0004 — Eager background delegation via fold-to-main

2026-07-10. Reverses the "on a branch → run inline" delegation default (/ardd-implement, /ardd-converge). Extended by 0005 — Background by default: solo /ardd-plan stopped authoring the branch the fold existed to undo, so the fold is now a recovery path, and the delegation/merge_policy knobs can remove the gate's prompts entirely.

The old default and why it was wrong

The solo-mode delegation gate used to skip the "delegate to a background subagent?" offer whenever on_default was false — i.e. whenever the run was already on a feature branch or worktree. The stated reason: "the run is already isolated on a branch, so all its state already rides that branch."

That conflated two different things. A branch handles state isolation; backgrounding is about execution locus — freeing the interactive session while long-running work proceeds elsewhere. Being on a branch is no reason to run in the foreground. In practice the gate defaulted to inline for the entire common case (every /ardd-plan→/ardd-tasks→/ardd-implement chain runs on the feature branch /ardd-plan created), so backgrounding was almost never offered. (Feedback feedback-eager-backgrounding-return-to-5cde.)

The obstacle that made inline look necessary

Agent isolation: "worktree" branches the subagent's worktree from origin/<default> (fresh base), and worktree-align.sh fast-forwards local <default> into it. There is no way to branch a delegated worktree from the current feature branch. So a delegated run can only see state that lives on local <default>. When you're on a feature branch, the tasks file and its progress live on that branch — a delegated worktree branched from <default> wouldn't see them. Inline was the path of least resistance, not a principled choice.

The decision

Offer backgrounding eagerly, regardless of on_default. When the user backgrounds while on a feature branch, first fold that branch into local <default> and return the focused session there — a new deterministic scripts/fold-to-main.sh does the fast-forward fold + checkout, refusing (folded=false reason=dirty|detached|diverged|checkout-failed) rather than resolving, exactly like worktree-align.sh. Then delegate: the subagent's worktree branches from <default>, align carries the just-folded state in, and the interactive session is left clean on <default>.

This ties the two feedback asks into one operation: return-to-main is the mechanism that makes eager backgrounding possible on a branch. fold-to-main and worktree-align are counterparts — one pushes branch state onto local <default>, the other pulls local <default> into the worktree.

Invariants preserved

• "No state-commit before the branch." A fast-forward authors no new commit — folding only advances the <default> ref to commits that already exist on the branch. The "nothing is committed in the delegation-gate step" note still holds.
• "Default branch = merged truth." At the gate the tasks file is still ready (the ready→in-progress flip happens later, in the worktree), so the fold carries only planned truth onto <default>; in-flight truth rides the subagent's worktree as before. /ardd-converge is the one case that may fold an already-in-progress tasks file, briefly placing in-progress state on <default> until the subagent's branch merges — an accepted, bounded exception.
• Refuse, never resolve. Same discipline as worktree-align.sh: a dirty/detached/diverged tree stops the run with a reason, never a hand-reconciled merge.

Scope: solo mode only. Collaborative mode already always moves work to a branch and merges via PR, and never eager-merges locally, so it is unchanged.