Summary

Replaces the basic /status output with a comprehensive session snapshot showing model, provider, token breakdown, cost, context window, and more.

• Token breakdown: input/output/cache read/cache write/reasoning with totals
• Cost display: estimated or actual cost with human-readable labels
• Context window: live usage from context_compressor when agent is running; idle fallback via get_model_context_length when agent is idle (inspired by feat(status): show model, provider, context, and cumulative tokens #4678)
• Cache hit rate: percentage of cache reads vs total input
• Compression tracking: compaction count from live agent OR persisted compression_count for idle sessions — idle sessions now show compression history instead of omitting it
• Platform categorization: separates chat platforms from services (API, webhook)
• Queue depth: pending messages and follow-up items
• Model override awareness: shows session-level model overrides
• Transport mode: webhook vs polling for Telegram

Fixes #5960 (Tokens: 0 regression — reads from SessionDB instead of stale SessionStore).
Closes #7317 (context/compression state in gateway-visible status).
Closes #7714 (usage display like OpenClaw).
Supersedes #4678 (model + context in /status — this PR is a strict superset).
Supersedes #5989 (get_session_token_totals — included here plus much more).

Changes

Example output

Gateway (Telegram/Discord) — live agent:

⚕ Hermes Agent v0.8.0
Model: `anthropic/claude-sonnet-4` · Provider: anthropic
Usage: 25,000 in · 610 out · 50,730 total · Cost: $0.1234 est.
Cache: 25,000 read · 0 write · 50% hit · 120 reasoning
Context: 28,000 / 200,000 (14%) · Compactions: 1
Session: `agent:main:telegram:dm:12345` · updated 3m ago
ID: `20260412_104559_e1760b66`
Runtime: Chat Completions · Reasoning high · Transport webhook
Queue: depth 0 · State: running
Chats: telegram, discord

Gateway — idle session (compression history persisted):

⚕ Hermes Agent v0.8.0
Model: `openai/gpt-4o` · Provider: openai
Usage: 100 in · 200 out · 3,210 total · Cost: $1.2345 est.
Cache: 10 read · 5 write · 9% hit
Context: 45,000 / 200,000 (22%) · Compactions: 3
Session: `agent:main:telegram:dm:12345` · updated 7m ago
ID: `20260412_104559_e1760b66`
Runtime: Reasoning medium
Queue: depth 0 · State: idle
Chats: telegram

Design decisions

• Shared helpers, not duplicated methods: hermes_cli/status_format.py holds all formatting logic. Both CLI and gateway import from it.
• Graceful degradation: Every data source is wrapped in try/except with safe defaults. Missing SessionDB, missing agent, missing cache — all handled.
• Idle context + compression fallback: When no live agent exists, uses get_model_context_length() + session_entry.last_prompt_tokens for context, and session_entry.compression_count for compaction history. Both are persisted via update_session() after each turn.
• Backward-compatible persistence: compression_count defaults to 0 in SessionEntry.from_dict(), so existing sessions.json files load cleanly.
• Backward-compatible attributes: Uses getattr() with defaults for optional runner attributes (_pending_hidden_turns, _agent_cache, etc.).

Test plan

• pytest tests/cli/test_cli_status_command.py tests/gateway/test_status_command.py tests/hermes_cli/test_status_format.py — 42 tests pass
• Live-tested on downstream fork with 59 regression tests passing
• Idle compression count verified via test with compression_count=3 on SessionEntry

🤖 Generated with Claude Code