Evidence — test runs

Work item: issue #172 · Testing plan rows T1, T2, T2b, T8, T10 · captured 2026-08-07, after the owner-review rebuild to the pullRequests[] endpoint model. (The first capture, for the link-record version this PR initially carried, is in this file's git history.)

Nothing here is redacted, and nothing needed to be: the outputs are pytest summaries and test names. No network call, no gh invocation, no secret store is involved.

Full suite

$ uv run --directory cli pytest -q
1423 passed, 1 skipped in 51.96s

The one skip is pre-existing and unrelated; the baseline before this work item was 1379 passed, 1 skipped.

T1 — the store and the resolver

$ uv run --directory cli pytest tests/test_routing.py -q
106 passed in 1.09s

New in this file:

test_link_pull_request_records_the_pr_on_the_work_items_record
test_link_pull_request_is_idempotent
test_link_pull_request_refuses_the_work_item_itself
test_record_owning_resolves_a_pr_to_its_work_items_record
test_session_for_prefers_the_prs_own_endpoint
test_a_closed_endpoint_falls_back_to_the_work_items_session
test_close_endpoint_leaves_the_record_live
test_touch_records_deliveries_per_endpoint
test_an_unreadable_pull_request_entry_does_not_take_the_record_down
test_a_nested_pull_request_tree_is_flattened_on_read
test_endpoints_survive_closing_and_reopening_the_record
test_pr_work_item_names_the_ref_extraction_emits_last
test_pr_work_item_is_none_for_events_that_concern_no_pull_request
test_pr_work_item_carries_the_host_off_the_payload
test_delivery_status_resolves_a_prs_endpoint
test_dispatcher_still_routes_pr_events_that_are_not_close
test_dispatcher_delivers_pr_events_into_the_work_items_session_when_collapsed

Three pin properties rather than behaviours:

• test_link_pull_request_records_the_pr_on_the_work_items_record asserts the registry directory holds exactly one file — the owner's "single session file for one work-item" — and reads it back through a fresh SessionRegistry (the restart property as a filesystem fact).
• test_pr_work_item_names_the_ref_extraction_emits_last asserts the ref a PR is recorded under equals extract_work_items(...)[-1] on both PR-shaped events, so recording and routing cannot drift.
• test_touch_records_deliveries_per_endpoint pins that dedup does not leak between a work item's conversations — the property that makes several sessions per work item safe against redelivery.

T2 — the ticket's reproduction, as an integration test

$ uv run --directory cli pytest tests/test_webhook_routing_integration.py -q
22 passed in 11.52s

The issue-172 scenarios, each with a Gherkin docstring and a Requirement: link:

test_pr_comment_reaches_the_linked_issues_work_item
test_pr_event_still_reaches_its_work_item_after_the_link_is_removed
test_a_recorded_pr_does_not_suppress_a_work_item_the_linkage_still_finds
test_spawning_for_a_linked_issue_records_the_binding
test_a_stop_on_an_unlinked_pr_stops_the_bound_session
test_a_pr_close_matched_through_a_binding_leaves_the_session_open
test_a_pr_with_its_own_session_is_still_auto_closed

The central one drives two signed POSTs: the first (with Closes #15) records PR 16 on issue 15's record and spawns the PR's own session; the second (linkage edited out) is delivered into that session, with no duplicate spawn and no record ever minted for the PR. The close scenario now also asserts the PR's endpoint is closed while the work item's session stays live — issue-101's rule expressed in the model.

Two fixture details recorded because they are real properties of the code: test_a_stop_on_an_unlinked_pr_stops_the_bound_session must pass authorized_users to the dispatcher's config as well as the router's (issue-106's deliberate stricter re-check), and the R2.3 both-records scenario runs in collapsed mode (sessionPerPr: false) so its both-deliver assertion stays sharp — under per-PR sessions the two records' endpoints would contend for the PR's one loop-<slug> tmux name (decision-064 § Known edge).

