fix: clear scheduled job context before each invocation (#301)

[kindralai]

Problem

Recurring scheduled jobs (e.g. scheduler:proactive-mail-triage, which runs hourly between 8 AM and 7 PM) accumulate conversation history across invocations. Each run adds its input and output to the same history key. Over time this causes:

• Growing token costs — each run starts with the full history of all previous runs loaded into context
• Cache invalidation — the long context means the model's prompt cache rarely hits
• Degraded relevance — the model sees old job outputs mixed with the current task
• Eventual slowdown — as the history grows, the LLM call takes longer and costs more

Issue #235 / PR #236 gave each job its own context key, but a single recurring job's own context still grows unbounded.

Solution

Send a /new command via agent.process() to clear the job's history key before each invocation. Applied to both run_agent_task() and run_subagent_task().

Note on respond=: The issue suggested respond=False, but the _process_impl respond-gate (line 1348) returns early before reaching the /new handler when respond=False — it just records the command as an inbound message. Using respond=True processes the /new command and clears context. The "Conversation cleared." response is discarded by the caller (not assigned or delivered).

Changes

File	Change
humux/core/scheduler.py	Add /new context clearing in run_agent_task() and run_subagent_task() before the actual task runs

Acceptance criteria

• Recurring job's context is cleared before each invocation
• One-shot jobs also clear context (consistent behaviour)
• run_subagent_task() gets the same treatment
• No "Conversation cleared." message is delivered to the user chat
• Existing fix from Scheduled jobs share the same conversation context — history leaks and jobs are serialised #235 (per-job context keys) is preserved
• Silent/email jobs are unaffected (they already suppress output)
• All existing scheduler tests pass (19/19)

Closes #301

[mattmezza]

LGTM

merge it @kindralai[bot]