Refresh agentd top: unified running view, fast startup, robust cursor restore, cheap LLM summaries

Diagnosis from current code:

• top only calls getRunningAgents over oneshot log directories and never loads persistent agents, so it misses currently running persistent sessions.
• main dispatch ignores the top --workspace argument (pattern is Top _ showAll -> getLogRoot ...), so explicit workspace/log-root override has no effect.
• runTop blocks before first render via initialAgents <- getRunningAgents ...
• getRunningAgents does one docker ps -f name=<run> per run directory; with many historical runs this dominates startup/refresh time.
• aiCurrentActivity is always Nothing, so there is no activity summary despite UI column.
• top uses customMainWithDefaultVty + neverShowCursor with no explicit final cursor restore fallback.

Proposed implementation plan: 1) Replace top data loading with a unified snapshot that merges oneshot + persistent running agents. 2) Fix top argument wiring so top --workspace is honored. 3) Eliminate per-run docker calls: query running containers once per refresh and intersect with run IDs. 4) Make refresh non-blocking (background worker + BChan updates); render immediately with loading state. 5) Add activity summary pipeline:

• deterministic fallback summary immediately from latest events/session lines,
• cheap LLM summary async via Omni.Agents.Summarize (haiku model),
• cache + fingerprint per agent so we only resummarize on new activity.

6) Add agent kind (oneshot/persistent) so Enter/l and k route correctly (logs and stop behavior). 7) Harden terminal teardown: explicit Vty lifecycle and finally block that always re-enables cursor. 8) Add tests for merge/filter logic + summary cache invalidation + mode-sensitive actions.

Implemented refresh of agentd top in Omni/Agentd.hs.

What changed:

• Removed top --workspace from CLI surface (per request). agentd top now only supports --all.
• Reworked top data model to unify both domains:
• oneshot container runs (from /var/log/agentd/<run>/events.jsonl)
• persistent agents (from Daemon.listPersistentAgents)
• Fixed the core visibility bug: running persistent sessions now appear in top.
• Removed per-run docker polling:
• old behavior did docker ps -f name=<run> per run directory
• new behavior does one docker ps --format {{.Names}} per refresh and intersects with run IDs
• Made startup responsive:
• top starts with an empty/loading state
• refreshes run via async snapshot jobs posted through BChan events
• Added activity pipeline:
• deterministic fallback activity from latest event/session lines
• async cheap LLM summary refinement via Omni.Agents.Summarize (haiku model)
• fingerprinted summary cache + in-flight dedupe to avoid redundant summarization
• graceful fallback mode if provider init fails
• Added kind-aware actions in top:
• Enter/l shows logs via oneshot or persistent backend correctly
• k stops oneshot containers or persistent sessions appropriately
• Added terminal teardown hardening:
• on top exit, always emit cursor-show escape in finally cleanup

Verification:

• typecheck.sh Omni/Agentd.hs
• lint Omni/Agentd.hs
• bild --test Omni/Agentd.hs
• agentd top --help now shows only [-a|--all].
• Interactive tmux smoke check confirms running persistent agents are now listed in top.

Adjusted top status footer to use fixed-width single-char refresh indicator to avoid horizontal jitter. Changed from 'refreshing/steady' to 'sync=R' (refreshing) and 'sync=.' (steady). Verified with typecheck.sh Omni/Agentd.hs.