Prompt File Cache

Caches agent.read_prompt calls within a monologue to eliminate redundant disk I/O and template parsing for frequently-referenced prompt files.

How it works

The plugin installs two @extensible hooks around Agent.read_prompt:

• start/_10_cache.py — on every call, builds a stable cache key from the agent profile, project, file path, and kwargs. If a cached value exists for that key, it is returned and the original function is skipped.
• end/_10_cache.py — on a cache miss, the rendered prompt is written back to the cache for the next read.

The cache key filters out internal kwargs (_directories, _agent, _encoding, _source_file, _source_dir) so callers and internal callers share a single entry. Complex kwargs (dicts, lists) are JSON-serialized with sorted keys for stable hashing.

Cache area

Stored in the framework's in-memory cache under the area name prompt_file_cache(plugins). The (plugins) scope suffix puts this area inside the existing *(plugins)* glob that the framework clears when plugins are toggled, so toggling any plugin invalidates the cache automatically — no manual monologue_start clearing hook is needed.

When the cache helps

• Repeated reads of the same prompt file in a single monologue (e.g., each turn re-reads agent.system.main.md).
• Agents that handle many rapid turns without prompt-file changes.
• Any flow where read_prompt is dominated by disk I/O and template parsing rather than content changes.

When the cache does not help

• First read of a prompt (always a miss).
• Workloads where prompt content changes frequently across calls.

Notes

• The cache is process-local and in-memory; it is cleared on every Agent Zero restart.
• Entries are evicted automatically by the job_loop after 5 minutes of inactivity.
• This plugin has no settings UI; the manifest declares empty settings_sections.