[JacobStephens2]

Summary

thirdshift architect [<focus>] --plan-only starts an Architect run from the Launch directory, with no Issue URL.

thirdshift architect "the Spec run" --plan-only
  preflight                      # origin is on GitHub, git identity, HEAD not detached,
                                 # Base branch on origin and not ahead; no Origin match
  launch pull                    # only with launch.pull = true
  worktree <repo>-architect      # detached at origin/<base>, no Issue branch
    Architecture review          # session prompt: prompt::architecture_review(base, focus)
  remove the worktree            # whatever the outcome
  read the final message's last line
    "Architecture review plan: <Issue URL>"
  check the plan                 # in this repository, open, created since the run started,
                                 # not labelled ready-for-human / needs-info / wontfix
  swap needs-triage -> ready-for-agent   # one PUT, other labels kept
  stdout: <plan URL>, exit 0

Any failure after the review starts exits 1 with the cause, then the session log, and changes no label.

 src/
+├── architect.rs     # the Architect run: review, Report (final line), plan checks, label swap
 ├── args.rs          # + Command::Architect, parsed only as the first argument
 ├── issue.rs         # + Repo::of_origin; the Origin match now uses it
 ├── worktree.rs      # + ReviewWorktree: detached, always removed
 ├── session.rs       # + Sessions::run_to_final_message; Logs named for a Run or an Architect run
 ├── prompt.rs        # + architecture_review, and the three final-line prefixes
 ├── github.rs        # + issue (state, labels, createdAt), set_labels
 ├── spec_run.rs      # unready_label now answers for any issue's labels
 └── main.rs          # + architect(), help text
 prompts/
+└── architecture-review.md   # generated
 tests/
+├── architect_run.rs # 20 end-to-end cases
 └── fakes/           # gh: issue create, labels, createdAt; claude: a final message

Three things to know before merging:

• --plan-only is required for now. Nothing dispatches a plan until Architect run dispatches its plan as a Spec run or a Run #232, so bare thirdshift architect is an argument error (exit 2) rather than a run that stops short without saying so.
• The Session prompt names skills that aren't here yet. /thirdshift:improve-codebase-architecture, to-spec, to-tickets and codebase-design arrive with Headless Factory skills for the Architecture review #229. Until then a real Architect run has no skills to load; the end-to-end tests use the fake claude.
• An idea or already-filed final line fails the run for now. The prompt defines all three final lines so Headless Factory skills for the Architecture review #229 and Architect run with no Strong candidate: idea issue or already filed #231 have one format to match, and the cause names the issue reported. Architect run with no Strong candidate: idea issue or already filed #231 turns both into exit 0.

Evidence

• Before: thirdshift architect --plan-only → not a GitHub issue URL: architect, exit 2.
  After: cargo test --test architect_run → 20 passed. The cases:

  plan-only marks the published Ticket ready and prints its URL
  plan-only marks a published Spec ready, leaving its Tickets and other labels
  the Session prompt names the Factory skills, the Base branch and the final line
  a focus goes into the Session prompt
  the review runs detached at origin's Base branch; the Launch directory is never touched
  progress lines cover the review starting, the plan it reported and the label swap
  a failing / interrupted session fails the run and leaves needs-triage
  no final line, or a malformed one, fails the run
  a closed plan / one older than the run / one with another unready label / one in another repository is refused
  detached HEAD, Base branch ahead of origin, no git identity, non-GitHub origin: rejected before any work
  launch pull brings the Launch directory's Base branch up to date first
  

• Full suite: cargo test → 490 passed, 0 failed. cargo fmt --check and cargo clippy --all-targets -- -D warnings are clean.

• Parser unit tests cover architect with and without a focus, --plan-only on either side, stray arguments, and architect anywhere but first.

Unaddressed findings

Standards

• "launch repository" in new src/worktree.rs docs (CONTEXT.md avoids "launch repo"): the file already uses that wording throughout; changing only the new lines would make it inconsistent, and rewording the whole file is outside this Ticket.
• run::pull_base_branch(&launch, Some(&base), &base) passes the same value twice: the two differ for a Run (a Continuation's base, a detached HEAD); one odd call is cheaper than a second function.
• ReviewWorktree's drop repeats Worktree's lock-and-remove shape: Worktree also deletes the Issue branch and can be kept; the shared part is three lines.
• log.exists().then_some(log) mirrors failed_run::fail: one expression; not worth a helper.
• An empty units slice stands for "sent by an Architect run" on the prompts page: one prompt has no unit; a sender type can come when a second does.
• (logs, log) travel with base/focus: mirrors the Run's code; bundling them here alone would diverge from it.
• One-field ArchitectArgs: parallels RunArgs, and Architect run dispatches its plan as a Spec run or a Run #232 adds its flags.

Spec

• --plan-only is mandatory, not optional: without dispatch (Architect run dispatches its plan as a Spec run or a Run #232) a bare architect would have to do the same thing silently; an argument error says so.
• A real run needs Headless Factory skills for the Architecture review #229's skills: out of scope here; called out above.
• Idea and already-filed lines are parsed, then fail the run: the format has to be defined in one place now; making them succeed is Architect run with no Strong candidate: idea issue or already filed #231.
• A fourth plan check, that the plan is in this repository: thirdshift is about to change labels on the issue the agent named, and Architect run: thirdshift finds and implements a Strong deepening opportunity unattended #27 asks that it "never dispatches the wrong issue".
• An empty or dash-led focus is rejected: keeps a mistyped flag from becoming the focus.
• architect merge --plan-only takes merge as the focus: Architect run dispatches its plan as a Spec run or a Run #232 owns merge, no-merge and parallel on architect, including rejecting them with --plan-only.
• Refactors on the Run path (Repo::of_origin, Logs constructors, main.rs helpers): needed so the Architect run shares the Run's code instead of copying it; the existing suite passes unchanged.
• "Created after the run started" compares GitHub's time with the local clock: a clock wrong by more than the length of a review is needed to misjudge it; a clock-free check is a bigger design than this Ticket.
• Nothing requires needs-triage to be present before the swap: the spec's check is "no unready triage label other than needs-triage"; a plan without it still ends ready-for-agent.
• A git worktree add that fails partway can leave <repo>-architect behind: same as a Run's worktree today; not reproduced.
• An interrupt during the label write reports interrupted though the swap may have landed: one request wide; not reproduced.

Merge Danger

Door: two-way

A new command and new code paths; reverting the PR removes them. The fixed-format final line is the one thing #229 and #231 will build on.

Blast Radius: contained

The Run path changes only by refactor: the Origin match goes through Repo::of_origin, log names through Logs::of_run, and spec_run::unready_label takes labels rather than a Ticket. thirdshift help gains a wider usage column and an architect paragraph.

Closes #230

🤖 Generated with Claude Code