Skip to content

feat(skills): skill bundles for coded LangGraph agents [PC-5187] - #1179

Draft
cristian-groza wants to merge 26 commits into
mainfrom
feat/skills-coded-agents
Draft

feat(skills): skill bundles for coded LangGraph agents [PC-5187]#1179
cristian-groza wants to merge 26 commits into
mainfrom
feat/skills-coded-agents

Conversation

@cristian-groza

@cristian-groza cristian-groza commented Oct 29, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

PR stack (skills for coded agents and the advanced harness), merge in this order

  1. [1/2] feat(skills): skill bundles for coded agents [PC-5187] uipath-python#1962: sdk.skills in uipath-platform 0.2.42 and skillBundle overrides in uipath 2.14.45.
  2. [2/2] UiPath/uipath-langchain-python (this PR): kept as a draft until [1/2] is released. Then it raises uipath-platform / uipath to those versions here and in samples/skills-agent/pyproject.toml, refreshes uv.lock and stamps the release notes. Until then CI cannot import uipath.platform.skills.
  3. [3/3] UiPath/uipath-agents-python#829: the low-code advanced harness, which now imports the middleware and prompt from this PR instead of keeping its own copy.

Summary

The LangChain side of skill bundles for coded agents. uipath.platform.skills resolves and downloads the bundles; this PR wires them into a LangGraph agent.

  • UiPathSkills: declares an agent's bundles at import time and resolves them once per process, safely across threads and event loops. It then hands out:
    • load_skill_tool(): a load_skill(name) tool that returns a skill's SKILL.md. Its output is capped to fit the context budget, and unknown names get a helpful error.
    • prompt_section(): the ## Skills system-prompt block.
    • advanced_middleware(backend) for deepagents agents.
  • Building blocks for an already-resolved SkillSet: create_load_skill_tool, skills_prompt_section.
  • Advanced (deepagents) agents: UiPathSkillsMiddleware, SkillLoadTraceMiddleware and create_skills_middleware. These moved here from uipath-agents-python (UiPath/uipath-agents-python#829) so coded and low-code agents share one implementation. The prompt is word for word the harness text, including D15: skill scripts are stored but cannot be executed. The module is imported lazily, so deepagents is only needed by agents that use it.
  • Sample samples/skills-agent: a LangGraph agent that binds two bundles (finance pinned to v3, ops latest), with local skills and bindings.json. Its README covers the local source, the orchestrator placeholder, overrides and traces.
  • Docs: docs/skills.md. The mkdocs nav entry goes into uipath-python once this page is merged.
  • Version: uipath-langchain 0.18.52.

Test plan

  • pytest tests/skills (110 passed), ruff check, ruff format --check, mypy src/uipath_langchain/skills tests/skills. Run locally against [1/2] installed in editable mode.
  • commitlint (commitlint.config.mjs).
  • End to end on the staging tenant with the sample:
    • uipath init generated the graph with its tools node.
    • Both bundles resolved through UIPATH_SKILLS_SOURCE=local.
    • Claude Haiku 4.5 called load_skill("ops/escalation-policy") and followed its steps, and skills_used was ["ops/escalation-policy"].
    • The run succeeded, and trace.jsonl holds the Skill bundles resolved and Skill load - ops/escalation-policy spans.
  • CI is green once [1/2] is released and the floors are raised.

🤖 Generated with Claude Code

cristian-groza and others added 11 commits October 28, 2026 22:02
@cristian-groza @claude
feat(skills): add load_skill tool and prompt section for coded agents…
e5979bf 
… [PC-5187]

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@cristian-groza @claude
fix(skills): keep the load_skill tool inside the run and the context …
20e6d20 
…budget [PC-5187]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
feat(skills): add UiPathSkills helper that resolves bound bundles onc…
f182aff 
…e per process [PC-5187]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
fix(skills): resolve bound bundles once across threads and loops [PC-…
02eba76 
…5187]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
feat(skills): move the deepagents skills middleware into uipath-langc…
ff5d971 
…hain [PC-5187]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
fix(skills): keep the same-loop fallback and cleanup off the shared f…
022a8a1 
…older [PC-5187]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
docs(skills): add the skills sample agent and guide [PC-5187]
ca10d9b 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
fix(skills): decide and store a finished resolution under one lock [P…
027985f 
…C-5187]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
test(skills): configure the SDK mock before building UiPathSkills [PC…
ef68d97 
…-5187]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
docs(skills): tighten the skills sample and guide [PC-5187]
3eae999 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
chore: bump version to 0.18.52 for skills support in coded agents [PC…
3e0ba1a 
…-5187]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza cristian-groza mentioned this pull request Oct 29, 2026
Open
5 tasks done
cristian-groza and others added 15 commits October 29, 2026 13:06
@cristian-groza @claude
fix(skills): say skill scripts cannot be executed, as the low-code ha…
c8e6909 
…rness does [PC-5187]

