Skip to content

Latest commit

 

History

History
68 lines (53 loc) · 5.34 KB

roadmap.md

File metadata and controls

68 lines (53 loc) · 5.34 KB

Roadmap

Taskrivet's next goal is qualified, source-authorized ticket execution: choose an appropriate model and provider from current evidence, run within explicit bounds, recover without duplicate work, and return verifiable results for source-owned review.

Delivery board · Milestones · Selected roadmap issues · Design discussions

Current foundation

The alpha has persistent terminal tabs, session reconnection after desktop quit, a task inspector, bounded context, a local tool catalog, failure records, and a supervised single-ticket CLI loop. GitHub activity observations and confirmed reports retain source links and delivery receipts.

Automatic source claims, verified provider selection, unattended ticket dispatch, service-crash/reboot recovery, and automatic acceptance are not complete. The execution contract describes those boundaries; the long-term plan retains the broader design.

Next milestone: qualified ticket execution

These are existing issue identities selected from the planning catalog. They are proposed contracts, not declarations that work is ready or authorized. Each needs current scope and prerequisite review before implementation. No release date is promised.

Order Outcome and acceptance focus Existing issue
1 Atomic source claims. Competing workers cannot both own a ticket; changed revisions or new holds refuse launch; an unknown claim reply is reconciled by stable attempt ID. #252
2 Lease renewal and revocation. Renew only while source authority is valid; expire or revoke without replaying a running attempt; report unresolved process state. #253
3 Model-specific qualification. Distinguish installed/catalogued tools from fresh generation evidence, exact provider/model identity, capability, entitlement, and quota. Unknown evidence remains unknown. #408
4 Explainable route selection. Bind the recommendation to the ticket revision and qualified launcher; show rejected alternatives and evidence confidence; invalidate changed conditions before launch. #452
5 Cost and fallback bounds. Explicitly allowed provider/model routes and spending limits; no automatic paid fallback from a local route; distinguish measured and declared usage. #454
6 Reviewable completion receipts. Ordered, idempotent attempt evidence tied to scope, source revision and artifacts; replay-safe delivery and separate source acceptance. #551
7 Continuous-work controls. Visible pause state prevents new admission; resume revalidates source authority; existing process state is reported independently of scheduling state. #701
8 Crash reconciliation. On restart, reconcile persisted attempt identities with actual service/process state; unknown outcomes never become successful completion or an automatic rerun. #3

A synthetic end-to-end acceptance run must cover competing claims, a hold changing during preparation, expired qualification, provider failure, pause, uncertain launch acknowledgement, restart, receipt replay, and source review refusal. A green unit suite alone does not qualify unattended execution on a real machine.

Contributor experience and maintenance

The maintenance milestone covers usable documentation, reproducible bug reports, small PRs, dependency and security checks, and truthful release notes. Releases remain intentional: a merged PR, package build, or project-card movement does not publish a desktop application or establish end-to-end acceptance.

Use bug, enhancement, and documentation for change type; area:* labels for the affected subsystem; needs-triage for scope/readiness review; and roadmap for the focused milestone. GitHub's project status describes development progress, not permission to execute ticket work. A good first issue should be assigned only after a maintainer identifies a genuinely bounded, independent contribution.

Planning catalog

The repository contains a large generated capability catalog, including repeated slices. Those entries preserve external identities and are not a ready queue or proof of a missing feature. Refine one issue against current source before taking it on. Link overlaps and prerequisites; do not bulk-close entries as completed without evidence. The focused board intentionally starts with a small set instead of importing every proposed slice as committed work.

Longer-term work includes workspace isolation, context retrieval and invalidation, measured outcomes by task class, bounded recovery strategies, accessible desktop controls, and signed distribution. Discuss changes to those priorities in Discussions and update the relevant issue and milestone together.