work-ledgerlisted

# The work ledger A run that tries to build everything at once takes an hour, cannot be interrupted, and gives the user nothing to look at until it is finished or broken. It also drifts: research on long-horizon agents is consistent that step-by-step agents lose the goal because each locally-best move pulls slightly away from it, while **plan-ahead agents hold**. The named failure modes are losing track of earlier decisions, declaring half-finished work done, and quietly changing what is being built. The fix is not a longer prompt. It is to make the plan an **artifact the agent re-reads**, rather than a memory it is trusted to keep. ## Not everything gets a ledger A plan is for work large enough that the plan earns its cost. A typo does not need slices, an approval loop and a journal — that is friction wearing rigour's clothes, and it teaches the user to route around the tool. `trivial` work skips all of this: a checkpoint, the change, a report. The entry command handles it before this skill is ever invoked. Everything `bounded` or larger gets a ledger. If work classified trivial turns out not to be, stop and re-classify rather than carrying on. A one-line change that becomes three files is a misclassification, and continuing is how a small change becomes an unreviewed large one. ## Two files, and the split is the point **`.devteam/plan.md` — the contract.** Drafted at intake, shown to the user in full, revised until they approve it, and never rewritten afterwards. T