D15: scripts are stored, never run. The advanced-agent prompt now matches the
uipath-agents-python harness word for word, so both runtimes give the same rule.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
docs(skills): run the sample on the local source, as the SDK drops `m…
70e1ca7 
…ock` [PC-5187]

uipath-platform no longer ships the Skills API client or the `mock`
source, and `orchestrator` is a placeholder until the API exists. The
sample, its README and the docs use `local`; the sample's override test
retargets a binding instead of pinning a version the folder lacks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
docs(skills): give the sample's override every binding field [PC-5187]
c304fb7 
A skillBundle resource overwrite now replaces the whole binding, as for the
other bindings: name, folderPath and versionMode, plus version when pinned. The
sample's example override and the override test carry the full entry, and the
README says an incomplete entry is skipped with a warning.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
refactor(skills): import the skills middleware from its submodule [PC…
f36473f 
…-5187]

uipath_langchain.skills re-exported the four middleware names through a
module __getattr__, so the load_skill path would not import deepagents. The
names now come from uipath_langchain.skills.middleware directly: the package
init loses the lazy loader and still never imports deepagents.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
docs(skills): shorten the skills package docstring [PC-5187]
8e1cc9c 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
test(skills): drop the prompt text tests [PC-5187]
9cf20c4 
test_prompt.py mostly asserted the wording of the skills prompts. The
remaining skills tests still exercise skills_prompt_section and the
SkillsMiddleware template through UiPathSkills and the middleware.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
test(skills): drop load_skill tests that repeat other tests [PC-5187]
2c95a7b 
The plain async call is covered by the async span test, the enum rejection by
pydantic and the ToolNode test, and "supporting files record no span" by the
SDK's SkillSet.read tests. The metadata test no longer pins the description
wording.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
test(skills): drop middleware tests that repeat other tests [PC-5187]
a160288 
The rendered-section test duplicated the modify_request test through private
methods, the async missing-folder test the shared error path, and the empty
deepagents_sources test the empty case of advanced_middleware.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
test(skills): trim the sample graph tests to the integration paths [P…
687c8b6 
…C-5187]

The bundle-source check moves into the main run test, which already ran the
same script, and the unit test of the sample's own skills_used helper goes.
The four tests left run UiPathSkills on the real SDK and local source inside a
LangGraph graph: a skill load, a failed load, an override and no skills.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
refactor(skills): resolve skills in the graph factory only [PC-5187]
4666ba8 
The runtime calls a graph factory once per entrypoint, on one event loop, so
UiPathSkills no longer coordinates loads across threads and loops. load_sync,
the lazy load_skill tool for module-level graphs and the machinery behind them
(the thread lock, futures, the cleanup generation and the same-loop fallback)
are gone; load is an asyncio.Lock around one cached resolution, and
load_skill_tool raises before load. The docs drop the module-level graph
section, and the tests keep the contract the factory path relies on.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
feat(skills): run a SKILL.md read inside its Skill load span [PC-5187]
738bb9c 
SkillLoadTraceMiddleware now decides from the call's arguments whether
read_file reads a bound SKILL.md, and runs the tool inside the SDK's
skill_load_span. The span is current while the tool runs, so the
read_file tool call nests under "Skill load - <bundle>/<skill>". A read
that returns an error, or a Command without a ToolMessage, leaves the
span with an error status; a Command carrying a ToolMessage counts as a
load. Tracing failures still never affect the tool call.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
Merge origin/main into feat/skills-coded-agents [PC-5187]
0762323 
main moved uipath-langchain to 0.18.55 and raised the uipath minimums.
Take main's dependencies and keep the skills bump on top: 0.18.56.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
Merge origin/main into feat/skills-coded-agents [PC-5187]
26d48f9 
Main moved to uipath 2.15 / uipath-platform 0.3 and released 0.19.0;
the branch version becomes 0.19.1 in pyproject.toml and uv.lock.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
chore(samples): require the skills releases in the skills sample [PC-…
ed533fa 
…5187]

The skills code ships in uipath 2.15.1 and uipath-langchain 0.19.1.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cristian-groza @claude
Merge origin/main into feat/skills-coded-agents [PC-5187]
3dd7490 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@sonarqubecloud

sonarqubecloud Bot commented Nov 5, 2026

Copy link
Copy Markdown

Quality Gate Failed Quality Gate failed

Failed conditions
0.0% Coverage on New Code (required ≥ 90%)

See analysis details on SonarQube Cloud

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

Copilot code review Copilot
Copilot will automatically review once the pull request is marked ready for review

At least 1 approving review is required to merge this pull request.

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@cristian-groza