Hook subsystem: session metrics in payloads + warm-reset primitive

Feature request: turn count / token usage / latency in hook payloads + in-place session-reset primitive

Context

Building a SaaS multi-tenant agent platform on Claude Code. Owner-mode (single user, technical operator) handles long-conversation performance degradation by manually restarting the session when slowness becomes obvious. SaaS subscribers will not do this. Subscribers expect always-on behaviour. "Claude is slow today" reads as "the product is broken" — direct churn signal.

What we tried

Wanted to build a hook that:

1. Detects long-conversation symptoms (turn count + cumulative tokens + per-turn latency).
2. At threshold, fires a save-state protocol (flush memory / close handoffs / write diary line).
3. Triggers an in-place session reset that preserves cwd + state pointers for the new session to read.
4. Surfaces to subscriber as "Refreshing your assistant — new session ready in 3 seconds."

What's blocked

Per Claude Code hooks reference (https://code.claude.com/docs/en/hooks.md):

• No turn count exposed to hooks. Stop / SessionStart / PreToolUse payloads don't include cumulative turn metrics.
• No token usage exposed. Same.
• No per-turn latency exposed. Same.
• No in-place session-reset primitive. Stop hooks can claude logout (kill the process) and let the user re-launch — but this loses the working directory, breaks the subscriber UX, and risks data loss if save-state hasn't run yet.

Workaround we shipped (acknowledged as fragile)

A Stop hook (monitor-length.ps1) that uses transcript-size-on-disk as a proxy for token count + line count as a proxy for turn count. Suggests a restart when thresholds cross. NEVER auto-restarts. NEVER kills the session. Surfaces via additionalContext from a UserPromptSubmit hook.

This works for owner mode (operator sees the suggestion, decides whether to restart). It does NOT work for SaaS subscribers — they'd just see "system suggests restart" and either ignore it (slowness continues) or click "restart" and lose context.

What we need

Three additions to Claude Code's hook subsystem:

1. Conversation metrics in hook payloads

Add to Stop / SessionStart / PreToolUse / UserPromptSubmit payloads:

{
  "session_metrics": {
    "turn_count": 47,
    "cumulative_input_tokens": 142000,
    "cumulative_output_tokens": 28000,
    "cumulative_total_tokens": 170000,
    "last_turn_latency_ms": 8420,
    "session_start_iso": "2026-05-07T16:42:00+08:00",
    "context_window_size": 200000,
    "context_utilisation_pct": 85
  }
}

This unblocks accurate threshold detection. The context_utilisation_pct is the killer metric for SaaS — at 85%, the next compaction is imminent and likely expensive.

2. In-place session-reset primitive

A hook output schema that lets a Stop hook trigger:

{
  "hookSpecificOutput": {
    "hookEventName": "Stop",
    "sessionAction": "warm_reset",
    "warmResetCarryover": {
      "preserve_cwd": true,
      "carry_state_files": ["MEMORY.md", "session-diary.md"],
      "user_facing_message": "Refreshing your assistant — new session ready in 3 seconds."
    }
  }
}

Semantics: same Claude Code process, same cwd, but conversation context is reset to empty + state files pointed to in carry_state_files are auto-read at the new SessionStart.

3. Optional: opinionated long-conversation-mitigation as platform feature

Even with the above two primitives, every SaaS-grade Claude Code deployment will need to build the same thing. Worth Anthropic shipping a platform-level "auto-mitigate long-conversation degradation" toggle that uses #1 + #2 internally and lets the operator just say "yes, autopilot." Reduces the build surface from N teams each implementing it to one.

Priority justification

• For a single-tenant deployment: medium. The fragile workaround is acceptable while owner-mode dominates.
• For a SaaS roadmap: hard blocker. Cannot take a paying subscriber across a degradation event without #1 and #2 in place.
• For Anthropic: any team building agent platforms on Claude Code hits this exact wall the moment they try to support always-on subscribers. Shipping Managed Agents at $0.08/session-hour with per-subscriber-per-session-hour billing makes long-conversation handling an Anthropic-side concern, not just a customer-side one.

How we'd help

Happy to share:

• Empirical metrics from owner-mode sessions — when the perceived-slow threshold actually lands (turn count, transcript size, etc.).
• Design sketches for the warm-reset semantics (carry_state_files, cwd preservation, sub-second user-facing message).
• A sample SaaS subscriber UX flow showing where the platform boundary should land.