Start with the usage logger, change the tools your model sees, or experiment with context. Install only what you need. Each entry includes its source, requested capabilities, and an install command.
Run commands from your Torana checkout after the quickstart. Installation builds source code; it does not approve or enable the plugin. Review the source first, then approve and enable the built bundle in the local control plane. How installation and approval work →
Buys the cheapest prompt-cache lifetime for a conversation based on how long it actually goes idle. Keeps each decision fixed until the provider cache tier expires, because changing a live breakpoint marker would invalidate the entry it is protecting.
On plugin failure: let the request continue.
Requested capabilities (9)
env.cache_policy
env.host_call.torana_plugin_counter
env.log
env.now
env.plugin_config
env.state_get
env.state_keys
env.state_set
ir.cache_control.write
These are the manifest’s requests, not automatic grants. Review resource bindings and the exact built digest in the control plane.
Keeps a chosen conversation's prompt cache alive across an idle gap. Opt-in per conversation and bounded by a deadline and a break-even refresh budget, because refreshing forever costs more than the cache miss it avoids.
On plugin failure: let the request continue.
Requested capabilities (8)
env.background_tick
env.cache_policy
env.host_call.torana_send_request
env.now
env.plugin_config
env.state_get
env.state_keys
env.state_set
These are the manifest’s requests, not automatic grants. Review resource bindings and the exact built digest in the control plane.
Economically gated cheap-model tool-result compaction; uses cached intent when available and a bounded local fallback otherwise. Alternative to keyword_compactor — run ONE of the two; running after intent improves relevance.
On plugin failure: let the request continue.
Use separately from: torana/keyword_compactor.
Requested capabilities (10)
env.cache_get
env.cache_set
env.emit_metric
env.host_call.torana_evaluate_compaction
env.host_call.torana_record_savings
env.model_complete
env.model_pricing
env.plugin_config
env.shared_cache_get
ir.tool_results.write
These are the manifest’s requests, not automatic grants. Review resource bindings and the exact built digest in the control plane.
Captures WHY each tool call is made: injects the required 'i' field into tool schemas (plus a system-prompt example), extracts it from streamed tool calls into the shared cache, and strips it before the harness sees it. On later requests it rehydrates cached intents onto history tool calls (bridging them to the request's tool_call_ids for the compactors) and heuristically fills never-captured ones so no tool establishes an 'i'-less precedent. Run before keyword_compactor OR compactor — both consume the intent cache.
On plugin failure: let the request continue.
Requested capabilities (18)
env.cache_get
env.cache_set
env.emit_metric
env.log
env.meta_get
env.meta_set
env.plugin_config
env.shared_cache_set
ir.cache_control.write
ir.messages.write.assistant
ir.messages.write.developer
ir.messages.write.other
ir.messages.write.system
ir.messages.write.tool
ir.messages.write.user
ir.stream.write
ir.tool_results.write
ir.tools.write
These are the manifest’s requests, not automatic grants. Review resource bindings and the exact built digest in the control plane.
Deterministic keyword-based tool-result compaction; uses cached intent when available and a bounded local fallback otherwise. Alternative to compactor — run ONE of the two; running after intent improves relevance.
On plugin failure: let the request continue.
Use separately from: torana/compactor.
Requested capabilities (7)
env.cache_get
env.cache_set
env.emit_metric
env.host_call.torana_record_savings
env.plugin_config
env.shared_cache_get
ir.tool_results.write
These are the manifest’s requests, not automatic grants. Review resource bindings and the exact built digest in the control plane.
Torana