Verification evidence — issue-228

Executing ../testing-plan.md. One section per activity; raw output in fenced blocks. Environment: this repository's checkout, uv workspace in cli/, Linux, Python 3.11. No credentials involved; all processes loopback-only against temp directories.

T1 — core.lifecycle unit tests

$ uv run pytest tests/test_core_lifecycle.py -q
10 passed

T2 — lifecycle command tests

$ uv run pytest tests/test_lifecycle_cmd.py -q
7 passed

T3 — re-pointed poller tests

$ uv run pytest tests/test_poll_daemon_integration.py tests/test_poll_command.py \
    tests/test_poll_status.py tests/test_poll_heartbeat.py -q
37 passed in 22.79s

One run of this combination failed a single process-table assertion while the full suite was executing concurrently on the same machine (CPU contention stretching a bounded wait); three immediate re-runs and all three full-suite runs passed. Recorded rather than hidden — the timing bound is the daemon tests' known sensitivity, not new to this change.

(test_poll_daemon_integration spawns real the-loop start subprocesses and interrogates /proc; test_poll_command drives daemon_entry poller --once; test_poll_status drives the-loop status text and JSON.)

T4 — restart API + MCP flag integration

$ uv run pytest tests/test_service_lifecycle_integration.py tests/test_mcp_integration.py \
    tests/test_api_routers_integration.py -q
19 passed in 9.15s     # incl. test_restart_schedules_a_detached_fixed_argv_process
                       # and test_mcp_can_be_disabled_per_config

T5 — contract parity

$ uv run pytest tests/test_api_contract_parity.py tests/test_config_schema_parity.py \
    tests/test_docs_parity.py tests/test_configschema.py -q
29 passed in 2.08s     # /api/v1/restart present in app + authored contract

T6 — schema copies, docs parity, config schema

(combined into the T5 run above: schema copies byte-identical, docs P1–P5 green over
the new/removed pages and the four new keys)

T7 — full suite

$ cd cli && uv run pytest -q
2034 passed, 1 skipped in 84.45s (final make check run; three full-suite runs total, all green)

The skip is the same pre-existing one main carries. The first run on this branch — code changed, tests not yet re-pointed — was 23 failed / 1991 passed / 1 skipped, all 23 in the expected categories (poll-command suites, docs parity, two config tests); the count then grew by the new lifecycle tests and shrank by the fork-specific scenarios that died with daemonize().

T8 — lint / format / types

$ make lint format-check typecheck
uv run ruff check cli hooks         → All checks passed!
npx markdownlint-cli2 "**/*.md"     → 658 file(s), 0 error(s)
uv run ruff format --check cli hooks → 212 files already formatted
uv run pyright cli                  → 0 errors, 0 warnings, 0 informations

T9 — markdown lint

Included in make lint above (markdownlint-cli2 over 658 files, 0 errors). The three errors it caught mid-work — unescaped | inside code spans in the new capability history rows — were fixed, which is self-review round 1's second finding.

T10 — config validation

$ uv run python scripts/validate_config.py
VALID   .the-loop/harness-config.yaml
VALID   skills/the-loop/templates/harness-config.yaml
VALID   cli/the_loop/harness-config.default.yaml
VALID   .the-loop/collaborators.yaml
VALID   skills/the-loop/templates/collaborators.yaml
VALID   .the-loop/cli-config.yaml
VALID   skills/the-loop/templates/cli-config.yaml

The template now carries the four new keys and validates against the updated schema; the repo's own cli-config.yaml (no enabled keys) validates unchanged — R5.3.

T11 — manual smoke (temp HOME, real service)

Script: fresh $HOME with a two-line cli-config.yaml (a free port, nothing else), run from an empty working directory. Transcript, verbatim:

$ the-loop start
service     started         [enabled]  started at http://127.0.0.1:34493; /mcp exposed
gh-webhook  disabled        [disabled]  webhooks.ghWebhook.enabled is false
poller      disabled        [disabled]  polling.enabled is false
exit=0

$ the-loop status
service     running (pid 7060) [enabled] — http://127.0.0.1:34493, healthy
gh-webhook  not running [disabled]
poller      not running [disabled]
            last cycle: unknown — no heartbeat recorded
exit=0

$ the-loop restart
stopping:
poller      not-running     [disabled]  poller is not running
gh-webhook  not-running     [disabled]  gh-webhook is not running
service     stopped         [enabled]  stopped (pid 7060)
starting:
service     started         [enabled]  started at http://127.0.0.1:34493; /mcp exposed
gh-webhook  disabled        [disabled]  webhooks.ghWebhook.enabled is false
poller      disabled        [disabled]  polling.enabled is false
exit=0

