[JacobStephens2]

Summary

A Spec run now takes every Ticket it can reach. When any Ticket isn't done, it ends as a Failed spec run.

 land_tickets
   push the Spec branch
-  while a Ticket is ready: run it, fail the Spec run if it fails
-  bail "no Ticket is ready, and #… still open"
+  while a Ticket is ready (open, not Unready, no sub-issues, blockers closed, no outcome yet):
+    run it, record Landed(PR) or Failed(cause, session log)
+    re-read the graph
+  if every Ticket is closed: go on to the Spec PR
+  else: report one line per Ticket, exit 1, no Spec PR, no Spec review

The summary on stderr:

thirdshift: #21 failed: claude exited 1 (session log: …/acme-widgets-issue-21-….jsonl)
thirdshift: #22 blocked by #21
thirdshift: #23 landed with https://github.com/acme/widgets/pull/1
thirdshift: #24 unready: labelled needs-info
thirdshift: #25 blocked by #99 (outside the Spec)
thirdshift: #26 in a cycle: #26 blocked by #27 blocked by #26
thirdshift: Tickets not done: #21, #22, #24, #25, #26, …

• The cause and session log of a failed Ticket come from the last lines of that Ticket's Run. The session log: prefix is now one constant (failed_run::SESSION_LOG), shared by the Run and the Spec run.
• Help and the README (new "Spec runs" section) explain Unready Tickets, and that a Spec run takes every Ticket it can reach.
• CONTEXT.md now says a Ticket with sub-issues is treated as an Unready Ticket.
• help_does_not_mention_the_flag_aliases now matches whole words, because -h appears inside ready-for-human.

Evidence

• Before: on main, a failed Ticket ended the Spec run at once, and an unready Ticket made it bail with no Ticket is ready. Tickets that didn't depend on them never ran.
  After: new end-to-end tests in tests/spec_run.rs, all passing. The full cargo test suite is green.
  • a_needs_info_ticket_and_what_it_blocks_are_not_run_while_an_independent_ticket_lands
  • each_unready_label_keeps_a_ticket_from_running: covers ready-for-human, wontfix and needs-triage; a ready-for-agent Ticket runs.
  • a_failed_ticket_stops_only_its_dependents_and_is_not_started_again: checks that the sessions are exactly ["21","23","24"], and checks the cause and the session log path.
  • an_open_outside_blocker_holds_a_ticket_back_and_a_closed_one_does_not
  • tickets_in_a_cycle_are_not_run_and_the_summary_names_the_cycle
  • a_ticket_with_its_own_sub_issues_is_unready
  • removing_a_needs_info_label_while_another_ticket_runs_lets_it_run
  • help_explains_unready_tickets_and_that_a_spec_run_takes_every_ticket_it_can_reach
  • Every Failed spec run test asserts exit 1, empty stdout, no Spec PR and no Spec review session.

Merge Danger

Door: two-way

Blast Radius: Spec runs only

Ordinary Runs are unchanged, apart from the help text and the shared session log: constant, whose output is byte-identical.

Unaddressed findings

Standards

• report prints lines instead of returning them, so it can't be unit-tested. Skipped: the Spec's testing decision is to test only through the end-to-end harness, which already covers every line.
• run_ticket parses the child's stderr inline. Skipped: the parse is five lines. The shared SESSION_LOG constant removes the hidden coupling, which was the real risk.
• Primitive Obsession: PR URL and log path are Strings. Skipped: both are read back as text only to be printed again.
• Duplicated #n formatting and joins. Skipped: each site joins with a different separator, so a helper would save little.
• report's cascade repeats some of next_ready's checks. Skipped: they answer different questions ("may it start?" and "why is it not done, in priority order"), and sharing one classifier would blur both.

Spec

• Only Tickets that landed in this Spec run are listed, not ones closed earlier. Skipped: "Landed Tickets are listed with their PR" reads as the Tickets this run landed. Their PR comes from the Ticket's Run.
• The cause is taken from the child's last stderr lines, which is fragile. Partly addressed: the prefix is now a shared constant, and the test asserts the cause text. A Run prints its error as one line, so a structured channel isn't needed yet.
• A spawn or read error in run_ticket aborts the Spec run with no summary. Skipped: it means the thirdshift binary itself can't be run, which would stop every Ticket the same way.

Closes #144

🤖 Generated with Claude Code