[JacobStephens2]

Summary

Each looping end-to-end test in the four slowest binaries now keeps one case end to end. The other cases moved to unit tests at the level where they actually differ.

 tests/spec_run.rs
-  sigterm_… / sighup_… / sigint_…  (3 Spec runs)
+  a_signal_to_the_spec_run_alone_…            (1, SIGTERM)   + interrupt::tests
-  each_unready_label_…            (3 Spec runs)
+  an_unready_label_…              (1, wontfix)               + spec_run::tests
-  parallel_1_… for [parallel, --parallel]    (2)
+  parallel_1_… --parallel         (1)                         + args::tests
 tests/run_notification.rs
-  interrupted: INT/TERM/HUP (3) · failed send: 2 replies × 2 outcomes (4)
-  credentials key: 3 · broken credentials: 3 · no-email spellings: 2
-  RESEND_API_KEY unset/empty: 2 · email.always absent/false: 2
+  one Run each                    + resend_key::tests, args::tests, config::tests
 tests/release.rs
-  agent fails / prints nothing (2 releases) · 7 bad versions · 3 low versions
-  3 kinds of "no bump PR summary"
+  one script run each             + tests that source the script and call
+                                    release_summary / is_plain_version /
+                                    is_higher / bump_summary directly
 tests/merge_run.rs
   budget and cap tests keep their real rounds, now with a comment saying so,
                                    + run::tests (Budgets accounting)

Other code changes:

• Scripts: release.sh and release-notes.sh call main only when run directly, so tests can source them. The version pattern moved into is_plain_version.
• resend_key.rs: the empty-RESEND_API_KEY rule moved into environment_key(value), so it can be tested without touching the process environment.
• Repetition kept: tests where the repetition is the behavior say so in a comment. These are the Repair cap and upstream-move budget tests, the Spec-run rerun, and the Run vs Merge run prompt comparison. So are the release-notes rerun and help/version/update.

Evidence

Full suite: cargo test passes, 24 of 24 binaries, including 76 unit tests (was 65).

How timings were taken: each test executable ran directly with libtest --report-time, base first, then this branch. The base was origin/issue-174 at dbbe8de, which already includes #205 and #204. The machine is shared, 4 cores. Load averages are noted; single-test times vary by about ±1s between runs.

Totals per binary. CPU is user+sys. "Σ tests" is the sum of the per-test times.

Binary	Load	Wall before → after	CPU before → after	Σ tests before → after
spec_run	4.3 / 4.8	103.8s → 99.8s	291.2s → 280.8s	400.6s → 386.2s
merge_run	2.4 / 3.0	31.1s → 30.4s	77.1s → 77.5s	110.4s → 110.3s
release	3.9 / 5.0	30.3s → 29.1s	80.2s → 79.6s	117.8s → 112.5s
run_notification	5.2 / 4.9	15.0s → 8.7s	34.0s → 22.5s	48.4s → 30.7s

What moved and what didn't:

• merge_run had no multi-Run loops; its changes are comments and unit tests only.
• spec_run's wall time is set by its long single-Run scenarios, so fewer Runs barely move its total.

Each test changed

spec_run

Test	Before	After
each_unready_label_keeps_a_ticket_from_running → an_unready_label_keeps_a_ticket_from_running	10.39s	4.72s
sigterm_while_a_ticket_runs_… → a_signal_to_the_spec_run_alone_fails_the_ticket_run_then_ends_the_spec_run_as_interrupted	1.71s	1.70s
sighup_while_a_ticket_runs_… (removed)	2.37s	—
sigint_to_the_spec_run_alone_is_passed_on_to_the_ticket_run (removed)	3.13s	—
parallel_1_runs_the_tickets_one_at_a_time	22.71s	10.80s
repairs_exhausted_on_the_spec_pr_send_it_back_to_draft_and_exit_1 (comment only)	13.99s	12.49s
rerunning_a_failed_spec_run_… (comment only)	11.84s	10.45s

merge_run (comments only)

