Pocket Watcher

AI coding agents are excellent at starting work, but they rarely understand whether the work they are starting can actually finish inside the compute available. Planning expands, refactors appear, optional features get started, and a usage window can end with several unfinished pieces instead of one shippable product. Pocket Watcher treats compute as a project constraint. A project receives a fixed Ship Contract containing an assigned compute budget, protected finishing reserves, feature priorities, and bounded feature wallets. Pocket Watcher runs features through IBM Bob, records real Bob session costs, and continuously reforecasts what can still fit. When actual execution becomes more expensive than estimated, Pocket Watcher does not silently increase the budget. It enters COMPRESS and blocks additional feature execution until optional scope is deferred and the remaining plan becomes affordable again. As the project approaches completion, LAND prevents new scope while preserving validation and repair. Completion is also separated from agent success messages. A Bob run returning successfully does not automatically mean a feature is finished. Pocket Watcher can run deterministic acceptance checks at zero Bobcoin cost before marking work complete. In our verified COMPRESS proof, the project began with a fixed 0.0900 Bobcoin envelope and 0.0400 Bobcoins protected for finishing. Real execution caused the original scope to become unsafe, Pocket Watcher blocked additional work, an optional feature was deferred, and the required work continued. Total Bob spend was 0.039908 Bobcoins, the assigned budget never increased, the protected reserve remained intact, and all four final project checks passed. Pocket Watcher is an installable TypeScript CLI with a Bob Skill as its conversational front door. The final implementation passes 152 automated tests and was also audited for Windows, macOS, and Linux portability. Pocket Watcher does not compress tokens. It compresses unnecessary scope, planning and tangents so limited compute is spent on finishing the most important useful outcome.