Cloud: updating an agent's system does not recompile the in-context system message #3446

Summary

On Letta Cloud (api.letta.com, client @letta-ai/letta-client 1.12.1), PATCH /v1/agents/{id} with a new system updates the agent record but the system message the model actually receives is unchanged: the oldest message of the agent (role system) still carries the previous prompt, and the Anthropic prompt-cache figures on later turns show the same cached prefix.

Repro

1. Create an agent with system = prompt A (ours contains a unique marker sentence per version).
2. Send a few messages so the agent has history.
3. client.agents.update(agentId, { system: promptB }) → GET /v1/agents/{id} returns system = prompt B (2,158 chars in our case).
4. GET /v1/agents/{id}/messages?limit=3&order=asc → the first message (message_type: system_message) is 10,829 chars and contains only prompt A's marker, not prompt B's.
5. Send three more messages. Each run's step usage (via runs.steps.list) shows cached_input_tokens ≈ the whole previous request (15.4k → 15.5k → 15.7k) and cache_write_tokens ≈ 120 per turn, i.e. the prefix sent to Anthropic did not change; a new system prompt would have invalidated it. The model's replies also kept following prompt A's rules and ignored prompt B's new instruction.

Observed today (19 Sep 2026) on agent agent-780de0e0-f3c5-4745-ad1c-79c7cebe68a5 (BYOK Anthropic, model handle cli-anthropic-key/claude-sonnet-5). Happy to share run ids.

Expected

Either the update recompiles the in-context system message (as memory block edits do), or the docs say system is creation-only on Cloud so callers rebuild. We now rebuild the agent from our own mirror on every prompt version bump, which works but destroys the message history.

Related

#3445 (tool_rules accepted but not persisted on Cloud).