ROADMAP.md
ROADMAP.md
Each of these was considered during the 0.3.0 work and left out on purpose. They are listed with the reason so a later release does not rediscover them.
.agentic/. Nothing prunes transactions, job logs, run transcripts or research sources. A long-lived workspace grows without bound and the user is the only cleanup mechanism. A policy has to decide what a rollback is still allowed to reach.semantic_search is lexical chunk ranking. An embedding index would need a model, a storage format and an invalidation story, none of which exist yet.taskkill /PID <pid> /T /F, which walks the whole tree. POSIX still sends SIGTERM to the direct child and SIGKILL two seconds later, so a grandchild spawned by an allowlisted script can outlive cancel. detached: true plus a kill(-pgid) would close it, at the cost of changing signal delivery for every POSIX job.close-never-fires case. Node emits close only once every inherited stdio pipe is released. A grandchild that survives its parent and keeps those pipes open leaves close pending, so the job never settles on its own; the tree kill covers this on Windows, and on POSIX the job stays running until the grandchild exits. Detecting it needs exit-plus-timeout bookkeeping separate from close.pending_approval. Wording it per mode would read better for the model but the schema block is re-sent every turn and is already ratcheted at its measured budget, so it would have to be paid for elsewhere.<agentic-mode> hint for small context windows. The hint is bounded but not scaled: an 8192-token model pays the same ~2 KB as a 32768-token one. A shorter variant below some context threshold is the obvious fix and needs a rule for which sentences may go.workspace_inspect changes lists command jobs by scanning .agentic/jobs/*/state.json. That is fine for tens of jobs and linear in the number of directories; a workspace with thousands wants an index file maintained on write.<agentic-mode> hint reports an approved plan's age once it is over an hour old. workspace_plan current and capabilities report the plan but not its age, so a model that reads them instead of the hint cannot tell a fresh approval from a 23-hour-old one.Each of these was considered during the 0.3.0 work and left out on purpose. They are listed with the reason so a later release does not rediscover them.
.agentic/. Nothing prunes transactions, job logs, run transcripts or research sources. A long-lived workspace grows without bound and the user is the only cleanup mechanism. A policy has to decide what a rollback is still allowed to reach.semantic_search is lexical chunk ranking. An embedding index would need a model, a storage format and an invalidation story, none of which exist yet.taskkill /PID <pid> /T /F, which walks the whole tree. POSIX still sends SIGTERM to the direct child and SIGKILL two seconds later, so a grandchild spawned by an allowlisted script can outlive cancel. detached: true plus a kill(-pgid) would close it, at the cost of changing signal delivery for every POSIX job.close-never-fires case. Node emits close only once every inherited stdio pipe is released. A grandchild that survives its parent and keeps those pipes open leaves close pending, so the job never settles on its own; the tree kill covers this on Windows, and on POSIX the job stays running until the grandchild exits. Detecting it needs exit-plus-timeout bookkeeping separate from close.pending_approval. Wording it per mode would read better for the model but the schema block is re-sent every turn and is already ratcheted at its measured budget, so it would have to be paid for elsewhere.<agentic-mode> hint for small context windows. The hint is bounded but not scaled: an 8192-token model pays the same ~2 KB as a 32768-token one. A shorter variant below some context threshold is the obvious fix and needs a rule for which sentences may go.workspace_inspect changes lists command jobs by scanning .agentic/jobs/*/state.json. That is fine for tens of jobs and linear in the number of directories; a workspace with thousands wants an index file maintained on write.<agentic-mode> hint reports an approved plan's age once it is over an hour old. workspace_plan current and capabilities report the plan but not its age, so a model that reads them instead of the hint cannot tell a fresh approval from a 23-hour-old one.