Test	Before	After
foreign_commits_that_keep_coming_spend_the_upstream_move_budget	5.26s	5.75s
a_review_repair_counts_against_the_repair_cap	6.69s	7.07s
merge_failures_that_keep_moving_the_base_branch_spend_the_base_move_budget	8.89s	7.84s
a_repair_needed_after_a_failed_merge_counts_against_the_repair_cap	7.37s	7.58s
a_merge_run_gives_the_agent_the_same_implement_prompt_as_a_run	4.67s	4.73s

release

Test	Before	After
when_the_agent_fails_or_prints_nothing_… → when_the_agent_fails_the_summary_is_the_generated_notes	7.77s	4.41s
the_summary_is_the_agents_unless_it_fails_or_prints_only_blank_space (new)	—	0.17s
without_a_bump_pr_summary_the_release_body_is_…	3.76s	2.51s
the_summary_is_only_that_of_a_pr_merged_as_the_tagged_commit (new)	—	0.10s
a_version_that_is_not_plain_semver_is_refused	3.04s	2.12s
only_plain_x_y_z_versions_with_no_leading_zeros_pass (new)	—	0.11s
a_version_not_higher_than_the_one_on_main_is_refused	2.33s	2.07s
a_version_is_higher_only_by_its_parts_as_numbers (new)	—	0.09s
the_release_body_is_the_bump_prs_summary_… (comment only)	5.01s	5.08s
when_the_bump_pr_cannot_be_read_… (comment only)	4.50s	4.20s

run_notification

Test	Before	After
an_interrupted_run_sends_one_notification_that_it_was_interrupted	3.48s	1.07s
without_resend_api_key_the_run_stops_before_any_work_and_sends_nothing	0.31s	0.11s
a_failed_send_is_a_warning_that_changes_neither_the_exit_code_nor_stdout	6.45s	1.89s
no_email_skips_the_notification_email_always_asks_for	4.08s	1.88s
without_resend_api_key_a_run_sends_with_the_key_in_the_credentials	5.40s	1.52s
broken_credentials_stop_a_run_that_asks_for_a_notification_before_any_work	0.48s	0.11s
help_version_and_update_never_read_the_credentials (comment only)	0.27s	0.23s
email_always_without_an_address_or_a_key_… → email_always_without_an_address_stops_the_run_before_any_work	0.23s	0.15s
without_email_always_a_run_without_the_flag_sends_nothing → email_always_false_makes_a_run_without_the_flag_send_nothing	3.49s	1.57s

New unit tests:

• interrupt::tests raises SIGINT, SIGTERM and SIGHUP in turn and checks each is recorded.
• run::tests covers the shared Repair cap and upstream-move budget.
• spec_run::tests checks each Unready label.
• resend_key::tests covers an empty RESEND_API_KEY and each way the Credentials can be broken.
• args::tests covers both spellings of no-email and parallel.
• config::tests covers email.to without email.always.

Merge Danger

Door: two-way

Tests only, apart from two small refactors made for testability. The scripts' main guard changes nothing when they are run as before (scripts/release.sh, and release-notes.sh from the workflow).

Blast Radius: tests

If a moved case was mis-covered, a regression in that case would only show in unit tests rather than end to end.

Unaddressed findings

Standards

• Testing past the seam in tests/release.rs (sourcing the scripts, overriding summary_input): skipped. The codebase-design skill allows internal seams used by a module's own tests, and bash offers no other level at which the cases differ.
• Middle Man, from_environment delegating to environment_key: skipped. It is the seam that keeps unit tests from touching the process environment.
• Assert-message wording differs between files ("stderr: {}" vs "{}"): skipped. Each file keeps its existing style.

Spec

• No unit test for a failed Run plus a failed send, only a comment: skipped. main settles the exit code and prints stdout before RunNotification::send, which returns (), so no seam exists without restructuring main. The reply shapes are covered by email::error_text's unit test.
• No unit test for the email.always plus Credentials key case: skipped. email.always only picks NotificationAsk::Send(None), which config::tests covers. The key is then found by the same Resend::new call as with the flag, which stays end to end.
• The no-email and parallel loops were argument-parsing work left to Unit tests for argument parsing; one Run per --email end-to-end test #205: kept. Unit tests for argument parsing; one Run per --email end-to-end test #205 merged without them and added the args seam, so two small unit tests were enough to meet this issue's one-Run-per-behavior criterion.
• The interrupt unit test checks the flag, not the outcome: skipped. The outcome is FailedRun.interrupted = interrupt::requested(), which doesn't depend on the signal. The remaining end-to-end tests show the flag reaching the outcome.
• Extra cases in the new release unit tests (blank space, agent success): kept. They are cheap and pin the boundary of the fallback.

Closes #206

🤖 Generated with Claude Code