Product or interface
CLI - interactive TUI
Use case and problem
Long-running MCode sessions change while they are already running: tools become relevant, instructions change, sessions resume, or context approaches compaction. Small runtime changes can force more prompt/tool state to be rebuilt or reprocessed than necessary, reducing prompt/KV-cache reuse and adding avoidable latency, token processing, and cost. The goal is to keep MCode’s behavior unchanged while making the runtime more cache-aware wherever the selected model/provider supports it.
Desired behavior
Allow MCode to preserve reusable cached context across small runtime changes: add/remove tools without unnecessarily losing the stable prompt prefix, introduce mid-conversation runtime instructions while preserving cache where supported, reuse or warm recent session context on resume, and let compaction/cache behavior adapt per model/provider. Unsupported providers should simply keep the current behavior.
Acceptance idea
existing cached session → small runtime change → reuse stable prefix → process only what changed
instead of:
existing session → small runtime change → rebuild/reprocess everything
Platform
Multiple platforms
Alternatives and additional context
Other agent harnesses are starting to expose ideas like dynamic tools without losing KV cache, mid-conversation system messages, cache warming, and faster resume. MCode already has prompt-cache awareness, context management, compaction, tools, MCPs, skills, and resumable sessions, so this feels like a natural next efficiency step rather than a change to how MCode fundamentally works.
Product or interface
CLI - interactive TUI
Use case and problem
Long-running MCode sessions change while they are already running: tools become relevant, instructions change, sessions resume, or context approaches compaction. Small runtime changes can force more prompt/tool state to be rebuilt or reprocessed than necessary, reducing prompt/KV-cache reuse and adding avoidable latency, token processing, and cost. The goal is to keep MCode’s behavior unchanged while making the runtime more cache-aware wherever the selected model/provider supports it.
Desired behavior
Allow MCode to preserve reusable cached context across small runtime changes: add/remove tools without unnecessarily losing the stable prompt prefix, introduce mid-conversation runtime instructions while preserving cache where supported, reuse or warm recent session context on resume, and let compaction/cache behavior adapt per model/provider. Unsupported providers should simply keep the current behavior.
Acceptance idea
existing cached session → small runtime change → reuse stable prefix → process only what changed
instead of:
existing session → small runtime change → rebuild/reprocess everything
Platform
Multiple platforms
Alternatives and additional context
Other agent harnesses are starting to expose ideas like dynamic tools without losing KV cache, mid-conversation system messages, cache warming, and faster resume. MCode already has prompt-cache awareness, context management, compaction, tools, MCPs, skills, and resumable sessions, so this feels like a natural next efficiency step rather than a change to how MCode fundamentally works.