Merge topologies

fab operator autopilot runs a queue of changes through the fab pipeline and opens a PR for each. How those PRs relate to main — and to each other — is controlled by the queue’s merge mode (named at queue start via --mode or natural language, else resolved from the autopilot.merge_mode config preference). There are three:

Mode	Shorthand	PRs merged	Same-repo chain shape
cherry-pick-ladder (default)	▂▄▆	held until “merge all”	each PR based on main, carrying cherry-picked copies of earlier PRs
merge-auto	░▒▓█	merged as each completes	none — every change rebases onto the latest main independently
stacked-prs	▄▀	held until “merge all”	each PR based on the previous PR’s branch (true stack, no cherry-pick commit)

cherry-pick-ladder

                    ┌───┐
            ┌───┐   │ C │
    ┌───┐   │ B │   ├╌╌╌┤
    │ A │   ├╌╌╌┤   │ b'│
    │   │   │ a'│   │ a'│
────┴───┴───┴───┴───┴───┴──▶ main
     PR1     PR2     PR3

Every PR stands on main; each successive diff is taller because it carries cherry-picked copies of its predecessors (a', b') below the dotted line. All PRs are held open together and merged base-first once you ask for “merge all”.

merge-auto

    ┌───┐          ┌───┐          ┌───┐
    │ A │          │ B │          │ C │
────┴─▼─┴●─────────┴─▼─┴●─────────┴─▼─┴●──▶ main
       merged         merged         merged

Nothing coexists and nothing is held: the operator merges each PR into main the moment it lands (▼ into ●), main advances, and the next change starts from the advanced line. No batch review, no re-stacking.

stacked-prs

                    ┌───┐
            ┌───┐   │ C │  PR3 · base: B
    ┌───┐   │ B │   └───┘
    │ A │   └───┘  PR2 · base: A
────┴───┴──────────────────▶ main
    PR1 · base: main

Uniform height: every diff shows only its own delta, because each dependent branch is created off its dependency’s branch rather than off main (no cherry-pick commit). The diagonal is load-bearing — merging a bottom box means re-seating the ones above it, so a “merge all” retargets and rebases each PR up the chain after the one below it lands.

Choosing a mode

• cherry-pick-ladder — the safe default. Every PR is independently reviewable against main, at the cost of restated diffs further up the chain.
• merge-auto — for a queue of genuinely independent changes you trust to land one at a time without a batch review step.
• stacked-prs — for a queue with real same-repo dependencies where you want each PR’s diff to show only its own change, at the cost of the merge-all choreography (base retarget + rebase per PR) when the stack comes down.

The default (cherry-pick-ladder) is settable as a standing preference — name a mode once, machine-wide, and every queue started without naming a mode adopts it: fab config set --system autopilot.merge_mode <name>. An explicit --mode flag or instruction at queue start always outranks the configured value.

All three share the same underlying pipeline (intake → apply → review → hydrate → ship → review-pr) and queue-ordering rules — mode only changes when PRs merge and how same-repo dependents are branched. See the fab-operator skill’s Coordination Patterns section for the full autopilot mechanics (dependency resolution, ordered merge, CI-failure handling).