fix(caching): honor prompt_caching.cache_ttl disable in config

[BB-light]

Summary

Allows operators to fully disable Anthropic prompt caching by setting cache_ttl to a falsy value in config.yaml.

The problem

The existing prompt_caching.cache_ttl config only accepts "5m" or "1h". Any other value (including false, null, "off") is silently ignored, and caching remains enabled at the "5m" tier. This means:

• Operators who set cache_ttl: false expecting to disable caching see no effect
• Cache writes continue, incurring write costs (1.25x for 5m tier) that bill against "extra usage" on OAuth subscription
• There is no config lever to disable prompt caching without code changes

The fix

After the existing "5m" / "1h" check, a new elif branch catches falsy values:

# config.yaml
prompt_caching:
  cache_ttl: false    # also accepts: null, "off", "disabled", "no", "none"

When detected, sets:

• agent._use_prompt_caching = False
• agent._use_native_cache_layout = False
• agent._cache_ttl = None

Changes

• Modified: agent/agent_init.py — falsy-value detection after existing TTL parsing
• Modified: hermes_cli/config.py — updated inline doc comment
• Modified: tests/run_agent/test_run_agent.py — parametrized test covering all 7 falsy variants

Test plan

• pytest tests/run_agent/test_run_agent.py -k cache_ttl passes (existing + new tests)
• cache_ttl: "5m" still works (existing test)
• cache_ttl: "1h" still works (existing test)
• cache_ttl: "30m" (invalid) still falls back to "5m" (existing test)
• cache_ttl: false disables caching (new test, 7 variants)

[teknium1]

Thanks for identifying a real current gap: main documents prompt caching as always-on, and cache_ttl currently honors only "5m" and "1h" (website/docs/user-guide/configuration.md:915-924; agent/agent_init.py:643-650).

Problems

• The PR disables caching only during initialization (agent/agent_init.py PR line 497). Current model switching re-derives both flags at agent/agent_runtime_helpers.py:2019-2027, and fallback activation does the same at agent/chat_completion_helpers.py:1596-1604; either path can re-enable markers.
• The added test covers only initialization (tests/run_agent/test_run_agent.py PR lines 934-955), not those re-derivation paths.
• The config surface needs deliberate resolution. The related prompt_caching.enabled implementation was reverted in #56126 for broader evaluation; current docs still define cache_ttl only as a "5m"/"1h" tier selector.

Suggested changes

• Put the chosen opt-out in the shared cache-policy decision path and test init, model-switch, and fallback behavior.
• Document one selected configuration contract rather than treating TTL aliases as a second disable API.

Automated hermes-sweeper review.

[teknium1]

This only affects initial construction. Current main overwrites both flags when switching models (agent/agent_runtime_helpers.py:2019-2027) and when activating a fallback (agent/chat_completion_helpers.py:1596-1604), so cache_ttl: false can be lost mid-session. Put the selected opt-out in the shared policy path and cover those re-derivations.

[kshitijk4poor]

Merged via PR #76015 — your fix was cherry-picked with authorship preserved. The model-switch/fallback survival gap (the same issue that caused #56105 to be reverted in #56126) was fixed on top by gating anthropic_prompt_cache_policy() on a _cache_disabled flag set during init. Thanks for the contribution!