Summary

Refs #526, supersedes #528.

Integrates Anthropic's server-side context management (beta: context-management-2025-06-27). When enabled, the API automatically clears old tool use/result pairs and thinking blocks after prompt cache lookup but before token counting — preserving cache prefixes while freeing context space.

How it works

The client sends the full unmodified message array + a context_management parameter:

1. Anthropic checks prompt cache → cache hit (messages unchanged)
2. Applies edits server-side (removes old thinking/tool results)
3. Counts tokens on edited version (now smaller)
4. Runs inference on edited version

This is impossible with client-side stripping, which modifies the message array and breaks cache prefixes.

Implementation

• anthropic_adapter.py: Added context-management-2025-06-27 beta header. build_anthropic_kwargs() builds context_management edits via extra_body. clear_thinking only included when reasoning is enabled (API requirement).
• run_agent.py: Pipes context_editing config through AIAgent to the adapter
• cli.py / gateway/run.py: Loads context_editing config from config.yaml
• hermes_cli/config.py: context_editing section in DEFAULT_CONFIG
• tests/test_context_editing.py: 8 tests covering all scenarios

Config (opt-in)

context_editing:
  enabled: true
  trigger_tokens: null          # auto: 60% of context window
  keep_tool_uses: 5             # recent tool pairs to preserve
  keep_thinking_turns: 2        # recent thinking turns to preserve
  exclude_tools:                # never cleared
    - memory
    - skill_manage
    - todo
  clear_tool_inputs: false
  clear_at_least_tokens: null   # auto: 10% of context window

Live tested with Anthropic API

• ✅ Single turn with context_management — accepted, response normal
• ✅ Multi-turn with tool calls + thinking + context editing — works end-to-end
• ✅ clear_thinking correctly omitted when reasoning is disabled
• ✅ Config plumbing verified through AIAgent._build_api_kwargs()
• ✅ 3437 tests pass (6 pre-existing failures)

Notes

• Direct Anthropic only — works with provider: anthropic. OpenRouter passthrough is unverified.
• Disabled by default — opt-in via context_editing.enabled: true
• Works alongside compression — context editing is first line of defense (cheap, cache-friendly), compression kicks in later if needed (expensive, cache-breaking)