fix: harden Discord sync and cron memory opt-in

[quantum-chaos]

Summary

• delete stale Discord slash commands before creating/updating desired commands to avoid command-cap failures
• allow cron jobs to explicitly opt into memory tools with allow_memory_tools
• preserve safe uploaded image filename metadata in API-server multimodal normalization for SKU-aware workflows

Test Plan

• uv run --with pytest python -m pytest tests/tools/test_cronjob_tools.py tests/cron/test_scheduler.py tests/gateway/test_api_server_normalize.py -q -o 'addopts='
• Result: 201 passed in 3.24s

[daimon-nous]

The Discord slash command sync fix (delete-before-create reorder) duplicates the approach in #24371, #29134, and #19366. The cron allow_memory_tools portion addresses #18885 and relates to #29301, #9802.

[teknium1]

Thanks for the focused regressions. The API filename and cron-tool access ideas still address behavior absent from current main, but the cron implementation needs rework; the Discord portion is already implemented.

Problems

• The added skip_memory=False path initializes the external provider, then current turn finalization automatically calls sync_all and queue_prefetch_all (run_agent.py:3416-3424). Suppressing system-prompt memory does not prevent automatic persistence, so this does not provide the stated deliberate-tools-only behavior.
• Discord deletion-before-create is already on main in plugins/platforms/discord/adapter.py:1897-1914, with the near-cap regression in tests/gateway/test_discord_sync_limit.py:61-140 (commit e9b86f352fc73db5ca3de6e3fb50ef57d774f8f9).

Suggested changes

• Re-scope the cron work to a tools-only provider mode that blocks automatic prefetch/sync, and add an integration test using a provider that records lifecycle calls. agent/memory_provider.py:74-76 specifically identifies cron as a context providers should skip for writes.
• Drop the superseded Discord hunk when salvaging.

This is an automated hermes-sweeper review.

[teknium1]

Setting skip_memory=False enables the full external-provider lifecycle, not only its tools. Current turn finalization unconditionally calls sync_all and queue_prefetch_all for an initialized manager (run_agent.py:3416-3424), so this path automatically persists cron turns despite the tools-only claim. Please separate provider-tool exposure from prefetch/sync lifecycle activation and cover that with an integration test.