[Improve] Fast sessions acknowledge users before substantive work

[roomote-roomote]

​Opened on behalf of @daniel-lxs. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Fast's system prompt now opens with a high-priority Turn Startup workflow before environments, integrations and Brain instructions, native tools, or other work policy.
• For substantive human-authored turns, the model is told to make its first action a user-visible acknowledgement, then continue the same turn through Brain recall or other work and finish normally.
• Immediate terminal answers, clarifications or reactions remain single responses. Eligible ambient conversation may still use ignore_event, and trusted platform events continue to follow their dedicated no-ack rules.
• Later subagent, task launch and steering, and custom automation guidance now points back to the same startup rule instead of carrying contradictory exemptions.
• The broad service, bridge, OpenCode permission, and event-monitor enforcement from the earlier revision has been removed. Only Roomote's pre-existing narrow MCP/cancel acknowledgement safeguard remains.

Why this change was made

Fast prompts are a first-class behavior surface, and the earlier fail-closed gate added substantial lifecycle machinery across every tool path. The prompt-first design makes the desired ordering prominent before the model sees Brain or work capabilities, while preserving conversational efficiency and avoiding new OpenCode permission/session coupling.

OpenCode 1.18.10 does not expose a public named first-tool choice that could replace this with a simpler native guarantee. Its session.prompt request type supports only boolean tool filtering; structured output is terminal; and tool.execute.before can block execution but cannot force selection and depends on plugin delivery (docs, #41422, #45367).

Impact

Substantive human Fast turns are strongly steered to acknowledge before Brain recall, delegation, integrations, or other work, without adding a second inference or a universal runtime interception layer. This is intentionally a softer behavioral contract: prompt ordering and content provide the primary guarantee, while the existing narrow MCP/cancel check remains as defense in depth. Prompt tests pin the rule ahead of Brain and tool catalogs and cover immediate, ambient, and platform-event exceptions; the full cloud-agents suite passes.

[roomote-community]

No new code issues found. See task

_{Reviewed 2f56c07}