diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 03a5215..2183182 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -4,11 +4,10 @@ on:
   push:
     tags:
       - "v*.*.*"
+  workflow_dispatch:
 
 env:
   PYTHON_VERSION: "3.13"
-  PYPI_PROJECT_URL: https://pypi.org/p/teslemetry-stream
-  ENVIRONMENT_NAME: pypi
 
 jobs:
   lint:
@@ -84,8 +83,8 @@ jobs:
     needs: build
     runs-on: ubuntu-latest
     environment:
-      name: ${{ env.ENVIRONMENT_NAME }}
-      url: ${{ env.PYPI_PROJECT_URL }}
+      name: pypi
+      url: https://pypi.org/p/teslemetry-stream
     permissions:
       id-token: write
     steps:
diff --git a/AGENTS.md b/AGENTS.md
index 03c3b7f..3e8600b 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -14,7 +14,7 @@ This file is the project's committed home for project-intrinsic agent knowledge:
 - Energy site events (`teslemetry_stream/energysite.py`) are shaped differently from vehicle signals: `live_status`/`site_info` are flat top-level envelopes (`{createdAt, site_id, isCache?, live_status|site_info}`), not nested under `data`, and the payload is a full opaque document rather than a field delta - there is no per-field config to enable, the server auto-polls subscribed sites. `tests/test_energysite_events.py` fixtures mirror the server schemas.
 - `energy_totals` is shaped differently again: the site id rides the `id` field, not `site_id` - filter on `id` and `totals`, not `site_id`. It carries a compact cumulative `totals` object (`EnergyHistoryTotals` in `const.py`) instead of a document, fires only when the server's periodic `calendar_history` poll detects a change (silence is not staleness), and has no snapshot-on-connect delivery. The wire payload is `id`/`createdAt`/`totals` plus `isCache` only when true; `Key.PRODUCT_TYPE`/`Key.TOPIC`/`Key.URL` in `const.py` remain defined for other event kinds but are not part of the energy_totals filter.
 - `site_info` events do not carry `tariff_content`/`tariff_content_v2`; the V2 tariff is its own `tariff_content_v2` event/listener (`listen_TariffContentV2`), same envelope shape as `site_info`, with a `None` body meaning an explicit server-side removal rather than "not received yet". Both share the same silence-means-no-change contract - freshness lives in REST, never in event cadence. There is deliberately no library helper recombining `site_info` and `tariff_content_v2` into one document - that would only ever cover the V2 tariff (legacy V1 `tariff_content` has no SSE topic and stays REST-only by design); a consumer wanting both tariffs together should use the REST site_info endpoint.
-- Releases (tag `v*.*.*`) go through `.github/workflows/release.yml` directly - it's the sole top-level workflow, triggered on the tag push: `lint` + the full `test` python-version matrix (mirrors `ci.yml`) must pass on the exact release SHA before `build` (single Python, build+twine) runs, and only then do the `pypi` environment's protection rules (and its trusted-publishing OIDC) allow `publish-to-pypi`. It must stay a top-level workflow, not a `workflow_call` reusable one - PyPI's trusted publisher is configured for the `release.yml` + `pypi` environment identity, and a reusable-workflow caller signs PEP 740 attestations under the caller's identity instead, which that publisher check rejects. The `pypi` GitHub environment itself (required reviewers, deployment branches) is admin-configured outside this repo's files.
+- Releases (tag `v*.*.*`) go through `.github/workflows/release.yml` directly - it's the sole top-level workflow, triggered on the tag push: `lint` + the full `test` python-version matrix (mirrors `ci.yml`) must pass on the exact release SHA before `build` (single Python, build+twine) runs, and only then do the `pypi` environment's protection rules (and its trusted-publishing OIDC) allow `publish-to-pypi`. It must stay a top-level workflow, not a `workflow_call` reusable one - PyPI's trusted publisher is configured for the `release.yml` + `pypi` environment identity, and a reusable-workflow caller signs PEP 740 attestations under the caller's identity instead, which that publisher check rejects. The `pypi` GitHub environment itself (required reviewers, deployment branches) is admin-configured outside this repo's files. `jobs.<id>.environment.name`/`.url` cannot reference the `env` context (only `github`, `inputs`, `vars`, `needs`, `secrets`, `strategy`, `matrix` resolve there) - referencing `env.*` there is a workflow-file parse error that fails the whole file at startup, before trigger filtering, so it fails every push (not just tags) with zero jobs and no logs. Use literal values or `vars.*` instead. `release.yml` also carries `workflow_dispatch` so a tag whose run failed before publish can be re-run manually without tag surgery.
 - `TeslemetryStream(topics=...)` is an optional exact SSE wire-event allowlist sent as the connection's `topics` query param; `SseTopic` in `const.py` is the closed set the server recognizes, and `SSE_VEHICLE_TOPICS`/`SSE_ENERGY_TOPICS`/`SSE_ALL_TOPICS` are client-side presets - flat per-product-kind lists of exact wire names, deliberately not further split by whether a topic happens to have a connect-time snapshot server-side; that's upstream server behavior, not something this library encodes. Omitting `topics` (`None`) is legacy-all forever - every applicable event delivered unfiltered. An explicitly empty iterable is rejected with `ValueError` at construction time rather than silently falling back to legacy-all - "no topics" must not mean "all topics", mirroring the server's own 400 on an empty `topics` value. A bare `str`/`SseTopic` is accepted as a single topic rather than iterated character-by-character - `topics` type-checks `str | Iterable[str] | None` precisely because a lone string also satisfies `Iterable[str]`, the classic footgun. `tests/test_sse_topics.py` covers the tariff listener, its null-removal signal, the `topics` param's URL construction, the empty-iterable rejection, and the bare-string/bare-`SseTopic` case.
 - `TeslemetryStream` owns exactly one `_listen_task`: `async_add_listener`'s zero-to-one transition only creates it when absent/done, and `listen()` itself checks `asyncio.current_task()` against `self._listen_task` - a second concurrent `listen()` call joins the owner via `await existing_task` instead of racing it for the connection. `connect()` serializes the actual GET behind `self._connect_lock` and re-checks `self.active` after acquiring both the lock and the response, discarding a response that arrived after a stop/supersede rather than publishing it. Internal reconnect paths (EOF, `ClientError`, unexpected exceptions in `__anext__`) call `_close_response()`, which only clears the response and notifies connection listeners - they must not call `close()`, which additionally flips `active=False` and cancels the owned task, i.e. a real stop. `listen()`'s `finally` calls `_close_response()` unconditionally, so task cancellation (however triggered) still releases the connection. `tests/test_stream_lifecycle.py` covers the add/remove/re-add race, duplicate `listen()` calls, cancellation while blocked reading content, close-during-connect, and close-during-backoff.