[JacobStephens2]

Summary

Without --plan-only, an Architect run now goes on from the plan it marked ready to the Spec run or Run that thirdshift <plan URL> would start, and ends as that run does.

 thirdshift architect [<focus>] [merge | no-merge] [parallel <n>] [--plan-only]
   architect::run                  # Architecture review, checks, needs-triage -> ready-for-agent
-  print the plan's URL            # --plan-only was required
+  --plan-only?
+    print the plan's URL, exit 0  # merge, no-merge, parallel: argument error (exit 2)
+  else
+    "dispatching the plan <url>, as thirdshift <url> would"
+    run::run(plan, goal, parallel)  # the entry point thirdshift <Issue URL> uses
+      plan has sub-issues -> Spec run
+      otherwise           -> Run   # parallel <n> fails here, as for any non-Spec
+    exit code, PR URL on stdout, last stderr line: the dispatched run's

• goal is merge/no-merge, else the User config's merge.always; parallel is parallel <n>, else spec.parallel (Parallel::new, now shared with the Run path).
• The dispatched run sends no Run notification yet, whatever email.always says: that is One Run notification per Architect run #233. The README and help say so.
• merge, no-merge and parallel are flags after architect, with or without dashes, so none of them can be the focus.

Seams tested: the thirdshift binary end to end through tests/support::Scenario with the fake claude and gh (tests/architect_run.rs, tests/command_surface.rs), and the argument parser, args::parse. The fake gh gains gh fake sub-issues <n> '<list>', so a fake Architecture review can publish a Spec.

Evidence

• Before: thirdshift architect → exit 2, architect needs --plan-only: it can't yet implement the plan it publishes.
  After: new end-to-end tests in tests/architect_run.rs, all passing:

a_single_ticket_plan_starts_a_run_on_it_and_ends_as_that_run_does
a_plan_with_tickets_starts_a_spec_run_on_it_and_ends_as_that_spec_run_does
merge_merges_the_runs_pull_request_or_the_spec_pr
the_user_configs_merge_default_applies_unless_no_merge_is_given      # both plan shapes
parallel_passes_through_to_the_spec_run                              # #10 needs #9 landed first
parallel_on_a_single_ticket_plan_fails_as_it_does_for_an_issue_that_is_not_a_spec
a_dispatched_run_that_fails_fails_the_architect_run_as_a_failed_run_does
progress_lines_show_the_dispatch_after_the_label_swap_and_before_the_run

• Dropping parallel from the dispatch makes both parallel tests fail (checked by hand, then restored).
• Parser unit tests cover the flags in both spellings, contradictory, repeated and malformed ones, and each flag with --plan-only; tests/command_surface.rs covers the same rejections on the binary (exit 2, help on stderr, no review started).
• Full suite: cargo test passes, every test binary (87 unit tests, 28 in architect_run, 61 in spec_run, and the rest); cargo clippy --all-targets and cargo fmt --check are clean.

Unaddressed findings

Reviewed with origin/issue-27 as the fixed point: the local issue-27 branch on this machine is behind it.

Standards

• "Dispatch" is not in the CONTEXT.md glossary. The Architect run entry already describes the step, and the word comes from Architect run: thirdshift finds and implements a Strong deepening opportunity unattended #27 and Architect run dispatches its plan as a Spec run or a Run #232; a glossary term for a verb used in one place seemed more than it needs.
• The Run notification glossary entry doesn't list the dispatched run among those that send none. That is the interim state until One Run notification per Architect run #233, which defines the one notification per Architect run and will rewrite the entry.
• Data Clumps: DispatchArgs { goal, parallel } repeats two fields of RunArgs. Embedding it would change RunArgs and the Run's parsing, which Architect run: thirdshift finds and implements a Strong deepening opportunity unattended #27 says stays unchanged; the separate type is what makes "flags with --plan-only" unrepresentable.
• Duplicated Code: the merge / no-merge / parallel match arms are in both parse and parse_architect. Three one-line arms over the shared ask_once and ask_parallel; a shared flag parser for two loops that otherwise differ would cost more than it saves.
• Duplicated Code: pr_from and agent_opens_pr in tests/architect_run.rs resemble helpers in tests/spec_run.rs. Test files here keep their own small helpers and agent scripts; moving them into tests/support would mean rewriting spec_run.rs call sites in this PR.
• Divergent Change: the Architect run is split between src/architect.rs (up to the plan) and architect() in src/main.rs (the dispatch). Deliberate: main.rs is where a Run's flags meet the User config's defaults, so dispatching from there is what keeps the plan's run identical to thirdshift <plan URL>.

Spec

• The README's note that the dispatched run sends no Run notification covers ground that belongs to One Run notification per Architect run #233. Kept: it is what the code does today, and without it "exactly as thirdshift <plan URL> would" would be wrong for anyone with email.always set.

Merge Danger

Door: two-way

Revert the two commits and architect needs --plan-only again.

Blast Radius: architect

thirdshift architect without --plan-only used to be an argument error and now implements the plan unattended, opening a pull request, or merging it with merge or merge.always. A bare merge, no-merge or parallel after architect is now a flag, not a focus. The Run path changes only in how it builds Parallel.

Closes #232

🤖 Generated with Claude Code