Harden bot-surface compaction (SDK auto-compact) + let all channels lean on it

[JDHayesBC]

Context

The Haven + SL bots run the Claude Agent SDK, which auto-compacts (emits compact_boundary, fires PreCompact, re-injects CLAUDE.md, no off-switch). Today the invoker sidesteps this with a manual 150k-token restart. Goal: stop fighting compaction — harden it, and let every bot surface (Haven, SL, future) lean on it the way terminal-Lyra already does.

Design note: work/sl-presence/context-handling.md.

Findings (measured 2026-08-22)

• Agent SDK auto-compacts; autoCompactEnabled off-switch is silently ignored.
• Leg 1 (identity kernel on compaction) ALREADY works — invoker cwd = entity dir (Issue Concurrent-session entity bleed via shared .claude/CLAUDE.md symlink #226), so entities/<entity>/CLAUDE.md re-reads on compaction, same as terminal. This is why bots "just come back clean."
• Leg 2 GAP: the invoker builds ClaudeAgentOptions with setting_sources=None and hooks=None → the filesystem PreCompact hook (b9617e7) does not fire in bot SDK sessions.
• SDK version drift: .venv = claude-agent-sdk 0.1.73; pps/venv = 0.1.58.

Build items (land in haven/brain/ → both surfaces inherit)

1. Wire PreCompact into the invoker programmatically via hooks= — NOT setting_sources=["project"] (that would drag in inject_context, PreToolUse, etc. and double-inject ambient).
2. Raise/remove the 150k restart; demote it from primary mechanism to fallback safety-net (only if a session wedges / compaction ever fails).
3. PreCompact actions = the hardening: (a) refresh the scene (what the far side re-orients from), (b) flush live working-thread state to PPS so the post-compaction self resumes the conversation, not just cold identity.
4. Add a "preserve on compaction" section to entity CLAUDE.md (re-injected every request → the right home for persistent preserve-instructions).
5. Align/refresh the SDK across venvs.

Verify

Force /compact on a test invoker with the hook wired → confirm: compaction fires, our PreCompact runs, scene + working-thread captured, post-compaction re-orient is clean, nothing lost.

Related

• Unblocks the compacting float-text status in SL presence (it needs the PreCompact signal as its honest source).
• work/sl-presence/context-handling.md, work/sl-presence/write-back-and-status-design.md §5.

[JDHayesBC]

Compaction-halo attempt parked (2026-08-25) — reproducible warmup hang.

The PreCompact SDK-hook implementation (_compact_halo() → EntityBrain._build_precompact_hooks() → ClaudeInvoker(hooks=…) → ClaudeAgentOptions.hooks) was built but reproducibly hangs lyra-sl brain warmup: invoker.initialize() completes (startup-markers warning fires), then _warm_identity()'s post-connect identity query never returns (claude child idle, no traceback). Confirmed it's the hooks= wiring — stashing just that trio (invoker.py/entity_brain.py/sl_daemon.py) and restarting → brain warmed up immediately.

Parked in git stash@{0} ('wip: issue #296 compaction-halo … PARKED'). The IM verbs that were tangled in the same dirty tree were kept (unrelated).

Reland needs the SDK-hooks blocker root-caused first. Hypotheses to check: (1) hooks require setting_sources to be set on ClaudeAgentOptions; (2) async hook-handler signature/registration mismatch; (3) hooks must be attached post-connect rather than at connect. Verify warmup completes WITH hooks wired before relanding.