subagent-driven-development adds progress.md ledger to stop rework

The workspace itself exists in a specific place for a specific reason: bash scripts/sdd-workspace PLAN_FILE prints an absolute path under the repository root, one directory per plan file base name. The script's comments explain that it is kept out of .git/ because Claude Code treats .git/ as a protected path and refuses agent writes, which would leave implementation subagents unable to write their reports.

The ledger's first line identifies which plan file the ledger belongs to, and that line drives what happens on a resume. If it points at your plan, tasks with a complete line are treated as done and not re-dispatched, and work restarts from the first task with no complete line. A final line that is a fix-round entry means the repair loop is mid-flight and continues from the next round. A ledger pointing at a different plan, or one left in the old .superpowers/sdd/progress.md path, is left alone as another plan's progress and you create a fresh ledger. Where two plans share a file name, the script records the owning plan's path in a plan-path marker per working directory, splitting directories by parent folder name or by a sequential number if those also collide.

Beyond completion entries, the ledger also carries a pre-start plan-collision scan table, per-round fix results, parked findings, and Ruling lines. A Ruling line records three things: what was decided, why, and what is lost if it turns out wrong, and every adjudication must leave a ledger entry — silently discarding a finding is not allowed. This pairs with the skill's design of not checking in with a human between tasks. The documentation's reasoning is that a wrong ruling costs visible rework, while a paused session asking a question costs the other person a full day. Only four situations are allowed to stop the controller: irreversible or destructive operations, security-related operations, side effects outside the worktree that customarily get prior confirmation (merges, pushes to shared branches, publishing), and a plan so broken that any path forward would be guesswork. At the end, the controller gathers every line containing "Ruling:" and lists them chronologically under "Rulings I made," so a developer can redo only the bad decisions.

Separately, the skill's routing logic matters for anyone reading the ledger format: when there is an implementation plan, the tasks are largely independent, and subagents are available, this skill applies; in-line execution or an environment without subagent tools branches to executing-plans. Both share the same working directory and ledger, so the way the ledger is read is common to both.