[JacobStephens2]

Summary

The Spec run's Ticket loop now reaches the outside world only through a private Outside seam. It has two adapters: ChildRunsAndGitHub in production, and a scripted fake in the module's tests. The fake drives a real Ticket board. Behaviour is unchanged.

 spec_run::run
   SpecPr::resume
-  land_tickets(spec, tickets, &worktree, parallel, &base_fix, &mut spec_pr)
-    worktree.push · child_run::start + channel · github::tickets
-    spec_pr.show/landed · interrupt::requested · progress::step
+  ChildRunsAndGitHub::new(spec, &worktree, &base_fix, &mut spec_pr)
+  land_tickets(&mut outside, tickets, parallel)
+    outside.push · start_ticket · next_ending · tickets
+    outside.show · landed · interrupted · step("starting #n")
   review_and_deliver           # unchanged

The #n landed, #n failed and #n interrupted lines stay in the adapter's waiting thread, so their timing on stderr is unchanged.

Evidence

• Before: the loop's rules were checked only end to end, by spawning child Runs against the fake gh and claude.
  After: 18 unit tests at the seam, one per rule in the acceptance criteria, run in about 0.01s together. The 10 end-to-end tests listed in the issue are deleted. The kept ones pass: cargo test --test spec_run gives 65 passed.
• The full cargo test passes. cargo clippy --all-targets is clean.

Merge Danger

Door: two-way

It refactors one module and its tests. A revert restores the old code and the deleted tests.

Blast Radius: Spec runs

A mistake in the production adapter would change how a Spec run starts, waits for or reports its Tickets. The kept end-to-end tests cover the real child Runs, concurrency, the Spec PR's writes and interrupts.

Unaddressed findings

Standards

• The Ticket ending lines are written outside the seam. #n landed, #n failed and #n interrupted still go straight to progress::step in finish_ticket. Skipped: the issue's Decisions keep them in the adapter's waiting thread, so their timing on stderr relative to the relayed lines does not change.
• The reached() and failed() test helpers duplicate others. Similar helpers exist in ticket_board.rs and run_ending.rs. Skipped: each test module keeps its own small builders, as the repo already does.
• "Ticket loop" is not a CONTEXT.md term. Skipped: the issue uses "Ticket loop" throughout and states that no CONTEXT.md change is needed. The term names a private function's job, not a domain concept.

Spec

None. Both findings were addressed: the first-error test now makes a Spec PR opening fail after the first error, and the start-error test checks that the Tickets not done lines are not printed.

Closes #380

🤖 Generated with Claude Code