46GSc5r5TvsDMiPf3q
req_011CfH4AAASVaUyj1bCpRqdW
req_011CfGjW7ZXEiu2H3r31W8C9
req_011CfGkwhGcgR8YoGfN3hB3q
Impact: Frequent forced session clears within a single working session, with loss of working context each time. Screenshots of the flagged exchanges are available if useful for review. Non recorded flags discussing plant nutrients and aerator pumps for water rooting fig cuttings

Another instance to add to the pattern above — this one has essentially zero hazard-adjacent content, which supports the idea that the trigger isn't purely content-based.
Context: mid-response discussing a friend helping dig up an established fig tree's root ball, and whether to bring a couple bags of topsoil as a polite gesture to patch the resulting hole in his yard. Flagged mid-sentence with no plausible unsafe subject matter nearby.
Request ID: req_011CfLQJLKSw7yK8vWy9ADEV

One more, same session: flagged mid-sentence again (req_011CfLRx1EfNiVYNa29iESP7) — this time while the assistant was literally in the middle of logging the previous false-positive's details for tracking. No hazard content anywhere near it.

Another completely benign one — req_011CfLSahFD3EdJu94a2rNyz, flagged responding to "trying to think of what to do next, did I forget to log anything or is there another education topic to add to the app." Casual planning talk, nothing remotely hazard-adjacent.