engineering-agent-loop-public · Nov 13, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -12,10 +12,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Changed
 - Capped `beaupy` below 5 so a lock refresh cannot take a major version silently.
-- Refreshed the bundled minimal workspace `uv.lock` to the latest within-constraint versions (`dlt` 1.31.0 → 1.32.0, `dlthub` 0.30.0 → 0.32.0, `dlthub-client` 0.28.7 → 0.28.8, `fastmcp` 4.0.11 → 4.1.6, `marimo` 0.25.1 → 0.25.2, `pyarrow` 25.0.1 → 26.0.1), plus 48 transitive updates, dropped `backports-tarfile`, `importlib-metadata`, `jaraco-classes`, `jaraco-context`, `jaraco-functools`, `jeepney`, `keyring`, `more-itertools`, `pywin32-ctypes`, `secretstorage`, `zipp`.
+- Refreshed the bundled minimal workspace `uv.lock` to the latest within-constraint versions (`dlt` 1.31.0 → 1.32.0, `dlthub` 0.30.0 → 0.32.0, `dlthub-client` 0.28.7 → 0.28.8, `fastmcp` 4.0.11 → 4.1.8, `marimo` 0.25.1 → 0.25.2, `pyarrow` 25.0.1 → 26.0.1), plus 55 transitive updates and 13 transitive removals.
 - Synced the bundled minimal workspace dependencies with dlt's `workspace_deps()`: changed `duckdb` to `duckdb>=0.9,<2`.
-- Refreshed the root `uv.lock` to the latest within-constraint versions (`beaupy` 3.12.0 → 4.0.1, `mypy` 2.4.0 → 2.4.1, `posthog` 7.64.1 → 7.84.3, `pytest` 9.1.1 → 9.2.0), plus 4 transitive updates, dropped `emoji`.
-- Refreshed the bundled skills to workbench `0c2b3a3`.
+- Refreshed the root `uv.lock` to the latest within-constraint versions (`beaupy` 3.12.0 → 4.0.1, `mypy` 2.4.0 → 2.4.1, `posthog` 7.64.1 → 7.91.0, `pytest` 9.1.1 → 9.2.0), plus 5 transitive updates and 3 transitive removals.
+- Refreshed the bundled skills to workbench `4594950`.
 
 ### Fixed
 - `make scaffold-deps-sync` works with dlt 1.32.0, where `WORKSPACE_DEPS` became `workspace_deps()` (#43).

diff --git a/scripts/generate_skills.py b/scripts/generate_skills.py
@@ -21,8 +21,8 @@
 
 WORKBENCH_REPO = os.environ.get("DLTHUB_WORKBENCH_REPO", "https://github.com/dlt-hub/dlthub-ai-workbench.git")
 WORKBENCH_BRANCH = "master"
-WORKBENCH_REF_SHORT = "0c2b3a3"
-WORKBENCH_REF = "0c2b3a340de9bab78bee8de91c8a99c051e88740"
+WORKBENCH_REF_SHORT = "4594950"
+WORKBENCH_REF = "45949505da3c8878cfd8713c3e7ae6ff3c675618"
 SKILL_TOOLKITS = ("init",)
 
 _KEEP = ".gitkeep"

diff --git a/skills/create-background-agent/SKILL.md b/skills/create-background-agent/SKILL.md
@@ -11,7 +11,7 @@ command line. An agent is an `AGENT.md`, written much like a `SKILL.md`. The wor
 `workbench/dlthub-platform/dlthub/agents/job-inspector/AGENT.md`, installed as
 `.claude/dlthub/agents/dlthub-platform/job-inspector/AGENT.md`.
 
-**Essential reading**: AI harness `https://dlthub.com/docs/hub/ai-harness/introduction.md`,
+**Essential reading**: dltHub AI Harness `https://dlthub.com/docs/hub/ai-harness/introduction.md`,
 triggers and scheduling `https://dlthub.com/docs/hub/pipeline-operations/triggers.md`, profiles
 `https://dlthub.com/docs/hub/pipeline-operations/profiles.md`
 
@@ -74,7 +74,7 @@ subagents. A workspace refers to an installed agent as
 `<toolkit>:<name>`, and the workspace's toolkit index (`.dlt/.toolkits`) travels with every
 deployment, so the reference resolves on the runner as it does locally.
 
-Shipping an agent in a toolkit means working in a clone of the AI harness repo
+Shipping an agent in a toolkit means working in a clone of the harness repo
 (`https://github.com/dlt-hub/dlthub-ai-harness`), since the folder lives beside the toolkit's
 skills and `make validate-toolkits` checks it there. The definition then travels to everyone who
 installs the toolkit, and upgrading it costs a change in the repo and a reinstall in every

diff --git a/skills/create-background-agent/agent-md-reference.md b/skills/create-background-agent/agent-md-reference.md
@@ -336,7 +336,7 @@ the `AGENT__MODEL` variable, which every agent job in that workspace reads. `run
 `model=`, and configuration outranks it, so a value in the deployment code is silently beaten by
 the variable; leave it out and the two cannot disagree.
 
-A definition shipped in an AI harness toolkit names no model. An alias resolves on Anthropic, OpenAI
+A definition shipped in a dltHub AI Harness toolkit names no model. An alias resolves on Anthropic, OpenAI
 and Google; an Azure workspace addresses a deployment on its own endpoint and has no alias, so a
 shipped `model: sonnet` is a default it cannot resolve. `make validate-toolkits` rejects one.
 

diff --git a/skills/create-background-agent/check_agent.py b/skills/create-background-agent/check_agent.py
@@ -172,7 +172,7 @@ def project_root(start):
 
 
 def component_path(root, kind, toolkit, name):
-    """An installed component, or the same component in an AI harness checkout."""
+    """An installed component, or the same component in a dltHub AI Harness checkout."""
     candidates = []
     for host in HOST_DIRS:
         if kind == "rules":

diff --git a/skills/create-background-agent/deployment.md b/skills/create-background-agent/deployment.md
@@ -321,7 +321,7 @@ docs give rather than as evidence of wiring. `trace["mcp_features"]` repeats the
 the file and says nothing about the server.
 
 The run above never starts a server: `_build_toolsets` returns `[]`, which keeps the check offline
-and quick, and it is what the harness repo's `tests/agents/conftest.py` does. Drop that override
+and quick, and it is what the dltHub AI Harness repo's `tests/agents/conftest.py` does. Drop that override
 and the real workspace MCP server starts over stdio. The trace reads the same either way, so the
 evidence is in the transcript, in the server's own two lines:
 
@@ -350,7 +350,7 @@ trace = result["trace"]
 report = result["result"]
 ```
 
-The AI harness repo wires the same loop into pytest in `tests/agents/conftest.py`, where the
+The harness repo wires the same loop into pytest in `tests/agents/conftest.py`, where the
 plugin registers on `Container()[PluginContext].manager` and the loop type comes from an
 `AGENT__LOOP` environment variable rather than `loop=`. A fixture there installs the toolkit into
 a temporary workspace, and `tests/agents/test_agents_run.py` is the shape of a test over it.