$ the-loop stop
poller      not-running     [disabled]  poller is not running
gh-webhook  not-running     [disabled]  gh-webhook is not running
service     stopped         [enabled]  stopped (pid 7070)
exit=0

$ the-loop status (after stop)
service     not running [enabled]
gh-webhook  not running [disabled]
poller      not running [disabled]
            last cycle: unknown — no heartbeat recorded
exit=1

Every claim on display: defaults compose service-only (R5.3), disabled rows name their keys (R1.2), restart is stop-then-start (R4.1), stop is idempotent about the not-running daemons (R3.1), and the final status exits 1 because an enabled service is down (R3.3). --with-upgrade was exercised at the plan level only (unit test with the planner faked, plus the real planner's own issue-152 suite) — executing a real upgrade would mutate this environment (testing-plan §Verification environment).

T12 — reference sweep

$ grep -rn "the-loop poll" README.md cli/README.md docs skills cli/the_loop \
    --include="*.md" --include="*.py" | grep -v docs/specs | grep -v docs/decisions
cli/the_loop/poller/daemon.py:3   # "``the-loop poll start`` used to be…" — the module
                                  # docstring describing the removal itself

Surviving matches are intended: the new daemon module's own docstring (which must name the old command to explain the move), the historical record under docs/specs/** and docs/decisions/**, and the rewritten test files whose docstrings describe what was re-pointed. Four code docstrings still naming the-loop poll status were caught by this sweep and fixed (self-review round 3).

Docs site build

$ cd docs && npm run docs:build
✓ building client + server bundles…
✓ rendering pages…
build complete in 41.95s

The new command pages, the edited option pages and the sidebar entries all render.

T13 — the PR #229 review round (the fold + the dashboard restart)

Second commit on the branch, after the owner's review instructions.

$ cd cli && uv run pytest -q          # gh-webhook/service commands removed,
2034 passed, 1 skipped in 76.80s      # receiver run loop relocated to webhook/daemon.py

$ cd ui && npx vitest run             # restart client (live+demo), Settings Service
Tests  91 passed (91)                 # card, ConfigEditor "Restart now" (2 new tests)
$ bun run lint                        # oxlint --type-aware: clean
$ npx tsc --noEmit                    # 0 errors
$ bun run build                       # vite build ok

The re-pointed scenarios of this round: test_cli.py (surface = the four lifecycle commands, nothing else), test_cli_config.py (both daemon modules resolve the config path per call; --config override honoured with no cached module path), test_routing.py (receiver strict/lenient config reads and hot reload via webhook.daemon), test_webhook_routing_integration.py (routing opt-in follows routing.enabled), test_service_lifecycle_integration.py (the-loop start/stop drive the real service), test_client.py (hints name the-loop start).

T14 — review round 2: single-process mode (issue-231)

Third commit on the branch, after the owner blocked the merge on issue-231 ("merging this PR will cause a regression").

$ cd cli && uv run pytest tests/test_core_lifecycle.py -q
15 passed                              # 5 new hosted-mode cases; 2 standalone-path
                                       # tests pinned with service.hostIngresses: false

$ uv run pytest tests/test_hosted_ingress_integration.py -q
2 passed in 7.15s                      # real processes: one pid holds all three locks,
                                       # hosted receiver answers /health on its own port,
                                       # status rows say hosted, stop releases everything;
                                       # a standalone daemon's lock is skipped, not stolen

$ uv run pytest -q
2041 passed, 1 skipped in 93.02s       # full suite after the hosting change

The first full-suite run of this round failed exactly one test — test_docs_parity.py::test_p4_every_schema_leaf_is_documented, because service.hostIngresses was in the schema before its section existed on docs/config/cli/service-options.md. Documented, re-ran: 5/5 parity tests green, then the clean full run above.

CI fix on head ec5fb76 — a transiently-held lock read as a started poller

CI (and only CI) failed test_poll_daemon_integration.py::test_start_reports_a_startup_failure_to_its_caller: start reported an unknown-provider poller as started (exit 0). The race: poller.daemon.run() took the single-instance lock before validating its sources, so a poller doomed by its own config held the lock for the few hundred milliseconds its validation took to fail — and start's wait-for-lock proof, sampling in that window, honestly saw a held lock. Locally the daemon always died before the first sample; CI's slower machine caught the window.

Fix: provider validation (a pure parse + construct, no side effects) moved before lock acquisition in run(), so a config-doomed poller exits without ever holding the lock — which is what the test's own Given ("exits before taking its lock") always described. The lock-first rule keeps fencing what it was built to fence: side effects (ledger, web terminal, dependency probes). The hosted path (_run_locked) is unaffected. Verified: the failing test 5× green in a row, the poller/lifecycle/hosted suites (32 passed), and the full suite re-run below.