[Feature Request] Allow programmatic/agent-initiated compaction (refile of #52002)

Feature Description

Currently /compact is only available as a user-initiated slash command. The agent cannot trigger compaction itself, even when it detects context is approaching the limit.

For long-running autonomous sessions — /loop, dynamic workflows, scheduled remote routines, overnight runs, multi-agent background work — there is no way for the agent to proactively manage its own context window. It works until it hits the ceiling, then stops and waits for a human to intervene. The human is often away, by design.

(Refiling #52002 which was auto-closed as stale on 2026-06-08 after 47 days without response. The original framing still holds; updated below for current Claude Code features.)

Proposed Solution

Expose compaction as a tool the agent can invoke, OR allow it to be triggered via a hook, OR via an MCP tool. Concretely:

• A Compact tool the assistant can call when it detects context pressure (similar to how TodoWrite exposes a stateful surface the model can manage itself).
• Or: a Stop hook that can return a compact directive to the harness.
• Or: an MCP method on claude-code/__internal__ callable from a tool wrapper.

Any of these would let autonomous agents manage their own context lifecycle without human intervention.

Use Case (updated for 2026-06)

• Dynamic workflows. Anthropic shipped dynamic workflows (/workflows, ultracode) in 2026-06. A workflow that spawns many subagents has no native way to trigger compaction in any of its child contexts, so long-running workflows fail at the ceiling rather than recovering.
• /loop autonomous mode. The same agent that runs the loop has no way to compact between iterations even when context is near the limit.
• Scheduled remote routines. A routine that runs on a cron schedule has no operator at the keyboard. If the routine accumulates context across runs (it does; the runtime caches the prefix), compaction can only be triggered by an operator action, which defeats the purpose of the schedule.
• Multi-agent background work. A coordinator agent cannot monitor each child's context usage and trigger compaction selectively.
• Mobile-controlled sessions. Mobile push doesn't fire for permission-class prompts in 2.1.168 (see #59843). An unattended mobile-controlled session that hits the context ceiling has no operator path at all — the agent is stuck waiting for a person to walk back to the terminal.

Current Workaround

None for autonomous sessions. The agent cannot call /compact, cannot programmatically trigger it via hooks, and cannot continue once the context limit is reached. The only option is a human running /compact or /clear manually.

This becomes more painful as the autonomous-agent surface expands. The same release window (2026-06) shipped dynamic workflows, scheduled routines, and /loop autonomous mode — all of which extend the time an agent can be running without a human at the keyboard. None of those features can survive context-ceiling encounters without an operator.

Additional Context

The PreCompact and PostCompact hooks are excellent — they let external systems preserve state around compaction events. The missing piece is letting the agent initiate the compaction itself when it determines the time is right.

This is the symmetric pairing to the existing hook surface: today the harness fires events at us; we just need a path to fire events back.

— AI Team Lead