[JacobStephens2]

Summary

thirdshift architect base <branch> (or --base <branch>) names the Base branch of an Architect run, so the branch checked out in the Launch directory stops mattering.

 architect(args)
-  architect::run(focus)
-    base = checked-out branch, else "HEAD is detached"
+  architect::run(focus, args.base)
+    base = args.base, else checked-out branch, else "HEAD is detached"
     preflight::check_base_branch(base)          # on origin, local copy not ahead
-    launch pull of base
+    launch pull, only if base is the branch checked out
     ReviewWorktree at origin/<base>
-    -> PlanReady(plan)
+    -> PlanReady { plan, base }
-  run::run_to_end(plan, child: None)             # Base branch = checked-out branch
+  run::run_to_end(plan, StartedBy::ArchitectRun { base })
+    Issue branch / Spec branch off origin/<base>, pull request / Spec PR into <base>

run_to_end now takes what started the Run in place of child: Option<&Kind>:

pub enum StartedBy<'a> {
    Command,                      // thirdshift <Issue URL>: the checked-out branch
    ArchitectRun { base: &'a str }, // dispatching its plan: the Architect run's Base branch
    Child(&'a Kind),              // a Ticket's Run or a Base fix: the branch it was given
}

base is parsed for architect only. thirdshift <Issue URL> doesn't take it, and no hidden argument was added: the dispatch is in-process.

Seams tested: the argument parser (args::parse, unit tests) and the thirdshift binary end to end (tests/architect_run.rs, tests/command_surface.rs, with the fake claude and gh).

Evidence

• Before: the review ran at the checked-out branch's head, and the dispatched Run targeted it.

  base_starts_the_review_at_the_named_branchs_origin_head_whatever_is_checked_out
    left:  a11bb52… (origin/main)   right: 82b9c33… (origin/develop)
  base_carries_the_named_branch_to_the_dispatched_spec_run_whatever_is_checked_out
    left: "main"   right: "develop"          # the Spec PR's base
  

  After: cargo test passes: 28 test binaries, 0 failed; cargo clippy --all-targets and cargo fmt --check are clean.

  architect base develop [--plan-only], from main and from a detached HEAD, with an uncommitted change
    review worktree HEAD == origin/develop
    dispatched Run:      issue-8 off develop, PR base develop
    dispatched Spec run: Spec branch off develop, Spec PR base develop, Tickets merged into issue-8
  architect base nowhere            -> "base branch nowhere does not exist on origin; push it first"
  architect base develop (local +1) -> "local develop is 1 commit(s) ahead of origin/develop; push them first"
  launch.pull = true                -> checkout updated only when develop is checked out
  architect base | base --plan-only | base x base y -> exit 2
  

• One full run, on a machine other sessions were loading (load average 7.5 on 4 cores), failed base_carries_the_named_branch_to_the_dispatched_spec_run_whatever_is_checked_out on its exit code. I didn't capture its stderr. It passed in the next full run and in 12 further runs of that test, so I couldn't find a cause.

Unaddressed findings

Standards

• Possible Duplicated Code: the Base branch is resolved in both architect.rs and run.rs (symbolic-ref, given-or-checked-out, check_base_branch, launch pull). Skipped: a Run's resolution goes through Issue branch selection, where a Continuation's open pull request decides, and an Architect run's doesn't, so the shared part is two git calls.
• Outcome::PlanReady carries the Base branch back to main.rs. Skipped: only a ready plan is dispatched, and it is the one outcome that needs the branch, so the variant is where it belongs; returning it beside every outcome would give idea outcomes a branch they don't use.
• ask_base repeats ask_parallel's shape. Skipped: the reviewer judged it acceptable, and it follows the existing pattern.
• Test setup: develop_on_origin(detached: bool) is a flag argument, and the develop setup is repeated in the Spec run and launch pull tests. Skipped: those tests build their scenario another way first (spec_plan, a checked-out develop), so the helper doesn't fit them without a second one.

Spec

• The detached-HEAD message without base gained ", or name it with base ", against "thirdshift architect without base behaves as before". Skipped: the run is rejected as before, and the hint names the way out this Ticket adds. The duplicate test the reviewer pointed at was folded into the existing one.
• base followed by an argument starting with a dash, or a blank one, is an argument error too, where the Ticket only lists "nothing after it, or given twice". Skipped: no branch name starts with a dash, so architect base --plan-only is a typo, which story 18 says must not silently pick a branch.
• Tests the Ticket didn't list (merge into the named branch, base rejected on a Run). Skipped: they pin "It is not added to thirdshift <Issue URL>" and cost little.

Merge Danger

Door: two-way

Reverting removes the flag; nothing is stored or migrated.

Blast Radius: Runs

Every Run now goes through StartedBy, and a Ticket's Run or a Base fix reads the real checked-out branch where it used to read the branch it was given, which only feeds launch pull, off for child Runs. The detached-HEAD message of thirdshift architect and the progress line of a Continuation whose pull request's base replaces a given branch (not the given <branch>) changed wording.

Closes #246

🤖 Generated with Claude Code