[JacobStephens2]

Summary

The Claim, the Sweep, and the Architect plan and idea now name only the labels to take off and the labels to put on. The new Edit in src/labels.rs handles the rest.

 Claim::make / release / remove_if_closed, sweep, mark_plan_ready, label_idea
-  github::issue_labels
-  github::ensure_labels   (only for the concept's own label)
-  Labels::swapped
-  github::set_labels | github::remove_label
-  github::*_label_command
+  Edit::read(issue, off, on) | Edit::of(issue, labels, off, on)
+    .takes_off(label) / .puts_on(label) / .labels()   # progress lines stay where they were
+    .apply()
+  labels::by_hand(issue, off, on)                      # needs no read

Edit::apply
  request(labels, off, on)     # private seam, unit-tested
    nothing changes, ignoring case        -> no request
    one label off, none on                -> DELETE, as the issue spells it
    otherwise                             -> ensure_labels(missing on-labels) then PUT

• Marking a plan ready still runs its checks first, and a failed check still changes no label. The labels it writes now come from a read made after the interrupt check.
• ready-for-agent and needs-triage are now added to a repository that lacks them, with their descriptions.
• One behaviour follows from the DELETE rule. A Claim on an issue that is already in-progress and still carries ready-for-agent now DELETEs ready-for-agent. It used to PUT the labels back. The issue's other labels therefore keep their order: tests/continuation.rs was updated for this.

Evidence

• Before: the Architect plan run onto a repository with only needs-triage and Architect-Plan wrote ready-for-agent with no description, and created no labels.
  After: marking_a_plan_ready_adds_ready_for_agent_with_its_description_to_a_repository_that_lacks_it and labelling_an_idea_adds_needs_triage_with_its_description_to_a_repository_that_lacks_it pass. a_repository_that_has_the_architect_plan_label_keeps_it_as_it_is now expects one gh label create ready-for-agent.
• New unit tests in labels::tests cover no request, a single DELETE as the issue spells the label, a PUT with missing labels added first, case-insensitive off and on, a label in both off and on, and the by-hand commands.
• cargo test, cargo clippy --all-targets and cargo fmt --check all pass.

Merge Danger

Door: two-way

A revert restores the old label handling. The repository labels the new code creates, ready-for-agent and needs-triage, are harmless.

Blast Radius: labels

Every label change thirdshift makes now goes through one path. A bug in Edit would affect Claims, the Sweep and the Architect run together. Progress lines, warnings and by-hand commands are unchanged.

Unaddressed findings

Standards

• Unit tests assert on the private Request enum: the issue specifies this internal seam ("a pure private function, unit-tested in the labels module").
• Edit::labels() widens the interface: the issue asks for it explicitly ("what the issue's labels were").
• Data Clumps (off/on travel together; by_hand could be a method): the issue specifies by_hand as a separate function that needs no read. Bundling off and on into a type would add a type with no rule of its own.
• Duplicated Code in remove_if_closed and sweep: the shared part is three lines, and the two handle failure differently (retry and warn vs. a warning per issue).
• Repeated Switches on the Claim's two flags: each branch prints a progress line that must stay word for word as it is today, so the branches can't be collapsed.

Spec

• No scenario test that a label added during the plan's checks survives: the issue doesn't ask for one, and the fake gh has no hook for changing labels in the middle of a run.

Closes #365

🤖 Generated with Claude Code