T12 — the two loops

$ uv run --directory cli pytest tests/test_graph_loops.py -q
20 passed in 0.86s

All new, pinning the seam from both sides (decision-065):

test_both_loops_ship_compiled_and_named
test_the_outer_implementation_node_awaits_the_inner_loops
test_no_inner_loops_passes_vacuously
test_an_unfinished_inner_loop_holds_the_outer_gate
test_all_inner_loops_complete_releases_the_gate
test_a_corrupt_inner_state_holds_the_gate_rather_than_passing_it
test_bootstrap_selects_the_loop_and_the_state_location
test_starting_an_inner_loop_leaves_the_outer_pointer_untouched
test_two_inner_loops_keep_separate_pointers
test_on_pr_spawn_enters_the_inner_loop_only
test_on_pr_event_advances_the_inner_loop_not_the_outer
test_a_merged_pr_completes_its_inner_loop_audited_as_forced
test_an_unmerged_close_leaves_the_inner_loop_where_it_was
test_an_inner_loop_prompt_claims_with_pr
test_deliver_assignment_skips_without_a_channel
test_deliver_assignment_renders_the_nodes_work_and_the_claim_command
test_deliver_assignment_addresses_the_inner_loop
test_a_failing_channel_never_gates_the_node
test_the_graph_assigns_on_entering_the_inner_loop
test_the_outer_loop_assigns_without_a_pr

Three pin invariants rather than behaviours: the outer pointer is untouched by an inner loop's start and events (the one-way flow, decision-065 D5); a merged PR's completion is recorded as forced with the reason, so check --recompute still shows which inner gates never ran (D6); and an unreadable inner state holds the outer gate naming the PR — a silent pass over a damaged record is the issue-124 shape. The P5 parity assertions (test_graph_parity.py) now iterate both shipped graphs, so an inner-loop node gating a section with nothing to read fails the build exactly as an outer one does.

T2b — the poll path

$ uv run --directory cli pytest tests/test_poller.py -q
107 passed in 0.24s

The two poll-path regressions found in self-review of the first draft remain covered in the rebuilt model:

Case	Consequence before the fix	Test
retry accounting asks about the PR's ref, but the delivery id lives on the PR's endpoint	a successful delivery reports unhandled; the poller re-forwards the same comment until maxRetries is spent	test_delivery_status_resolves_a_prs_endpoint
first-sight detection asks about the PR's ref	a running PR is treated as first sight: its whole thread is baselined as read and a spawn armed against it	test_a_recorded_pr_suppresses_presence_when_the_linkage_is_gone

T2/T2b (negative) — the regression tests against the un-fixed code

A pytest plugin rebinds SessionRegistry.record_owning to the bare find_by_work_item and _record_pr_binding to a no-op — the pre-issue-172 behaviour, everything else identical:

$ PYTHONPATH=… uv run pytest tests/test_webhook_routing_integration.py tests/test_poller.py \
    -q -p prefix_plugin -k "…issue-172 scenarios…"
FAILED …::test_pr_comment_reaches_the_linked_issues_work_item
FAILED …::test_pr_event_still_reaches_its_work_item_after_the_link_is_removed
FAILED …::test_a_recorded_pr_does_not_suppress_a_work_item_the_linkage_still_finds
FAILED …::test_spawning_for_a_linked_issue_records_the_binding
FAILED …::test_a_stop_on_an_unlinked_pr_stops_the_bound_session
FAILED …::test_a_pr_close_matched_through_a_binding_leaves_the_session_open
FAILED …::test_a_recorded_pr_suppresses_presence_when_the_linkage_is_gone
7 failed, 122 deselected in 28.22s

All seven fail, as they must — the checks check something.

T8 — the abuse cases

Run as part of T1, one per boundary in design.md § Security design:

Case	Test
a hand-edited entry (../../etc/passwd as a ref) is skipped per entry; the work item's own session survives	test_an_unreadable_pull_request_entry_does_not_take_the_record_down
a nested pullRequests tree is flattened on read — no structure to walk	test_a_nested_pull_request_tree_is_flattened_on_read
a work item cannot be recorded as its own PR	test_link_pull_request_refuses_the_work_item_itself
a closed/unspawnable endpoint falls back to the record — an event never reaches nowhere	test_a_closed_endpoint_falls_back_to_the_work_items_session

T10 — reset, migration, and the state classification

$ uv run --directory cli pytest tests/test_reset.py tests/test_state_portability.py -q
32 passed in 0.13s

The rebuild made this row smaller: the PR entries live inside local/<slug>.json, so there is no new generated path to classify, no new .gitignore line, and no new reset piece — deleting the record takes its PRs with it. test_state_portability.py passes unmodified, which is the assertion that nothing about the on-disk taxonomy changed.

Migration is the absent-key case: every record written before issue-172 has no pullRequests key, parses to an empty list, and round-trips byte-identically (to_dict omits the key when the list is empty). Every pre-existing test runs against such records and passes unchanged.