`langsmith==0.7.34` (langchain-ai/langsmith-sdk#2778, released
2026-04-23) is now on PyPI, so the lockfile can be regenerated cleanly
against the pin bumped in 4a25d8a. Un-blocks the `Verify uv.lock
files` CI check and the evals test jobs.

- `uv lock --upgrade-package langsmith` in libs/evals
- `uv run --group test ruff check` clean on changed files
- `uv run --group test ty check` clean (no "unused ignore" warnings)
- `uv run --group test pytest tests/unit_tests/` - 253 passed

Made-with: Cursor

…2888)

Migrates the LangSmith harbor env (and its integration-test fixture)
from the deprecated template API to shared-per-image snapshots, matching
the CLI migration shipped in #2824. `start()` now resolves snapshots by
name via `list_snapshots(name_contains=...)` /
`create_sandbox(snapshot_name=...)` and routes per-trial vCPU / memory /
filesystem limits onto `create_sandbox`; `stop()` never deletes the
shared snapshot.

Touches `libs/deepagents` (integration-test fixture only) alongside
`libs/evals` because the shared snapshot fixture has to move to the new
API in lock-step with the harbor env.

## SDK dependency

The Python-side new-kwargs support was added in
langchain-ai/langsmith-sdk#2769 (Python `snapshot_name` on
`create_sandbox` + `name_contains`/`limit`/`offset` on `list_snapshots`)
and released in `langsmith==0.7.34` on PyPI on 2026-04-23 (release PR:
langchain-ai/langsmith-sdk#2778). JS parity (#2770) shipped earlier in
`[email protected]` (#2774) but is not consumed here.

`libs/evals/pyproject.toml` now pins `langsmith>=0.7.34`; the two `# ty:
ignore[unknown-argument]` suppressions that the original draft mentioned
have been removed (`ty check` is clean); `libs/evals/uv.lock` has been
regenerated.

## How did you verify your code works?

- `uv run --group test ruff check` clean on changed files
- `uv run --group test ty check deepagents_evals/ deepagents_harbor/
tests/unit_tests/` — clean (no "unused ignore" warnings)
- `uv run --group test pytest tests/unit_tests/ -q` in `libs/evals` —
**253 passed**
- Integration test
`libs/deepagents/tests/integration_tests/test_langsmith_sandbox.py`
against live LangSmith prod — **81 passed, 4 xfailed** (the markers
explicitly declared in the file). Exercises the full new API end-to-end:
`list_snapshots(name_contains=...)` → `create_snapshot` (when missing) →
`create_sandbox(snapshot_name=...)` → `delete_sandbox(...)`. Verified
post-run that the shared `deepagents-cli` snapshot is preserved
(`status='ready'`) and the sandbox was cleaned up (0 leftovers).
- One pre-existing `@pytest.mark.xfail(strict=True)` from #2694 on
`test_awrite_aread_adownload_large_text_with_escaped_content` surfaces
as an XPASS against the 0.7.34 SDK — orthogonal to this PR; a follow-up
cleanup will drop that stale marker.

---

_Used Cursor (Claude Opus 4.7) to draft the migration following an
existing plan; all code reviewed and tested locally._

Made with [Cursor](https://cursor.com)

Bumps langsmith Python from 0.7.34 → 0.7.35 to publish the hub methods
merged in #2746.

Unblocks deepagents PR langchain-ai/deepagents#2898, which imports
`AgentEntry`/`FileEntry`/`SkillEntry` from `langsmith.schemas` and calls
`Client.pull_agent`/`push_agent`.

Same two-file scope as #2778.

…angchain-ai#2888)

Migrates the LangSmith harbor env (and its integration-test fixture)
from the deprecated template API to shared-per-image snapshots, matching
the CLI migration shipped in langchain-ai#2824. `start()` now resolves snapshots by
name via `list_snapshots(name_contains=...)` /
`create_sandbox(snapshot_name=...)` and routes per-trial vCPU / memory /
filesystem limits onto `create_sandbox`; `stop()` never deletes the
shared snapshot.

Touches `libs/deepagents` (integration-test fixture only) alongside
`libs/evals` because the shared snapshot fixture has to move to the new
API in lock-step with the harbor env.

The Python-side new-kwargs support was added in
langchain-ai/langsmith-sdk#2769 (Python `snapshot_name` on
`create_sandbox` + `name_contains`/`limit`/`offset` on `list_snapshots`)
and released in `langsmith==0.7.34` on PyPI on 2026-04-23 (release PR:
langchain-ai/langsmith-sdk#2778). JS parity (langchain-ai#2770) shipped earlier in
`[email protected]` (langchain-ai#2774) but is not consumed here.

`libs/evals/pyproject.toml` now pins `langsmith>=0.7.34`; the two `# ty:
ignore[unknown-argument]` suppressions that the original draft mentioned
have been removed (`ty check` is clean); `libs/evals/uv.lock` has been
regenerated.

- `uv run --group test ruff check` clean on changed files
- `uv run --group test ty check deepagents_evals/ deepagents_harbor/
tests/unit_tests/` — clean (no "unused ignore" warnings)
- `uv run --group test pytest tests/unit_tests/ -q` in `libs/evals` —
**253 passed**
- Integration test
`libs/deepagents/tests/integration_tests/test_langsmith_sandbox.py`
against live LangSmith prod — **81 passed, 4 xfailed** (the markers
explicitly declared in the file). Exercises the full new API end-to-end:
`list_snapshots(name_contains=...)` → `create_snapshot` (when missing) →
`create_sandbox(snapshot_name=...)` → `delete_sandbox(...)`. Verified
post-run that the shared `deepagents-cli` snapshot is preserved
(`status='ready'`) and the sandbox was cleaned up (0 leftovers).
- One pre-existing `@pytest.mark.xfail(strict=True)` from langchain-ai#2694 on
`test_awrite_aread_adownload_large_text_with_escaped_content` surfaces
as an XPASS against the 0.7.34 SDK — orthogonal to this PR; a follow-up
cleanup will drop that stale marker.

---

_Used Cursor (Claude Opus 4.7) to draft the migration following an
existing plan; all code reviewed and tested locally._

Made with [Cursor](https://cursor.com)